Mechanical arm optimization control method and device based on coupled physical neural network

Abstract

The invention relates to a mechanical arm optimization control method and device based on a coupled physical neural network, and belongs to the field of mechanical arm optimization control. The method comprises the following steps that S1, a mechanical arm optimization problem is extracted; s2, constructing accurate solution data; s3, establishing a constraint condition network and training; s4, establishing a target function network; s5, taking the objective function as a loss function, and establishing a dual-network coupling architecture; s6, training a dual-network coupling architecture; s7, outputting a predicted optimal solution result; s8, performing deviation correction on the predicted optimal solution by using a deviation correction method to obtain an optimal solution; and S9, mechanical arm parameters are adjusted according to the optimal solution to achieve control. According to the method, a multi-task objective function can be quickly coped with, so that a large number of repeated solving data sets are avoided, and the target required by a product is responded more timely and accurately; meanwhile, the innovative training method not only ensures the global search capability, but also effectively balances the calculation cost and the solving precision, and provides efficient technical support for solving a complex optimization problem.

Description

TECHNICAL FIELD

[0001] The present application relates to a mechanical arm optimization control method and device based on a coupled physical neural network, belonging to the field of deep learning and mechanical arm optimization control, and is especially suitable for the coordinated control of a mechanical arm based on a coupled physical neural network. BACKGROUND

[0002] With the wide application of mechanical arms in industrial production, the demand for their functional diversification (such as high-precision assembly, rapid handling, flexible grasping, etc.) is also increasing. Mechanical arms need to respond quickly and accurately to a variety of complex task objectives (such as end trajectory precision, joint torque smoothness, energy minimization, obstacle avoidance, etc.) under different working conditions, and these objectives are often coupled or even conflicting. This makes the multi-task, nonlinear motion optimization control problem a frontier challenge for the intelligent development of mechanical arms.

[0003] Existing global optimization methods have significant limitations in dealing with such complex engineering problems: deterministic methods are usually only suitable for simplified problems with obvious convexity or monotonicity characteristics, and are difficult to handle the highly nonlinear dynamics model of the mechanical arm; random methods such as degenerate algorithms and genetic algorithms cannot capture the complete information of continuous intervals because they are limited to discrete point solutions.

[0004] Neural network methods provide an effective way for global search of motion optimization control problems. Raissi et al. proposed a physical information neural network (PINN) that can embed the dynamics equation of the mechanical arm into the loss function, significantly reducing the dependence on sample data. Given the dual advantages of physical information neural networks in accurately fitting sample points and smoothing the solution space constraints, the present application chooses to explore this technical route in depth to solve the high-efficiency and high-precision solution of motion optimization control problems. However, when dealing with complex optimization problems such as multi-task objectives and nonlinear systems, there are still three major challenges: (1) computational efficiency bottleneck: due to the specificity of the task objectives, existing methods need to be trained separately for each task, resulting in a large consumption of computational resources. (2) global optimization problem: how to ensure the optimal solution of the motion optimization control problem on a global level. (3) model training efficiency problem: the training of neural network models relies on a large amount of data, and in order to express the motion optimization control problem with high precision, data needs to be selected strategically.

[0005] In summary, in the application of mechanical arm control systems, it is often necessary to perform fast and efficient optimization solution under the same physical constraint conditions for different task requirements. Especially with the increase in the degrees of freedom of the mechanical arm and the consideration of more refined control parameters, the dimension and complexity of the optimization problem rise sharply, resulting in extremely low solution efficiency of existing methods. SUMMARY

[0006] Therefore, in order to realize the optimization control problem of the robot arm in the multi-task scene with high efficiency and high precision, the application provides a robot arm optimization control method and device based on a coupled physical neural network, and a double-network coupling architecture is provided: through the cooperative work of two physical information neural networks, the constraint conditions and the objective function of the optimization problem are respectively modeled and solved, so that the accurate control of the robot arm in the changing multi-task target scene can be quickly responded, and a large amount of repeated data sets are avoided. Meanwhile, in order to further improve the training efficiency and precision, a physical information neural network training method based on a genetic algorithm for the double-network coupling architecture is provided, which not only ensures the global search ability, but also effectively balances the calculation cost and the solving precision, and provides efficient technical support for solving the complex robot arm optimization control problem.

