Loading [MathJax]/jax/element/mml/optable/SuppMathOperators.js
Research article

Wearable on-device deep learning system for hand gesture recognition based on FPGA accelerator

  • Gesture recognition is critical in the field of Human-Computer Interaction, especially in healthcare, rehabilitation, sign language translation, etc. Conventionally, the gesture recognition data collected by the inertial measurement unit (IMU) sensors is relayed to the cloud or a remote device with higher computing power to train models. However, it is not convenient for remote follow-up treatment of movement rehabilitation training. In this paper, based on a field-programmable gate array (FPGA) accelerator and the Cortex-M0 IP core, we propose a wearable deep learning system that is capable of locally processing data on the end device. With a pre-stage processing module and serial-parallel hybrid method, the device is of low-power and low-latency at the micro control unit (MCU) level, however, it meets or exceeds the performance of single board computers (SBC). For example, its performance is more than twice as much of Cortex-A53 (which is usually used in Raspberry Pi). Moreover, a convolutional neural network (CNN) and a multilayer perceptron neural network (NN) is used in the recognition model to extract features and classify gestures, which helps achieve a high recognition accuracy at 97%. Finally, this paper offers a software-hardware co-design method that is worth referencing for the design of edge devices in other scenarios.

    Citation: Weibin Jiang, Xuelin Ye, Ruiqi Chen, Feng Su, Mengru Lin, Yuhanxiao Ma, Yanxiang Zhu, Shizhen Huang. Wearable on-device deep learning system for hand gesture recognition based on FPGA accelerator[J]. Mathematical Biosciences and Engineering, 2021, 18(1): 132-153. doi: 10.3934/mbe.2021007

    Related Papers:

    [1] Liuchao Xiao, Wenbo Li, Leilei Wei, Xindong Zhang . A fully discrete local discontinuous Galerkin method for variable-order fourth-order equation with Caputo-Fabrizio derivative based on generalized numerical fluxes. Networks and Heterogeneous Media, 2023, 18(2): 532-546. doi: 10.3934/nhm.2023022
    [2] Tingfu Yao, Changlun Ye, Xianbing Luo, Shuwen Xiang . A variational MAX ensemble numerical algorism for a transient heat model with random inputs. Networks and Heterogeneous Media, 2024, 19(3): 1013-1037. doi: 10.3934/nhm.2024045
    [3] Shi Jin, Min Tang, Houde Han . A uniformly second order numerical method for the one-dimensional discrete-ordinate transport equation and its diffusion limit with interface. Networks and Heterogeneous Media, 2009, 4(1): 35-65. doi: 10.3934/nhm.2009.4.35
    [4] Peng E, Tingting Xu, Linhua Deng, Yulin Shan, Miao Wan, Weihong Zhou . Solutions of a class of higher order variable coefficient homogeneous differential equations. Networks and Heterogeneous Media, 2025, 20(1): 213-231. doi: 10.3934/nhm.2025011
    [5] Salim Meddahi, Ricardo Ruiz-Baier . A new DG method for a pure–stress formulation of the Brinkman problem with strong symmetry. Networks and Heterogeneous Media, 2022, 17(6): 893-916. doi: 10.3934/nhm.2022031
    [6] Yaxin Hou, Cao Wen, Yang Liu, Hong Li . A two-grid ADI finite element approximation for a nonlinear distributed-order fractional sub-diffusion equation. Networks and Heterogeneous Media, 2023, 18(2): 855-876. doi: 10.3934/nhm.2023037
    [7] Peter Bella, Arianna Giunti . Green's function for elliptic systems: Moment bounds. Networks and Heterogeneous Media, 2018, 13(1): 155-176. doi: 10.3934/nhm.2018007
    [8] Farman Ali Shah, Kamran, Dania Santina, Nabil Mlaiki, Salma Aljawi . Application of a hybrid pseudospectral method to a new two-dimensional multi-term mixed sub-diffusion and wave-diffusion equation of fractional order. Networks and Heterogeneous Media, 2024, 19(1): 44-85. doi: 10.3934/nhm.2024003
    [9] Gaziz F. Azhmoldaev, Kuanysh A. Bekmaganbetov, Gregory A. Chechkin, Vladimir V. Chepyzhov . Homogenization of attractors to reaction–diffusion equations in domains with rapidly oscillating boundary: Critical case. Networks and Heterogeneous Media, 2024, 19(3): 1381-1401. doi: 10.3934/nhm.2024059
    [10] Kexin Li, Hu Chen, Shusen Xie . Error estimate of L1-ADI scheme for two-dimensional multi-term time fractional diffusion equation. Networks and Heterogeneous Media, 2023, 18(4): 1454-1470. doi: 10.3934/nhm.2023064
  • Gesture recognition is critical in the field of Human-Computer Interaction, especially in healthcare, rehabilitation, sign language translation, etc. Conventionally, the gesture recognition data collected by the inertial measurement unit (IMU) sensors is relayed to the cloud or a remote device with higher computing power to train models. However, it is not convenient for remote follow-up treatment of movement rehabilitation training. In this paper, based on a field-programmable gate array (FPGA) accelerator and the Cortex-M0 IP core, we propose a wearable deep learning system that is capable of locally processing data on the end device. With a pre-stage processing module and serial-parallel hybrid method, the device is of low-power and low-latency at the micro control unit (MCU) level, however, it meets or exceeds the performance of single board computers (SBC). For example, its performance is more than twice as much of Cortex-A53 (which is usually used in Raspberry Pi). Moreover, a convolutional neural network (CNN) and a multilayer perceptron neural network (NN) is used in the recognition model to extract features and classify gestures, which helps achieve a high recognition accuracy at 97%. Finally, this paper offers a software-hardware co-design method that is worth referencing for the design of edge devices in other scenarios.


    In this paper we consider the numerical approximations of the following problem

    C0Dαtu(x,t)+2x2(ω(x)2u(x,t)x2)+κu(x,t)=f(x,t),0<x<L,0<tT, (1.1)
    u(x,0)=φ(x),0<x<L, (1.2)
    u(0,t)=α1(t),u(L,t)=α2(t),0tT, (1.3)
    2u(0,t)x2=β1(t),2u(L,t)x2=β2(t),0tT, (1.4)

    where κ0 is given constant, φ(x),α1(t),α2(t),β1(t),β2(t) and f(x,t) are given sufficiently smooth functions satisfying φ(0)=α1(0),φ(L)=α2(0),φ(0)=β1(0) and φ(L)=β2(0), C0Dαtu(x,t) denotes Caputo fractional derivative defined by

    C0Dαtu(x,t)=1Γ(1α)t0u(x,s)s1(ts)αds,0<α<1.

    And we suppose that there exist two constants C1 and C2 such that 0<C1ω(x)C2 for 0xL.

    More and more attention has been paid to the fractional differential equations (FDEs) due to its application foreground in chemistry, physics, finance and hydrology in the past twenty years [1,2,3,4]. As we know, the analytic solutions of FDEs are very difficult to obtain, some efficient numerical methods should be considered, especially fast algorithms with high order accuracy. Some essential definitions and properties of fractional derivatives can refer to monograph [5].

    This target problem in Eq (1.1) is frequently employed to simulate some phenomena in physics, such as wave propagation in beams, brain warping, ice formation and designing special curves on surfaces and so on, e.g., [6,7,8,9,10,11] and their references.

    Up to now considerable works have been done from theoretical and numerical point of view for fourth-order fractional diffusion equations. For instance, Hu and Zhang successively presented a finite difference scheme for the fourth-order fractional diffusion-wave and sub-diffusion equations, and a compact difference scheme for the former, see [12,13]. Ji et al.[14] constructed a compact difference scheme for the fourth-order fractional sub-diffusion equation under the fist Dirichlet boundary conditions. Zhang and Pu [15] presented a compact difference scheme for such equation by L21σ formula [16]. Ran and Zhang [17] presented a new compact difference schemes for the such equation of the distributed order.

    However, most of the work focus on the constant coefficient case. Recently, Zhao and Xu [18] presented a compact difference scheme for the time fractional sub-diffusion equation with the variable coefficient under the Dirichlet boundary conditions. Subsequently, based on the subtle decomposition of the coefficient matrices, Vong, Lyu and Wang [19] presented a compact difference scheme to solve the equations under Neumann boundary conditions. But the above works has only accuracy of order 2α in time.

    In this paper, our attention will be paid on the higher order difference scheme for solving the variable coefficient equations under the second Dirichlet boundary conditions For this purpose, we use the L21σ formula to approximate the Caputo fractional derivative. Unlike the integer order case, the time fractional derivative requires all history information. In order to reduce the computational complexity, we also construct a fast difference scheme. The stability and convergence of both schemes are proved in detail.