[0007] To achieve the above purpose, the application provides the following technical scheme:

[0008] The robot arm optimization control method based on the coupled physical neural network combines Figure 1 , and comprises the following steps:

[0009] S1: performing dynamic analysis on the robot arm to obtain a constraint condition equation, and establishing an objective function according to a control strategy of the robot arm, and then extracting an optimization problem of the robot arm;

[0010] S2: constructing an accurate solution data set corresponding to the constraint condition equation according to the dynamic characteristics of the robot arm;

[0011] S3: establishing a constraint condition network by using a physical information neural network, and training the constraint condition network by using the accurate solution data set;

[0012] S4: establishing an objective function network by using a deep learning network;

[0013] S5: taking the objective function as a loss function, and embedding the constraint condition network into the objective function network to establish a double-network coupling architecture;

[0014] S6: training the double-network coupling architecture by using a random training set;

[0015] S7: outputting a predicted optimal solution result by using the trained objective function network;

[0016] S8: performing a bias correction method on the predicted optimal solution result to obtain a high-precision optimal solution result;

[0017] S9: setting the robot arm to adjust parameters according to the high-precision optimal solution result to realize a control strategy.

[0018] Further, the optimization problem is in the form of:

[0019]

[0020] Among them, the independent variable parameter The dependent variable x = x(t,p) = (x1(t,p),…,x n (t,p)), dynamic equation Robotic arm interval process control strategy robotic arm end-effector control strategy obj corresponds to the objective function equation, st corresponds to the constraint equation; m is the parameter dimension, a positive integer; n is the dependent variable dimension, a positive integer; t0, t f p represents the initial and final times. L p U These are the lower and upper bounds of the parameter p, respectively.

[0021] Preferably, in step S2, the construction of the exact solution dataset corresponding to the constraint equation is carried out by manually fixing the parameter p and measuring the dependent variable x of the robotic arm.

[0022] Preferably, the exact solution dataset corresponding to the constraint equation described in step S2 is obtained by solving the dependent variable x of the dynamic equation Φ using a randomly fixed parameter p.

[0023] Furthermore, the constraint network is composed of a deep learning network connected in series with constraint equations; the objective function network is composed of a deep learning network connected in series with a parameter normalization equation; the dual-network coupling architecture is composed of an objective function network connected in series with a constraint network, and then connected in series with an objective function equation.

[0024] Furthermore, combined with Figure 2 The training process of the dual-network coupled architecture described in step S6 is as follows:

[0025] S601: A set of N randomly generated m-dimensional sample points Y;

[0026] S602: Input the sample point set Y into the deep learning network of the objective function network to obtain the predicted value. Where α is the deep learning network of the objective function network. The parameters to be trained;

[0027] S603: Using parameter normalization equations For predicted values The process is performed to obtain the parameter value p;

[0028] S604: Input the parameter value p and the independent variable t into the constraint network to obtain the predicted value of the dependent variable. wherein θ is a constraint condition network to be trained;

[0029] S605: randomly selecting a slack variable in the interval [-γ, γ] obtaining a dependent variable value by relaxing the dependent variable wherein γ = max{|E(ε)-3·D(ε)|,|E(ε)+3·D(ε)|}, and ε is a residual error of the constraint condition network

[0030] S606: taking the independent variable t, the parameter value p and the dependent variable value x as inputs, and calculating a loss value Loss = J(p) of the loss function as the objective function;

[0031] S607: adjusting the parameter α of the deep learning network of the objective function network by gradient descent of the loss value;

[0032] S608: repeating steps S602-S607 until the loss value meets the threshold value to complete the training.