    The structure of this paper is as follows: In Section 2, some necessary notations and lemmas are first introduced and a second-order difference scheme for the target problem (1.1)–(1.4) is constructed. In Section 3, an important priori estimate is first proved, and the unconditional stability and convergence of scheme are obtained. In Section 4, a fast second-order difference scheme is presented, and the corresponding unconditional stability and convergence are also strictly proved.

    In Section 5, a difference scheme based on nonuniform time grids is first presented, and some numerical examples are provided to verify the theoretical results. A brief conclusion is given finally.

    Let h=L/M and τ=T/N, where M, N are two positive integers. Denote xi=ih,0iM,tn=nτ,0nN, Ωh={xi0iM},Ωτ={tn0nN}. Let Vh={vv=(v0,v1,,vM)} be grid function space on Ωh, and ˚Vh={vvVh,v0=vM=0}. Also we denote σ=1α2,tn+σ=(n+σ)τ and ω(xi)=ωi.

    For uVh, we define

    δxui+12=1h(ui+1ui),δ2xui=1h2(ui+12ui+ui1).

    For any u,v˚Vh, we define the inner products

    (u,v)=hM1i=1uivi,(δxu,δxv)=hM1i=0(δxui+12)(δxvi+12),(u,v)ω=hM1i=1uiviωi,

    and norms

    u=(u,u),uω=(u,u)ω,δxu=(δxu,δxu),(δ2xu,δ2xv)=hM1i=1δ2xuiδ2xvi.

    In [16], Alikhanov developed a new second order difference formula (called L21σ formula) for the Caputo fractional derivative, which can be expressed in the following lemma.

    Lemma 2.1 ([16]). Suppose α(0,1),σ=1α2 and u(t)C3[0,T]. It holds

    C0Dαtu(t)t=tn1+σDατ,σun∣=O(τ3α),

    where

    Dατ,σun=ταΓ(2α)[C(n)0unn1j=1(C(n)nj1C(n)nj)ujC(n)n1u0],

    in which C(n)0=a0=σ1α for n=1, and

    C(n)k={a0+b1,k=0,ak+bk+1bk,1kn2,akbk,k=n1

    for n2, where aj=(j+σ)1α(j1+σ)1α and bj=12α[(j+σ)2α(j1+σ)2α]12[(j+σ)1α+(j1+σ)1α] for all j1.

    Let v(x,t)=2ux2. Then the problem Eqs (1.1)–(1.4) can be written in the equivalent system

    C0Dαtu(x,t)+2x2(ω(x)v(x,t))+κu(x,t)=f(x,t),0<x<L,0<tT, (2.1)
    v(x,t)=2u(x,t)x2,0<x<L,0<tT, (2.2)
    u(x,0)=φ(x),0<x<L, (2.3)
    u(0,t)=α1(t),u(L,t)=α2(t),v(0,t)=β1(t),v(L,t)=β2(t),0tT. (2.4)

    Suppose u(x,t)C(6,3)x,t([0,L]×[0,T]). Define

    Uni=u(xi,tn),Vni=v(xi,tn),0iM,0nN.

    Considering the Eqs.(2.1)–(2.2) at the point (xi,tn1+σ), we obtain

    C0Dαtu(xi,tn1+σ)+2x2(ω(xi)v(xi,tn1+σ))+κu(xi,tn1+σ)=f(xi,tn1+σ), (2.5)
    v(xi,tn1+σ)=2u(xi,tn1+σ)x2. (2.6)

    Using Taylor expansion

    u(xi,tn1+σ)=σUni+(1σ)Un1i+O(τ2)=Un1+σi+O(τ2),

    where Un1+σi=σUni+(1σ)Un1i. Then we obtain

    2u(xi,tn1+σ)x2=δ2xUn1+σi+O(τ2+h2),

    and

    2x2(ω(xi)v(xi,tn1+σ))=δ2x(ωiVn1+σi)+O(τ2+h2).

    Using Lemma 2.1, it follows from Eq (2.5), Eq (2.6) that

    Dατ,σUni+δ2x(ωiVn1+σi)+κUn1+σi=fn1+σi+(R1)ni,1iM1,1nN, (2.7)
    Vn1+σi=δ2xUn1+σi+(R2)ni,1iM1,1nN, (2.8)

    and there exists a constant Cr such that

    (R1)ni+(R2)ni∣≤Cr(τ2+h2),1iM1,1nN. (2.9)

    Omitting the small terms (R1)ni and (R2)ni in Eq (2.7) and Eq (2.8), we present the difference scheme (called L21σ scheme) for the equivalent system (2.1)–(2.4) as follows

    Dατ,σuni+δ2x(ωivn1+σi)+κun1+σi=fn1+σi,1iM1,1nN, (2.10)
    vn1+σi=δ2xun1+σi,1iM1,1nN, (2.11)
    u0i=φ(xi),0iM, (2.12)
    un0=α1(tn),unM=α2(tn),vn0=β1(tn),vnM=β2(tn),1nN, (2.13)

    where the initial-boundary conditions Eq (2.3), Eq (2.4) have been used.

    Theorem 2.2. The above difference scheme (2.10)–(2.13) is equivalent to

    kun1+σ1+Dατ,σun1+1h2(ω0β1(tn1+σ)+ω2δ2xun1+σ22ω1δ2xun1+σ1)=fn1+σ1, (2.14)
    Dατ,σuni+δ2x(ωiδ2xun1+σi)+kun1+σi=fn1+σi,2iM2, (2.15)
    kun1+σM1+Dατ,σunM1+1h2(ωMβ2(tn1+σ)+ωM2δ2xun1+σM22ωM1δ2xun1+σM1)=fn1+σM1, (2.16)
    u0i=φ(xi),0iM, (2.17)
    u0=α1(tn),unM=α2(tn). (2.18)

    Proof. Since

    δ2xω1vn1+σ1=1h2(ω0vn1+σ02ω1vn1+σ1+ω2vn1+σ2),δ2xωM1vn1+σM1=1h2(ωMvn1+σM2ωM1vn1+σM1+ωM2vn1+σM2).

    It follows from Eq (2.11) and Eq (2.13) that

    δ2xω1vn1+σ1=1h2(ω0β1(tn1+σ)2ω1δ2xun1+σ1+ω2δ2xun1+σ2),δ2xωM1vn1+σM1=1h2(ωMβ2(tn1+σ)2ωM1δ2xun1+σM1+ωM2δ2xun1+σM2).

    This together with Eq (2.10), we get Eq (2.14) and Eq (2.16). Eq (2.15) can be obtained by substituting Eq (2.11) into Eq (2.10). This proof is completed.

    The above equivalent form Eqs (2.14)–(2.18) will be used only in calculation.

    We first introduce the following essential lemmas.

    Lemma 3.1 ([16]). Suppose α(0,1) and C(n)k is defined in Lemma 2.1. It holds that

    C(n)0>C(n)1>C(n)2>>C(n)n2>C(n)n1,andC(n)k>1α2(k+σ)α.

    Lemma 3.2 ([16]). Suppose u={un0nN} is a grid function defined on Ωτ. It holds that

    (σun+(1σ)un1)Dατ,σun12Dατ,σ(un)2.

    Lemma 3.3 ([20,21]). For any u˚Vh, it holds that

    uL6δxu,δxuL6δ2xu.

    The following Lemma will be used in the analysis of the difference scheme.

    Lemma 3.4. For any u˚Vh, it holds that

    C1u2u2ωC2u2,C1δ2xu2δ2xu2ωC2δ2xu2.

    Proof. The proof is straightforward from the definition of |||| and ||||ω.

    We next show the priori estimate of the scheme (2.10)–(2.13).

    Theorem 3.5. Suppose {wni0iM,0nN} and {zni0iM,0nN} satisfy the following difference scheme

    Dατ,σwni+δ2x(ωizn1+σi)+κwn1+σi=pn1+σi,1iM1,1nN, (3.1)
    zn1+σi=δ2xwn1+σi+qn1+σi,1iM1,1nN, (3.2)
    wni=φ(xi),0iM, (3.3)
    wn0=0,wnM=0,zn0=0,znM=0,1nN. (3.4)

    Then, it holds that

    wn2w02+2TαΓ(1α)(L418C1max1nNpn1+σ2+2C2max1nNqn1+σ2). (3.5)

    Proof. Taking the inner product of Eq (3.1) by wn1+σ, we get

    (Dατ,σwn,wn1+σ)+(δ2x(ωzn1+σ),wn1+σ)+κwn1+σ2=(pn1+σ,wn1+σ). (3.6)

    Taking the inner product of Eq (3.2) by ωzn1+σ, we get

    (zn1+σ,ωzn1+σ)=(δ2xwn1+σ,ωzn1+σ)+(qn1+σ,ωzn1+σ). (3.7)

    From Eq (3.6) and Eq (3.7), it yields that

    (Dατ,σwn,wn1+σ)+(δ2x(ωzn1+σ),wn1+σ)+κwn1+σ2+(zn1+σ,ωzn1+σ)=(pn1+σ,wn1+σ)+(δ2xwn1+σ,ωzn1+σ)+(qn1+σ,ωzn1+σ). (3.8)

    Applying the discrete Green formula gives that

    (δ2x(ωzn1+σ),wn1+σ)=(δx(ωzn1+σ),δxwn1+σ)=(δ2xwn1+σ,ωzn1+σ). (3.9)

    Substituting Eq (3.9) into Eq (3.8), we obtain

    (Dατ,σwn,wn1+σ)+κwn1+σ2+zn1+σ2ω=(pn1+σ,wn1+σ)+(qn1+σ,zn1+σ)ω. (3.10)

    From Eq (3.2), we have

    (zn1+σi)2=(δ2xwn1+σi+qn1+σi)2. (3.11)

    Multiplying Eq (3.11) by hωi and summing up for i from 1 to M1, we get

    ||zn1+σ||2ω=||δ2xwn1+σ||2ω+2(δ2xwn1+σ,qn1+σ)ω+||qn1+σ||2ω. (3.12)

    Substituting Eq (3.12) into Eq (3.10), we obtain

    (Dατ,σwn,wn1+σ)+12zn1+σ2ω+12δ2xwn1+σ2ω+12qn1+σ2ω+κwn1+σ2=(pn1+σ,wn1+σ)+(qn1+σ,zn1+σ)ω(δ2xwn1+σ,qn1+σ)ω. (3.13)

    Using Cauchy-Schwarz inequality, we have

    (δ2xwn1+σ,qn1+σ)ω14δ2xwn1+σ2ω+qn1+σ2ω, (3.14)

    and

    (qn1+σ,zn1+σ)ω12zn1+σ2ω+12qn1+σ2ω, (3.15)

    From Eq (3.14), Eq (3.15) and Eq (3.13), we obtain

    (Dατ,σwn,wn1+σ)+14δ2xwn1+σ2ω(pn1+σ,wn1+σ)+qn1+σ2ω. (3.16)

    Based on Lemma 3.3 and Lemma 3.4, we have

    w2L436C1δ2xw2ω,qn1+σ2ωC2qn1+σ2. (3.17)

    Applying Cauchy inequality, we get

    (pn1+σ,wn1+σ)9C1L4wn1+σ2+L436C1pn1+σ214δ2xwn1+σ2ω+L436C1pn1+σ2. (3.18)

    Substituting Eq (3.18) into Eq (3.16) yields that

    Dατ,σwn2L418C1pn1+σ2+2C2qn1+σ2.

    where Lemma 3.2 has been used. That is,

    C(n)0wn2n1k=1(C(n)nk1C(n)nk)wk2+C(n)n1w02+μ(L418C1pn1+σ2+2C2qn1+σ2), (3.19)

    where μ=Γ(2α)τα. According to Lemma 3.1, we have

    C(n)n1>1α2(n1α2)α>1α2(nα2)α,1nN,

    and

    μ=ταΓ(2α)=TαNαΓ(1α)(1α)<Tα(nα2)αΓ(1α)(1α)<2C(n)n1TαΓ(1α). (3.20)

    Substituting Eq (3.20) into Eq (3.19) gives that

    C(n)0wn2n1k=1(C(n)nk1C(n)nk)wk2+C(n)n1[w02+2TαΓ(1α)(L418C1pn1+σ2+2C2qn1+σ2)].

    Denote

    J=w02+2TαΓ(1α)(L418C1max1nNpn1+σ2+2C2max1nNqn1+σ2).

    Now, we prove by the mathematical induction method that

    wn2J. (3.21)

    It holds obviously when n=0. Assuming Eq (3.21) is valid for n=1,2,,m1, then we have

    C(m)0wm2m1k=1(C(m)mk1C(m)mk)wk2+C(m)m1Jm1k=1(C(m)mk1C(m)mk)J+C(m)m1J=C(m)0J.

    This proof is completed.

    Applying the Theorem 3.5, we can immediately obtain the stability result.

    Theorem 3.6 (Stability). The difference scheme (2.10)–(2.13) is unconditionally stable with respect to the initial value φ and the source term f.

    Similarly, from Theorem 3.5, we can easily prove the solvability of the proposed scheme.

    Theorem 3.7 (Solvability). The difference scheme (2.10)–(2.13) is uniquely solvable.

    Proof. It suffices to prove the homogeneous linear system

    Dατ,σuni+δ2x(ωivn1+σi)+κun1+σi=0,1iM1,1nN,vn1+σi=δ2xun1+σi,1iM1,1nN,u0i=0,0iM,un0=unM=0,vn0=vnM=0,1nN,

    has only a trivial solution. Applying Theorem 3.1, we have ||un||2||u0||2=0. So uni0 for 0iM, which completes the proof.

    Next, we focus on the convergence of the difference scheme (2.10)–(2.13). Denote

    eni=u(xi,tn)uni,˜eni=v(xi,tn)vni,0nN,0iM.

    Theorem 3.8 (Convergence). Assume that u(x,t)C6,3x,t([0,L]×[0,T]) and {uni} are solution of the problem (1.1)–(1.4) and the difference scheme Eqs (2.10)–(2.13) respectively. Then there exists a positive constant C such that

    ||en||C(τ2+h2),0nN. (3.22)

    Proof. From Eq (2.7), Eq (2.8) and Eqs (2.10)–(2.13), we have the error equations as

    Dατ,σeni+δ2x(ω˜en1+σ)i+κen1+σi=(R1)ni,1iM1,1nN,˜en1+σi=δ2xen1+σi+(R2)ni,1iM1,1nN,e0i=0,0iM,en0=0,enM=0,˜en0=0,˜enM=0,1nN.

    Applying Theorem 3.5, we get

    en22TαΓ(1α)(L418C1max1nNRn12+2C2max1nNRn22),1nN.

    Noticing Eq (2.9), we get

    en22TαΓ(1α)(L418C1+2C2)Cr2(τ2+h2)2,1nN,

    which shows that Eq (3.22) is valid with

    C=Cr2TαΓ(1α)(L418C1+2C2).

    This proof is completed.

    Although the L21σ scheme (2.10)–(2.13) has accuracy of second order in time, it is not conducive to calculation due to it needs all history data to get the solution at current time point. Also, here we present a fast scheme by applying the sum-of-exponentials approximation to the kernel function tα.

    The sum-of-exponentials approximation reads as:

    Lemma 4.1 ([22]). For the given α(0,1), tolerance error ε, cut-off time step size ˜τ and final time T, there are one positive integer Nexp, positive points sj and corresponding positive weights wj(j=1,2,,Nexp) satisfying

    tαNexpj=1wjesjt∣≤ε,tϵ[˜τ,T],

    and the number of exponentials needed is of the order

    Nexp=O(log(1ε(loglog1ε+logT˜τ+log1˜τ(loglog1ε+logT˜τ)).

    The fast evaluation of Caputo derivative, FL21σ formula, is given as follows:

    FDαtun+σ=Nexpj=1˜wj˜Vnj+λa0(un+1un), (4.1)

    where λ=ταΓ(2α), ˜wj=1Γ(1α)wj, and ˜Vnj can be got form the following recursive relation

    ˜Vnj=esjτ˜Vn1j+Aj(unun1)+Bj(un+1un),j=1,2,,Nexp,n=1,2,, (4.2)

    with ˜V0j=0,(j=1,2,,Nexp) and

    Aj=(2+τsj)eτsj(2+3τsj)2(τsj)2e(τsj(σ+1)),Bj=(τsj2)eτsj+(2+τsj)2(τsj)2e(τsj(σ+1)),j1.