[0033] Preferably, the method for training the constraint condition network in step S3 is a physical information neural network training method based on a genetic algorithm, which is specifically:

[0034] S301: initializing a population: randomly generating an initial population containing M individuals from the parameter interval [p L , p U ];

[0035] S302: traversing the parameter p selected from replacing the parameter p in Φ, and solving the corresponding dependent variable value x at the time of the independent variable t by using a numerical solution method * ;

[0036] S303: storing in the exact solution training data set;

[0037] S304: training the PINN network using the exact solution training data set;

[0038] S305: if , then generating new parameter individuals by using selection, crossover and mutation operations and repeating steps S302-S305 until and outputting the trained network or reaching the maximum number of iterations and reporting an error; wherein α is an absolute error threshold.​

[0039] Further, the selection operation is to select all individuals corresponding to the residual high sampling area as new parameter individuals of the next generation directly, and the parameter individuals do not need to be solved again. The parameter individuals are solved again. * When the number of individuals does not reach M after the selection operation, the crossover operation and the mutation operation are performed in proportion; the crossover operation is to perform the crossover operation on the selected individuals to generate new individuals with a certain crossover probability; and the mutation operation is to generate new individuals in the parameter interval [p L , p U ] with a certain mutation probability.

[0040] Specifically, the strategy realizes automatic identification of the residual high sampling area and weight optimization of the key parameters through the selection, crossover and mutation operations, effectively avoids unordered expansion of the differential algebraic equation data set while ensuring the prediction accuracy of the constraint condition network, and improves the training efficiency of the random optimization.

[0041] Preferably, all elements in gamma can take alpha for the convenience of calculation, and alpha is greater than gamma in the probability sense, which can ensure the equivalence of the solution.

[0042] Preferably, steps S1 to S4 can be used to build and train the constraint condition network and the target function network offline, and only the operations of steps S5 to S8 need to be adjusted according to the target function online, which can greatly improve the calculation efficiency.

[0043] Preferably, the bias correction method is the Newton iteration method or the random walk method.

[0044] The application further discloses an electronic device, including at least one processor; and a memory connected with the at least one processor in communication; wherein,

[0045] The memory stores a computer program executed by the at least one processor,

[0046] The computer program is executed by the at least one processor to enable the at least one processor to execute the mechanical arm optimization control method based on the coupled physical neural network.

[0047] Finally, the application further discloses a computer readable storage medium, which stores computer instructions for enabling a processor to execute the mechanical arm optimization control method based on the coupled physical neural network.

[0048] ​The equivalence of the solution of the double coupling network architecture of the application and the solution of the optimization problem: when the approximation precision of the PINN network is high enough, the corresponding residual is close to the zero vector, then E(∈)→0 and D(∈)→0. Therefore, after introducing the relaxation variable γ, solving formula (1.1) and solving the optimization problem are equivalent.

[0049]

[0050] The application provides a mechanical arm optimization control method and device based on a coupled physical neural network, and a double network coupling architecture is innovatively proposed: through the cooperative work of two physical information neural networks, the constraint conditions and the objective function of the optimization problem are respectively modeled and solved, the constraint condition network only needs to be trained once to generalize and call the decoupling expression of the nonlinear system, and the objective function network introduces a relaxation variable of the upper limit of the global error, which theoretically guarantees the equivalence of the network architecture and the solution of the optimization problem. The application can quickly cope with multi-task objective functions to avoid a large amount of repeated solving data sets, so that the target of responding to product demand is more timely and accurate. In order to further improve the training efficiency and precision, the application also proposes a physical information neural network training method based on a genetic algorithm, which not only ensures the global search ability, but also effectively balances the calculation cost and the solving precision, and provides efficient technical support for the solution of complex optimization problems. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to make the purpose, technical scheme of the application more clear, the application provides the following drawings for description:

[0052] Figure 1 The method flowchart of the application;

[0053] Figure 2 The double coupling network architecture diagram of the method of the application;