    The recursive relation (4.2) shows that the FL21σ formula reduces significantly the computational complexity. Noticing that Eq (4.2) can be equivalently rewritten as the following summation form

    ˜Vnj=e(n1)τsjAj(u1u0)+n1i=1(e(ni1)τsjAj+e(ni)τsjBj)(ui+1ui)+Bj(un+1un),

    thus we have

    FDαtun+σ=nk=0Fg(n+1,α)k(uk+1uk), (4.3)

    in which Fg(1,α)0=λa0, and for n1,

    Fg(n+1,α)k={Nexpj=1˜wje(n1)sjτAj,k=0,Nexpj=1˜wj(e(nk1)sjτAj+e(nk)sjτBj),1kn1,Nexpj=1˜wjBj+λa0,k=n. (4.4)

    The equivalent expression (4.3) is more applicable in stability and convergence analysis.

    With respect to the FL21σ formula, we have the following some results.

    Lemma 4.2 ([22]). For any α(0,1), and u(t)C3[0,T], it holds that

    C0Dαtu(t)t=tn+σFDαtun+σ∣=O(τ3α+ε).

    Lemma 4.3 ([22]). Suppose α(0,1),Fg(n+1,α)k is defined by Eq (4.4), then it holds that

    Fg(n+1,α)n>Fg(n+1,α)n1>>Fg(n+1,α)0

    Lemma 4.4 ([22]). Suppose u = \{ {u^n}\mid0 \le n \le N-1\} is a grid function defined on {{\Omega}_\tau } , then it holds that

    \begin{array}{l} ({\sigma}{u^{n+1}}+(1-{\sigma}){u^{n}}) {}^\mathcal{F} D_t^{\alpha}{u^{n+\sigma}} \ge \frac{1}{2} {}^\mathcal{F} D_t^{\alpha}{(u^{n+\sigma})^2}. \end{array}

    Similar to the derivation of the \mathcal{L}2 - 1_{\sigma} scheme (2.10)–(2.13), it follows from Eq (2.1), Eq (2.2) we have

    \begin{align} &{}^\mathcal{F} D_t^{\alpha}U_i^{n+\sigma} + \delta_x^2{({\omega}_i{V_i^{n+{\sigma}}})} + {\kappa}U_i^{n+{\sigma}} = f_i^{n+{\sigma}} + {}^\mathcal{F} ({R_1})_i^n, \; 1 \le i \le M - 1, 0 \le n \le N-1, \end{align} (4.5)
    \begin{align} &V_i^{n+{\sigma}} = \delta_x^2U_i^{n+{\sigma}}+{}^\mathcal{F} ({R_2})_i^n, 1 \le i \le M-1, 0 \le n \le N-1, \end{align} (4.6)

    and there exists a constant {}^\mathcal{F} C_r such that

    \begin{align} \mid {}^\mathcal{F} ({R_1})_i^n\mid+\mid{}^\mathcal{F} ({R_2})_i^n\mid \le {{}^\mathcal{F} C_r}({\tau ^2}+{h^2}+\varepsilon), 1 \le i \le M-1, 0 \le n \le N-1. \end{align} (4.7)

    Omitting the small terms {}^\mathcal{F} ({R_1})_i^n and {}^\mathcal{F} ({R_2})_i^n in Eq (4.5) and Eq (4.6), from the boundary and initial conditions (2.3)–(2.4), we obtain the \mathcal{FL}2 - 1_{\sigma} scheme for the problem (2.1)–(2.4) as follows

    \begin{align} & {}^\mathcal{F} D_t^{\alpha}u_i^{n+\sigma}+\delta_x^2{({\omega}_i{v_i^{n+{\sigma}}})}+{\kappa}u_i^{n+{\sigma}} = f_i^{n+{\sigma}}, \; 1 \le i \le M-1, 0 \le n \le N-1, \end{align} (4.8)
    \begin{align} & v_i^{n+{\sigma}} = \delta_x^2u_i^{n+{\sigma}}, \; 1 \le i \le M-1, 0 \le n \le N-1, \end{align} (4.9)
    \begin{align} & u_i^0 = \varphi ({x_i}), \; 0 \le i \le M, \end{align} (4.10)
    \begin{align} & u_0^n = {{\alpha}_1}({t_n}), u_M^n = {{\alpha}_2}({t_n}), v_0^n = {{\beta}_1}({t_n}), v_M^n = {{\beta}_2}({t_n}), \; 1 \le n \le N. \end{align} (4.11)

    Next, we focus on the solvability, stability and convergence of the \mathcal{FL}2 - 1_{\sigma} scheme.

    Before the discussion, we first prove the following priori estimate.

    Theorem 4.5. Suppose \{w_i^n, z_i^n\mid0\le i\le M, 0\le n\le N\} satisfy the difference scheme

    \begin{align} & {}^\mathcal{F} D_t^{\alpha}w_i^{n+\sigma}+\delta_x^2{({\omega}_i{z_i^{n+{\sigma}}})}+{\kappa}w_i^{n+{\sigma}} = p_i^{n+{\sigma}}, \; 1 \le i \le M-1, 0 \le n \le N-1, \end{align} (4.12)
    \begin{align} & z_i^{n+{\sigma}} = \delta_x^2w_i^{n+{\sigma}}+q_i^{n+{\sigma}}, \; 1 \le i \le M-1, 0 \le n \le N-1, \end{align} (4.13)
    \begin{align} & w_i^n = \varphi ({x_i}), \; 0 \le i \le M, \end{align} (4.14)
    \begin{align} & w_0^n = 0, w_M^n = 0, z_0^n = 0, z_M^n = 0, \; 1 \le n \le N. \end{align} (4.15)

    Then, we have

    \begin{align} \|{w^n}\|{^2} \le \|{w^0}\|{^2}+\frac{{{1}}}{{{}^\mathcal{F} C}}(\frac{{{L^4}}}{{18{C_1}}}\mathop {\max }\limits_{1 \le n \le N} \|{p^{n-1+{\sigma}}}\|{^2}+2{C_2}\mathop {\max }\limits_{1 \le n \le N} \|{q^{n-1+{\sigma}}}\|{^2}). \end{align} (4.16)

    Proof. Similar to the proof of the Theorem 3.5, we can obtain from Eq (4.12) and Eq (4.13) that

    \begin{array}{l} {}^\mathcal{F} D_t^{\alpha}\|{w^{n+\sigma}}\|^2 \le \frac{{{L^4}}}{{18{C_1}}}\|{p^{n+{\sigma}}}\|{^2}+2{C_2}\|{q^{n+{\sigma}}}\|{^2}. \end{array}

    Noticing that

    \begin{align} {}^\mathcal{F} D_t^{\alpha}||{w^{n+\sigma}}||^2 = {}^\mathcal{F} g_n^{(n+1, \alpha)}||{w^{n+1}}||^2 - \sum\limits_{k = 1}^{n} ({{}^\mathcal{F} g_k^{(n+1, \alpha)} - {}^\mathcal{F} g_{k-1}^{(n+1, \alpha)})||{w^{k}}||^2 - {}^\mathcal{F} g_0^{(n+1, \alpha)}||{w^0}||^2}, \end{align} (4.17)

    we get

    \begin{equation} {}^\mathcal{F} g_n^{(n+1, \alpha)}\|{w^{n+1}}\|{^2}\!\leq\!\sum\limits_{k = 1}^{n} ({{}^\mathcal{F} g_k^{(n+1, \alpha)}\!-\!{}^\mathcal{F} g_{k-1}^{(n+1, \alpha)})\|{w^k}\|{^2} + {}^\mathcal{F} g_0^{(n+1, \alpha)}\|{w^0}\|{^2}}\!+\!(\frac{{{L^4}}}{{18{C_1}}}\|{p^{n+{\sigma}}}\|{^2}\!+\!2{C_2}\|{q^{n+{\sigma}}}\|{^2}). \end{equation} (4.18)

    From Lemma 4.4, we can further obtain

    \begin{equation*} {}^\mathcal{F} g_n^{(n+1, \alpha)}\|{w^{n+1}}\|{^2} \!\leq\!\sum\limits_{k = 1}^{n} ({{}^\mathcal{F} g_k^{(n+1, \alpha)}\!-\!{}^\mathcal{F} g_{k-1}^{(n+1, \alpha)})\|{w^k}\|{^2}\!+\!{}^\mathcal{F} g_0^{(n+1, \alpha)}\big[\|{w^0}\|{^2}}\!+\!\frac{{{1}}}{{{}^\mathcal{F} C}}(\frac{{{L^4}}}{{18{C_1}}}\|{p^{n+{\sigma}}}\|{^2}\!+\!2{C_2}\|{q^{n+{\sigma}}}\|{^2})\big]. \end{equation*}

    Denote

    \begin{array}{l} G = \|{w^0}\|{^2}+\frac{{{1}}}{{{}^\mathcal{F} C}}(\frac{{{L^4}}}{{18{C_1}}}\mathop {\max }\limits_{1 \le n \le N} \|{p^{n+{\sigma}}}\|{^2}+2{C_2}\mathop {\max }\limits_{1 \le n \le N} \|{q^{n+{\sigma}}}\|{^2}). \end{array}

    Now, we prove by the mathematical induction that

    \begin{align} \|{w^n}\|{^2} \le G. \end{align} (4.19)

    It holds obviously when n = 0 . Assuming Eq (4.19) is valid for n = 1, 2, \cdots, m-1 , then we have

    \begin{array}{l} {}^\mathcal{F} g_{m}^{(m+1, \alpha)}\|{w^{m+1}}\|{^2}& \leq \sum\limits_{k = 1}^{m} ({{}^\mathcal{F} g_k^{(m+1, \alpha)}-{}^\mathcal{F} g_{k-1}^{(m+1, \alpha)})\|{w^k}\|{^2}+ {}^\mathcal{F} g_0^{(m+1, \alpha)}G}\\ &\leq \sum\limits_{k = 1}^{m} ({{}^\mathcal{F} g_k^{(m+1, \alpha)}-{}^\mathcal{F} g_{k-1}^{(m+1, \alpha)})G+ {}^\mathcal{F} g_0^{(m+1, \alpha)}G} = {}^\mathcal{F} g_{m}^{(m+1, \alpha)}G. \end{array}

    This proof is completed.

    Based on Theorem 4.5, we can obtain the following stability theorems.

    Theorem 4.6 (Stability). The \mathcal{FL}2 - 1_{\sigma} scheme Eqs (4.8)–(4.11) is uniquely solvable, and unconditionally stable with respect to the initial value \varphi and the source term f .

    Theorem 4.7 (Convergence). Assume that u(x, t)\in C^{6, 3}_{x, t}([0, L]\times[0, T]) and \{ u_i^n\} are solutions of the problem (1.1)–(1.4) and the \mathcal{FL}2 - 1_{\sigma} scheme (4.8)–(4.11), respectively. Then there exists a positive constant C such that

    \begin{align} ||{e^n}|| \le C({\tau ^2}+{h^2}+\varepsilon), \; \; 0 \le n \le N. \end{align} (4.20)

    Proof. From Eq (2.7), Eq (2.8) and Eqs (4.8)–(4.11), we have the error equations as

    \begin{array}{l} &{}^\mathcal{F} D_t^{\alpha}e_i^{n+\sigma}+\delta_x^2{({\omega}{{\tilde e}^{n+{\sigma}}})_i}+{\kappa}e_i^{n+{\sigma}} = ({R_1})_i^n, \; 1 \le i \le M-1, 0 \le n \le N-1, \\ &\tilde e_i^{n+{\sigma}} = \delta_x^2e_i^{n+{\sigma}}+({R_2})_i^n, \; 1 \le i \le M-1, 0 \le n \le N-1, \\ &e_i^0 = 0, \; 0 \le i \le M, \\ &e_0^n = 0, e_M^n = 0, \tilde e_0^n = 0, \tilde e_M^n = 0, \; 1 \le n \le N. \end{array}

    Applying Theorem 4.5, we get

    \begin{array}{l} \|{e^n}\|{^2} \le \frac{{{1}}}{{{}^\mathcal{F} C}}(\frac{{{L^4}}}{{18{C_1}}}\mathop {\max }\limits_{1 \le n \le N} \|{}^\mathcal{F} R_1^n\|{^2}+2{C_2}\mathop {\max }\limits_{1 \le n \le N} \|{}^\mathcal{F} R_2^n\|{^2}), \; \; 1 \le n \le N. \end{array}

    Noticing Eq (4.7), we get

    \begin{array}{l} \|{e^n}\|{^2} \le \frac{{{1}}}{{{}^\mathcal{F} C}}(\frac{{{L^4}}}{{18{C_1}}}+2{C_2}){{}^\mathcal{F} C}^2({\tau ^2}+{h^2})^2, \; \; 1 \le n \le N, \end{array}

    which shows that Eq (4.20) is valid with C = {{}^\mathcal{F} C}\sqrt {\frac{{{1}}}{{{}^\mathcal{F} C}}(\frac{{{L^4}}}{{18{C_1}}}+2{C_2})} .

    It should be pointed out that the proposed difference schemes are based on assumptions that the solution of problem is sufficiently smooth. But the singularity of the time fractional derivative may lead to weak singularity near the initial time which may influence the accuracy of numerical method. Thus, in order to overcome the possible singularity of the solution near t = 0 , some related techniques have been developed, such as the initial correction techniques, non-uniform discretization and so on [23,24,25,26]. Because of this, a analogously scheme for the problem (1.1)–(1.4) based on the uniform mesh in space and graded mesh in time is first given as follows:

    \begin{align} & \Delta_{N}^{\alpha}u_i^n+\delta_x^2{({\omega}_i{v_i^{n}})}+{\kappa}u_i^{n} = f_i^{n}, \; 1 \le i \le M-1, 1 \le n \le N, \end{align} (5.1)
    \begin{align} & v_i^{n} = \delta_x^2u_i^{n}, \; 1 \le i \le M-1, 1 \le n \le N, \end{align} (5.2)
    \begin{align} & u_i^0 = \varphi ({x_i}), \; 0 \le i \le M, \end{align} (5.3)
    \begin{align} & u_0^n = {{\alpha}_1}({t_n}), u_M^n = {{\alpha}_2}({t_n}), v_0^n = {{\beta}_1}({t_n}), v_M^n = {{\beta}_2}({t_n}), \; 1 \le n \le N, \end{align} (5.4)

    where

    \begin{align} \Delta_{N}^{\alpha}u_i^n = \frac{d_{n, 1}}{\Gamma(2-\alpha)}u_i^n-\frac{d_{n, n}}{\Gamma(2-\alpha)}u_0^n+\frac{1}{\Gamma(2-\alpha)}\sum\limits_{k = 1}^{n-1}u_i^{n-k}(d_{n, k+1}-d_{n, k}), \end{align} (5.5)

    and

    \begin{align} d_{n, k} = \frac{(t_n-t_{n-k})^{1-\alpha}-(t_n-t_{n-k+1})^{1-\alpha}}{\tau_{n-k+1}}, \end{align} (5.6)

    with x_i = ih, t_n = (n/N)^rT, \tau_n = t_n-t_{n-1} , where the constant mesh grading exponent r \geq 1 . It should be noted that the graded mesh will be simplified to a uniform grid when r = 1 .

    In this subsection, we rely on two numerical examples to verify the availability of the proposed methods.

    Let

    \begin{array}{l} E(h, \tau) = \mathop {\max }\limits_{1 \le n \le N} ||u^n-U^n||_2, \; \; \mbox{Ord} = \log_2\Big(\frac{{{E(2h, 2\tau)}}}{{{E(h, \tau)}}}\Big). \end{array}

    Example 5.1. First, we consider the following problem

    \begin{array}{l} &\; _0^CD_t^{\alpha}u(x, t)+ \frac{{{{\partial}^2}}}{{{\partial}{x^2}}}\Big(\omega(x)\frac{{{{\partial}^2}u(x, t)}}{{{\partial}{x^2}}}\Big)+u(x, t) = f(x, t), \; \; 0 < x < 1, \ 0 < t \le 1, \\ & u(x, 0) = \cos(\pi x), \; \; 0 < x < 1, \\ & u(0, t) = t^{3+\alpha}+1, \; u(1, t) = -(t^{3+\alpha}+1), \; \; 0 \le t \le 1, \\ & \frac{{{{\partial}^2}u(0, t)}}{{{\partial}{x^2}}} = -\pi^2(t^{3+\alpha}+1), \; \frac{{{{\partial}^2}u(1, t)}}{{{\partial}{x^2}}} = \pi^2(t^{3+\alpha}+1), \; \; 0 \le t \le 1, \end{array}

    where \omega(x) = x^2+1 and f(x, t) = \cos(\pi x)\frac{{{\Gamma(4+\alpha)}}}{{{6}}}t^3+(t^{3+\alpha}+1)\big[\cos(\pi x)-2\pi ^2\cos(\pi x)+4x\pi^3\sin(\pi x)+(x^2+1)\pi^4\cos(\pi x)\big].

    It is not difficult to verify that the exact solutions of the problems 5.1 is u(x, t) = \cos(\pi x)(t^{3+\alpha}+1) , which satisfies the smoothness requirement in Theorems 3.8 and 4.7.