[0054] Figure 3 The structural schematic diagram of the electronic equipment in embodiment 3 of the application. DETAILED DESCRIPTION

[0055] Embodiment 1: Refer to the granted invention patent "A position and force hybrid control method of a redundancy robot arm" (publication number CN112828894A). The position and force of the redundancy robot arm are controlled, the rotation angle of each joint in the robot arm is controlled by the driving motor, and the input signal of the driving motor and the rotation angle of the output joint are related to the PID parameters and the feedback adjustment amplification parameters, etc. The vector composed of all parameters is denoted as p, and the vector composed of all joint rotation angle variables is θ(t, p). In order to realize efficient and accurate control of the redundancy robot arm, the present application provides a "robot arm optimization control method based on coupled physical neural network", specifically, comprising the following steps:

[0056] Step one, the constraint condition equation is obtained by analyzing the dynamics of the redundancy robot arm, and the objective function is established according to the control strategy of the robot arm, and then the optimization problem of the redundancy robot arm is extracted.

[0057] Among them, the kinematics equation of the redundancy robot arm is obtained by means of neural dynamics method And the nonlinear equation of the joint angle of the redundancy robot arm F(t, θ(t, p), p) = r(t, p) describes the force error and position error when the end effector of the redundancy robot arm contacts with the environment in the form of explicit differential equation, which is specifically described as: Wherein, e = [e1; e2], is the derivative of e with respect to time, is the joint angle of the redundancy robot arm, is the joint angular velocity of the redundancy robot arm, m is the degree of freedom of the robot arm, is the expected end effector velocity, is the Jacobian matrix of the redundancy robot arm, Diagonal matrix Δ = diag(0, 0, 1); is the position of the end effector of the redundancy robot arm; is the rotation matrix of the redundancy robot arm; is the stiffness coefficient of the end effector of the redundancy robot arm; the superscript T is the transpose operation of the matrix.

[0058] The normal vector of the environment contact surface in the end effector coordinate system can be expressed as Wherein, is the normal vector of the environment contact surface in the base coordinate system; considering that the end effector attitude needs to be kept in the direction of the environment contact surface normal vector, the formula for keeping the direction of the end effector kinematics relationship is Wherein, is the vector the Jacobian matrix of θ.

[0059] According to the demand, the minimum performance index is set as the objective function The constraint equation Φ is And Wherein, Λ is a non-zero coefficient matrix, and And ω is the joint angular velocity feasible region of the redundant manipulator determined by the objective to be optimized.

[0060] In summary, the optimization problem of the redundant manipulator is:

[0061]

[0062] Step two, according to the dynamic characteristics of the manipulator, a random fixed parameter p is adopted to measure the dependent variable x of the manipulator, and the accurate solution data set corresponding to the constraint condition equation is constructed.

[0063] Step three, the constraint condition network is established by using the physical information neural network, and the accurate solution data set is used for training. The constraint condition network is a deep learning network in series with the constraint condition equation; specifically:

[0064] S301: initialize the population: randomly generate an initial population containing M individuals from the parameter interval [p L , p U ]

[0065] S302: traverse the parameter selected from Replace the parameter p in Φ, and solve the dependent variable value x corresponding to the independent variable t at this time by using the numerical solution method * ;

[0066] S303: store in the accurate solution training data set;

[0067] S304: train the PINN network using the accurate solution training data set;

[0068] S305: if , then generate new parameter individuals by using selection operation, crossover operation and mutation operation and repeat steps S302-S305 until and output the trained network or reach the maximum number of iterations and report an error; wherein α is the absolute error threshold.

[0069] Further, the selection operation is to select corresponding to All individuals selected are directly used as new parameter individuals for the next generation; these parameter individuals do not need to be recalculated numerically for the dependent variable x. * After the selection operation is completed, if the number of individuals does not reach M, then crossover and mutation operations are performed proportionally. The crossover operation involves pairing the selected individuals and performing a crossover operation with a certain probability to generate new individuals. The mutation operation involves performing a mutation operation with a certain probability within the parameter range [p]. L ,p U New individuals are generated within the [ ].