    The numerical accuracy of both schemes are tested with respect to \alpha = 0.25, 0.5, 0.75 , respectively. In calculation, we take \varepsilon = 10^{-13} , which is much less than \tau^{2} . The errors and convergence orders of the suggested two schemes are showed in Table 1. We can observe that the values of \mbox{Ord} are always close to 2 , which means that the \mathcal{L}2 - 1_{\sigma} scheme and the \mathcal{FL}2 - 1_{\sigma} scheme have second order accuracy both in space and time for different \alpha\in (0, 1) . Table 2 lists the convergence orders of both schemes when \tau = h and CPU time with \alpha = 0.5. Obviously, the \mathcal{FL}2 - 1_{\sigma} scheme is faster than the \mathcal{L}2 - 1_{\sigma} scheme, especially for small \tau .

    Table 1.  The errors and convergence orders for Example 5.1.
    \mathcal{FL}2 - 1_{\sigma} scheme \mathcal{L}2 - 1_{\sigma} scheme
    \alpha {h=\tau} Nexp E(h, \tau) Ord E(h, \tau) Ord
    0.25 1/10 39 1.0510e-02 1.0510e-02
    1/20 42 2.5986e-03 2.0160 2.5986e-03 2.0160
    1/40 46 6.4786e-04 2.0040 6.4786e-04 2.0040
    1/80 49 1.6185e-04 2.0010 1.6185e-04 2.0010
    1/160 53 4.0456e-05 2.0002 4.0456e-05 2.0002
    0.5 1/10 39 1.0500e-02 1.0500e-02
    1/20 42 2.5959e-03 2.0161 2.5959e-03 2.0161
    1/40 46 6.4719e-04 2.0040 6.4719e-04 2.0040
    1/80 49 1.6169e-04 2.0010 1.6169e-04 2.0010
    1/160 53 4.0416e-05 2.0002 4.0416e-05 2.0002
    0.75 1/10 39 1.0472e-02 1.0472e-02
    1/20 43 2.5911e-03 2.0149 2.5911e-03 2.0149
    1/40 46 6.4598e-04 2.0040 6.4598e-04 2.0040
    1/80 50 1.6139e-04 2.0009 1.6139e-04 2.0009
    1/160 53 4.0340e-05 2.0003 4.0340e-05 2.0003

     | Show Table
    DownLoad: CSV
    Table 2.  The errors and convergence orders for Example 5.1 when \alpha = 0.5 .
    \mathcal{FL}2 - 1_{\sigma} scheme \mathcal{L}2 - 1_{\sigma} scheme
    {h=\tau} Nexp E(h, \tau) Ord CPU(s) E(h, \tau) Ord CPU(s)
    1/250 55 1.6549e-05 4.25 1.6551e-05 49.93
    1/500 58 4.1136e-06 2.0083 17.91 4.0771e-06 2.0213 208.36
    1/1000 62 1.0772e-06 1.9331 91.50 1.0253e-06 1.9915 924.27

     | Show Table
    DownLoad: CSV

    From the Tables 1, 2, we can see that these numerical results are consistent with the previous theoretical results. It shows the \mathcal{L}2 - 1_{\sigma} scheme (2.10)–(2.13) and the \mathcal{FL}2 - 1_{\sigma} scheme (4.8)–(4.11) are convergent with second order accuracy in space and time, and the \mathcal{FL}2 - 1_{\sigma} scheme is more practical.

    Example 5.2. Now, we consider the following problem

    \begin{array}{l} &\; _0^CD_t^{\alpha}u(x, t)+ \frac{{{{\partial}^2}}}{{{\partial}{x^2}}}\Big(\omega(x)\frac{{{{\partial}^2}u(x, t)}}{{{\partial}{x^2}}}\Big) = f(x, t), \; \; 0 < x < \pi, \ 0 < t \le 1, \\ & u(x, 0) = 0, \; \; 0 < x < \pi, \\ & u(0, t) = 0, \; u(1, t) = 0, \; \; 0 \le t \le 1, \\ & \frac{{{{\partial}^2}u(0, t)}}{{{\partial}{x^2}}} = 0, \; \frac{{{{\partial}^2}u(1, t)}}{{{\partial}{x^2}}} = 0, \; \; 0 \le t \le 1, \end{array}

    where \kappa = 0, \omega(x) = e^{x} and

    \begin{array}{l} f(x, t) = (\Gamma(1+\alpha)+\frac{3\Gamma(3)t^{3-\alpha}}{\Gamma(4-\alpha)})\sin x-2e^2(t^{\alpha}+t^3)\cos x. \end{array}

    The exact solution of the example 5.2 is u(x, t) = (t^{\alpha}+t^3)\sin x.

    The error and numerical accuracy of scheme (5.1)–(5.6) are listed in Tables 35 with respect to \alpha = 0.4, 0.6, 0.8 and some values of grading exponent r , respectively. We keep M = 2N in calculation. These results show that the scheme (5.1)–(5.6) has accuracy of order \alpha when r = 1 , and accuracy of order 2-\alpha when r \geq r_c = (2-\alpha)/\alpha . The reason for this result is that the smoothness requirement of the solution in Theorems 3.8 and 4.7 is not satisfied.

    Table 3.  The errors and convergence orders for Example 5.2 when \alpha = 0.5 .
    r=1 r=r_c r=2r_c
    N E(h, \tau) Ord E(h, \tau) Ord E(h, \tau) Ord
    32 3.3961e-02 4.6082e-03 1.3810e-02
    64 2.8987e-02 2.2847e-01 1.8881e-03 1.2873 5.6845e-03 1.2806
    128 2.4345e-02 2.5178e-01 7.1277e-04 1.4054 2.1522e-03 1.4012
    256 2.0108e-02 2.7586e-01 2.5719e-04 1.4706 7.7724e-04 1.4694
    512 1.6354e-02 2.9813e-01 8.8456e-05 1.5398 2.7075e-04 1.5214

     | Show Table
    DownLoad: CSV
    Table 4.  The errors and convergence orders for Example 5.2 when \alpha = 0.6 .
    r=1 r=r_c r=2r_c
    N E(h, \tau) Ord E(h, \tau) Ord E(h, \tau) Ord
    32 2.2240e-02 6.2089e-03 1.6603e-02
    64 1.6383e-02 4.4096e-01 2.7026e-03 1.2001 7.1400e-03 1.2174
    128 1.1717e-02 4.8360e-01 1.1103e-03 1.2832 2.9149e-03 1.2925
    256 8.1872e-03 5.1716e-01 4.4202e-04 1.3288 1.1560e-03 1.3343
    512 5.6228e-03 5.4208e-01 1.7147e-04 1.3662 4.5076e-04 1.3587

     | Show Table
    DownLoad: CSV
    Table 5.  The errors and convergence orders for Example 5.2 when \alpha = 0.8 .
    r=1 r=r_c r=2r_c
    N E(h, \tau) Ord E(h, \tau) Ord E(h, \tau) Ord
    32 9.0995e-03 1.0251e-02 2.3075e-02
    64 5.6954e-03 6.7599e-01 4.8305e-03 1.0855 1.0763e-02 1.1003
    128 3.5455e-03 6.8381e-01 2.1936e-03 1.1389 4.8645e-03 1.1457
    256 2.1527e-03 7.1984e-01 9.7736e-04 1.1663 2.1620e-03 1.1699
    512 1.2788e-03 7.5136e-01 4.3093e-04 1.1814 9.5272e-04 1.1822

     | Show Table
    DownLoad: CSV

    Example 5.3. Finally, we consider the following space-time variable coefficient problem

    \begin{array}{l} &\; _0^CD_t^{\alpha}u(x, t)+ \frac{{{{\partial}^2}}}{{{\partial}{x^2}}}\Big(((xt)^2+1)\frac{{{{\partial}^2}u(x, t)}}{{{\partial}{x^2}}}\Big)+u(x, t) = f(x, t), \; \; 0 < x < 1, \ 0 < t \le 1, \\ & u(x, 0) = \cos(\pi x), \; \; 0 < x < 1, \\ & u(0, t) = t^{3+\alpha}+1, \; u(1, t) = -(t^{3+\alpha}+1), \; \; 0 \le t \le 1, \\ & \frac{{{{\partial}^2}u(0, t)}}{{{\partial}{x^2}}} = -\pi^2(t^{3+\alpha}+1), \; \frac{{{{\partial}^2}u(1, t)}}{{{\partial}{x^2}}} = \pi^2(t^{3+\alpha}+1), \; \; 0 \le t \le 1, \end{array}

    where

    f(x, t) = \cos(\pi x)\frac{{{\Gamma(4+\alpha)}}}{{{6}}}t^3+(t^{3+\alpha}+1)\big[\cos(\pi x)-2t^2\pi ^2\cos(\pi x)+4xt^2\pi^3\sin(\pi x)+(x^2t^2+1)\pi^4\cos(\pi x)\big].

    The exact solution of above problem is also u(x, t) = \cos(\pi x)(t^{3+\alpha}+1) , while the variable coefficient function \omega(x, t) = (xt)^2+1 which depends on the variables x and t .

    Similar to the spatially variable coefficient problem, we apply the \mathcal{L}2-{1_\sigma} scheme and the \mathcal{FL}2-{1_\sigma} scheme to solve the problem in Example 5.3. Table 6 presents the numerical results. In calculation, we take \varepsilon = 10^{-11} . It is shown that the \mathcal{L}2-{1_\sigma} scheme and the \mathcal{FL}2-{1_\sigma} scheme are convergent with second order accuracy in space and time.

    Table 6.  The errors and convergence orders for Example 5.3.
    \mathcal{FL}2 - 1_{\sigma} scheme \mathcal{L}2 - 1_{\sigma} scheme
    \alpha {h=\tau} Nexp E(h, \tau) Ord E(h, \tau) Ord
    0.25 1/10 33 1.2114e-02 1.2114e-02
    1/20 36 3.0215e-03 2.0033 3.0215e-03 2.0033
    1/40 39 7.5667e-04 1.9975 7.5667e-04 1.9975
    1/80 42 1.8946e-04 1.9978 1.8946e-04 1.9978
    1/160 45 4.7412e-05 1.9986 4.7412e-05 1.9986
    0.5 1/10 33 1.3545e-02 1.3545e-02
    1/20 36 3.3864e-03 1.9999 3.3864e-03 1.9999
    1/40 39 8.4892e-04 1.9961 8.4892e-04 1.9961
    1/80 42 2.1266e-04 1.9971 2.1266e-04 1.9971
    1/160 45 5.3229e-05 1.9983 5.3229e-05 1.9983
    0.75 1/10 33 1.4683e-02 1.4683e-02
    1/20 36 3.6621e-03 2.0034 3.6621e-03 2.0034
    1/40 39 9.1663e-04 1.9983 9.1663e-04 1.9983
    1/80 42 2.2943e-04 1.9983 2.2943e-04 1.9983
    1/160 45 5.7401e-05 1.9989 5.7401e-05 1.9989

     | Show Table
    DownLoad: CSV

    In this paper, we propose two second order difference schemes in both space and time for solving the variable coefficient fourth-order fractional sub-diffusion equation subject to the second Dirichlet boundary conditions. The \mathcal{L}2-{1_\sigma} formula and \mathcal{FL}2-{1_\sigma} formula are applied to approximation the time Caputo fractional derivative. Compared with \mathcal{L}2-{1_\sigma} scheme, the \mathcal{FL}2-{1_\sigma} scheme shows the better computational efficiency. The unconditional stability, solvability and convergence of the two schemes are strictly proved by the discrete energy method. The nonuniform L_1 approximation for the such problem is also given. Numerical examples are given to verify the effectiveness of both schemes. It should be pointed out that the results in this paper can be directly extended to time-space variable coefficient problems if we constrain the coefficient function \omega(w, t) satisfying that 0 < C_1\le \omega(w, t) \le C_2 .

    This work described in this paper was supported by the Sichuan Science and Technology Program (Grant No. 2020YJ0110, Grant No. 2022JDTD0019), the National Natural Science Foundation of China (Grant No. 11801389) and the Laurent Mathematics Center of Sichuan Normal University and National-Local Joint Engineering Laboratory of System Credibility Automatic Verification (Grant No. ZD20220105).

    The authors declare there is no conflict of interest.