[0070] Specifically, this strategy achieves automatic identification of high residual sampling regions and optimization of key parameter weights through selection, crossover, and mutation operations. While ensuring the prediction accuracy of the constrained network, it effectively avoids the unordered expansion of the differential algebraic equation dataset and improves the training efficiency of stochastic optimization.

[0071] Step 4: Establish the objective function network using a deep learning network. The objective function network consists of a deep learning network connected in series with a parameter normalization equation.

[0072] Step 5: Using the objective function as the loss function, embed the constraint network into the objective function network to establish a dual-network coupled architecture. This dual-network coupled architecture consists of the objective function network connected in series with the constraint network, and then the objective function equation connected in series.

[0073] Step Six: Train the dual-network coupled architecture using a random training set. Specifically,

[0074] S601: A set Y of N randomly generated m-dimensional sample points;

[0075] S602: Input the sample point set Y into the deep learning network of the objective function network to obtain the predicted value. Where α is the deep learning network of the objective function network. The parameters to be trained;

[0076] S603: Using parameter normalization equations For predicted values The parameter value p is obtained by processing it.

[0077] S604: Input the parameter value p and the independent variable t into the constraint network to obtain the predicted value of the dependent variable. Where θ is the constraint network The parameters to be trained;

[0078] S605: Randomly select slack variables within the interval [-γ, γ]. The dependent variable value is obtained by relaxing the dependent variable. wherein, γ=max{|E(∈)-3·D(∈)|,|E(∈)+3·D(∈)|}, ∈ is a constraint condition network the residual of the constraint condition network;

[0079] S606: taking the independent variable t, the parameter value p and the dependent variable value x as inputs, calculating the loss value Loss=J(p) of the loss function of the objective function;

[0080] S607: adjusting the parameter α of the deep learning network of the objective function network by gradient descent of the loss value;

[0081] S608: repeating steps S602-S607 until the loss value meets the threshold value to complete the training.

[0082] In order to facilitate calculation, all elements in γ can take α, wherein α is greater than γ in a probabilistic sense, which can ensure the equivalence of the solution.

[0083] Step seven, outputting a predicted optimal solution result by using the trained objective function network.

[0084] Step eight, using a random walk method as a bias correction method to correct the predicted optimal solution result to obtain a high-precision optimal solution result.

[0085] Steps one to four can be used to build and train the constraint condition network offline, and the objective function network is built; online only needs to adjust the operation of steps five to eight according to the objective function; this can greatly improve the calculation efficiency.

[0086] Step nine, setting the redundancy of the mechanical arm according to the high-precision optimal solution result to adjust the parameter p to realize the hybrid control strategy of force and position.

[0087] Embodiment 2: In order to illustrate the change of the calculation time of the method of the application with the dimension, the beneficial effects of the method of the application are shown. In combination with the literature [1], the optimization solving problem of the end position control of the dynamics system of the first-order linear mechanical arm system is considered as follows:

[0088]

[0089] wherein, I is an n-dimensional unit diagonal matrix, p=(p1, p2, …, p n ) T is an n-dimensional parameter vector, t∈[0,1].

[0090] When the end position corresponding to the objective function is variable, the optimization system is efficiently solved multiple times. The application provides a "mechanical arm optimization control method based on coupled physical neural network", which specifically comprises the following steps:

[0091] Step one, the user inputs the optimization problem to be solved of the end position control of the dynamics system of the first-order linear mechanical arm system of different scales, including constraint condition equations and objective function equations; the electronic device uses lexical analysis and semantic analysis techniques to extract independent variables and dependent variables.

[0092] Step two, the constraint condition is solved by using the ode45 solver of Matlab to obtain numerical solutions, and the numerical solutions are used to construct an accurate solution data set.

[0