    [1] C. Wolf, A mathematical model for the propagation of a hantavirus in structured populations, Discrete Continuous Dyn. Syst. Ser. B, 4 (2004), 1065-1089. doi: 10.3934/dcdsb.2004.4.1065
    [2] J. Wu, R. Jafari, Orientation Independent Activity/Gesture Recognition Using Wearable Motion Sensors, IEEE Internet Things J., 6 (2018), 1427-1437.
    [3] P. K. Pisharady, M. Saerbeck, Recent methods and databases in vision-based hand gesture recognition: A review, Comput. Vis. Image Underst., 141 (2015), 152-165. doi: 10.1016/j.cviu.2015.08.004
    [4] H. S. Hasan, S. Kareem, Human computer interaction for vision based hand gesture recognition: A survey, Artif. Intell. Rev., 43 (2015), 1-54. doi: 10.1007/s10462-012-9356-9
    [5] H. I. Lin, Hsien-I., M. H. Hsu, W.-K. Chen, Human hand gesture recognition using a convolution neural network, In IEEE Int. Conf. Autom. Sci. Eng. (CASE), IEEE, 2014, 1038-1043.
    [6] O. K. Oyedotun, A. Khashman, Deep learning in vision-based static hand gesture recognition, Neural Comput. Appl., 28 (2017), 3941-3951. doi: 10.1007/s00521-016-2294-8
    [7] Z. Lu, X. Chen, Q. Li, X. Zhang, P. Zhou, A Hand Gesture Recognition Framework and Wearable Gesture-Based Interaction Prototype for Mobile Devices, IEEE Trans. Hum. Mach. Syst., 44 (2014), 293-299. doi: 10.1109/THMS.2014.2302794
    [8] Y. Huang, L. Gao, Y. Zhao, X. Guo, C. Liu, P. Liu, Highly flexible fabric strain sensor based on graphene nanoplatelet-polyaniline nanocomposites for human gesture recognition, J. Appl. Polymer Sci., 134 (2017), 45340. doi: 10.1002/app.45340
    [9] M. Panwar, Hand gesture recognition based on shape parameters, Int. Conf. Comput., Commun. Appl., Dindigul, Tamilnadu, 2012, 1-6.
    [10] C. Weng, Y. Li, M. Zhang, K. Guo, X. Tang, Z. Pan, Robust Hand Posture Recognition Integrating Multi-cue Hand Tracking. In Int. Conf. Technol. E-learn. Digital Entertain., Springer, Berlin, Heidelberg, 497-508.
    [11] D. H. Kim, J. Lee, H. S. Yoon, J. Kim, J. Sohn, Vision-based arm gesture recognition for a long-range human-robot interaction, J. Supercomput., 65 (2013), 336-352. doi: 10.1007/s11227-010-0541-9
    [12] J. Li, H. Huai, J. Gao, D. Kong, L. Wang, Spatial-temporal dynamic hand gesture recognition via hybrid deep learning model, J. Multimodal User Interfaces, 13 (2019), 363-371. doi: 10.1007/s12193-019-00304-z
    [13] Z. Lu, S. Qin, X. Li, L. Li, D. Zhang, One-shot learning hand gesture recognition based on modified 3d convolutional neural networks, Mach. Vis. Appl., 30 (2019), 1157-1180. doi: 10.1007/s00138-019-01043-7
    [14] A. Sarkar, A. Gepperth, U. Handmann, T. Kopinski, Dynamic Hand Gesture Recognition for Mobile Systems Using Deep LSTM, Int. Conference Intell. Hum. Comput. Interact., Springer, Cham, 2017.
    [15] T. Gonzalez-Sanchez, D. Puig, Real-time body gesture recognition using depth camera, Electron. Lett., 47 (2011), 697-698. doi: 10.1049/el.2011.0967
    [16] S. Diego, F. Bruno, B. Byron, HAGR-D: A Novel Approach for Gesture Recognition with Depth Maps, Sensors, 15 (2015), 28646-28664. doi: 10.3390/s151128646
    [17] B. Fang, F. Sun, H. Liu, 3D human gesture capturing and recognition by the IMMU-based data glove, Neurocomputing, 277 (2018), 198-207. doi: 10.1016/j.neucom.2017.02.101
    [18] P. Rouanet, P. Y. Oudeyer, F. Danieau, D. Filliat, The Impact of Human-Robot Interfaces on the Learning of Visual Objects, IEEE Trans. Robot., 29 (2013), 525-541. doi: 10.1109/TRO.2012.2228134
    [19] J. Zhu, L. G. Blumberg, Y. Zhu, M. Nisser, E. L. Carlson, X. Wen, et al., CurveBoards: Integrating Breadboards into Physical Objects to Prototype Function in the Context of Form, In Proc. CHI Conf. Hum. Factors Comput. Syst., 2020, 1-13.
    [20] D. Y. Lee, S. H. Lee, I. Oakley, Nailz: Sensing Hand Input with Touch Sensitive Nails, In Proc. CHI Conf. Hum. Factors Comput. Syst., 2020, 1-13.
    [21] A. Akl, C. Feng, S. Valaee, A Novel Accelerometer-Based Gesture Recognition System, IEEE Trans. Signal Process., 59 (2011), 6197-6205. doi: 10.1109/TSP.2011.2165707
    [22] M. Lee, K. Hwang, J. Park, S. Choi, S. Shin, W. Sung, FPGA-based low-power speech recognition with recurrent neural networks, IEEE Int. Workshop Signal Process. Syst. (SiPS), 2016,230-235.
    [23] V. Rybalkin, A. Pappalardo, M. M. Ghaffar, G. Gambardella, N. Wehn, M. Blott, FINN-L: Library extensions and design trade-off analysis for variable precision LSTM networks on FPGAs. In 28th Int. Conf. Field Program. Logic Appl. (FPL). IEEE, 2018, 89-897.
    [24] R Xie, J Cao, Accelerometer-Based Hand Gesture Recognition by Neural Network and Similarity Matching, IEEE Sensors J., 16 (2016), 4537-4545. doi: 10.1109/JSEN.2016.2546942
    [25] F. Ge, N. Wu, H. Xiao, Y. Zhang, F. Zhou, Compact Convolutional Neural Network Accelerator for IoT Endpoint SoC, Electronics, 8 (2019), 497. doi: 10.3390/electronics8050497
    [26] M. Roukhami, M. T. Lazarescu, F. Gregoretti, Y. Lahbib, A. Mami, Very Low Power Neural Network FPGA Accelerators for Tag-Less Remote Person Identification Using Capacitive Sensors, IEEE Access, 7 (2019), 102217-102231. doi: 10.1109/ACCESS.2019.2931392
    [27] J Chauhan, S Seneviratne, Y Hu, A. Misra, Breathing-Based Authentication on Resource-Constrained IoT Devices using Recurrent Neural Networks, Computer, 51 (2018), 60-67. doi: 10.1109/MC.2018.2381119
    [28] S. Okada, S. Ishibashi, T. Nishida, On-Line Unsupervised Segmentation for Multidimensional Time-Series Data and Application to Spatiotemporal Gesture data, In Int. Conference Ind. Eng. Other Appl. Applied Intell. Syst., IEA/AIE 2010, Proceedings, Part I. DBLP, 2010.
    [29] O. Dehzangi, V. Sahu, IMU-Based Robust Human Activity Recognition using Feature Analysis, Extraction, and Reduction, 24th Int. Conf. Pattern Recognition (ICPR), IEEE, 2018, 1402-1407.
    [30] J. H. Kim, G. S. Hong, B. G. Kim, D. P. Dogra, deepGesture: Deep learning-based gesture recognition scheme using motion sensors, Displays, 55 (2018), 38-45. doi: 10.1016/j.displa.2018.08.001
    [31] D. Jeong, B.-G. Kim, S.-Y. Dong, Deep Joint Spatiotemporal Network (DJSTN) for Efficient Facial Expression Recognition, Sensors, 20 (2020), 1936. doi: 10.3390/s20071936
    [32] L. García-Hernández, M. Pérez-Ortiz, A. Araúzo-Azofra, L. Salas-Morera, C. Hervás-Martínez, An evolutionary neural system for incorporating expert knowledge into the UA-FLP, Neurocomputing, 135 (2014), 69-78. doi: 10.1016/j.neucom.2013.01.068
    [33] E. Keogh, S. Chu, D. Hart, M. Pazzani, An online algorithm for segmenting time series, Proc. 2001 IEEE Int. Conf. Data Min., IEEE, 2001,289-296.
    [34] S. Mallat, Wavelets for a vision, Proc. IEEE, 84 (1996), 604-614. doi: 10.1109/5.488702
    [35] L. Perez, J. Wang, The Effectiveness of Data Augmentation in Image Classification using Deep Learning, arXiv preprint arXiv: 1712.04621.
    [36] J. Sola, J. Sevilla, Importance of input data normalization for the application of neural networks to complex industrial problems, IEEE Trans. Nucl. Sci., 44 (1997), 1464-1468. doi: 10.1109/23.589532
    [37] J. Han, J. Pei, M. Kamber, Data mining: concepts and techniques. Elsevier, 2011.
    [38] S. T. Chakradhar, M. Sankaradass, V. Jakkula, S. Cadambi, A dynamically configurable coprocessor for convolutional neural networks, 37th Int. Symp. Comput. Archit. (ISCA 2010), ACM, 2010.
    [39] T. Chen, Z. Du, N. Sun, J. Wang, C. Wu, Y. Chen, et al., DianNao: A Small-Footprint High-Throughput Accelerator for Ubiquitous Machine-Learning, ACM SIGARCH Comput. Architect. News, 42 (2014), 269-284. doi: 10.1145/2654822.2541967
    [40] P. Sayyah, T. L. Mihai, S. Bocchio, E. Ebeid, G. Palermo, D. Quaglia, et al., Virtual Platform-Based Design Space Exploration of Power-Efficient Distributed Embedded Applications, ACM Trans. Embed. Comput. Syst., (TECS), 14 (2015), 1-25.
    [41] J. G. Proakis, D. G. Manolakis, Digital signal processing: principles, algorithms, and applications, 1996.
    [42] U. Meyer-Baese, Digital Signal Processing with Field Programmable Gate Arrays, Springer, 2007.
    [43] J. Cho, S. Mirzaei, J. Oberg, R. Kastner, FPGA-Based Face Detection System Using Haar Classifiers, In Proc. ACM/SIGDA Int. Symp. Field Program. Gate arrays, 103-112.
    [44] J. Hegarty, J. Brunhaver, Z. Devito, J. Ragan-Kelley, N.Cohen, S. Bell, et al., Darkroom: Compiling high-level image processing code into hardware pipelines, ACM Trans. Graph., 33 (2014), 1-11.
    [45] J. Kim, B. Kim, P. P. Roy, D. Jeong, Efficient facial expression recognition algorithm based on hierarchical deep neural network structure, IEEE Access, 7 (2019), 41273-41285. doi: 10.1109/ACCESS.2019.2907327
    [46] H. Akima, A New Method of Interpolation and Smooth Curve Fitting Based on Local Procedures, J. ACM, 17 (1970), 589-602. doi: 10.1145/321607.321609
    [47] M. Shi, A. Bermak, S. Chandrasekaran, A. Amira, S. Brahim-Belhouari. A Committee Machine Gas Identification System Based on Dynamically Reconfigurable FPGA, IEEE Sensors J., 8 (2008), 403-414. doi: 10.1109/JSEN.2008.917124
    [48] M. Hamouda, H. F. Blanchette, K. Al-Haddad, F. Fnaiech. An Efficient DSP-FPGA-Based Real-Time Implementation Method of SVM Algorithms for an Indirect Matrix Converter, IEEE Trans. Ind. Electron., 58 (2011), 5024-5031. doi: 10.1109/TIE.2011.2159952
    [49] N. Attaran, A. Puranik, J. Brooks, T. Mohsenin, Embedded Low-Power Processor for Personalized Stress Detection, IEEE Trans. Circuits Syst. Ⅱ-Express Briefs, 65 (2018), 2032-2036. doi: 10.1109/TCSII.2018.2799821
    [50] Z. Li, L. Wang, S. Guo, Y. Deng, Q. Dou, H. Zhou, et al., Laius: An 8-bit fixed-point CNN hardware inference engine, In Proc. IEEE Int. Symp. Parallel Distrib. Process. Appl. and 2017 IEEE Int. Conf. Ubiquitous Computi. Commun. (ISPA/IUCC), 143-150.
    [51] Y. Zhang, N. Wu, F. Zhou, M. R. Yahya, Design of Multifunctional Convolutional Neural Network Accelerator for IoT Endpoint SoC, In Proc. World Congress Eng. Comput. Sci. 2018, 16-19.
    [52] C. M. Morales, U. Flores, M. A. Medina, M. D. Saazar, J. A. Caballero, D. C. Cruz, et al., Digital Artificial Neural Network Implementation on a FPGA for data classification, IEEE Latin Am. Trans., 13 (2015), 3216-3220. doi: 10.1109/TLA.2015.7387224
  • This article has been cited by:

    1. Yang Cao, Zhijun Tan, A fast and high-order localized meshless method for fourth-order time-fractional diffusion equations, 2025, 142, 10075704, 108586, 10.1016/j.cnsns.2024.108586
  • Reader Comments
  • © 2021 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(8046) PDF downloads(419) Cited by(21)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog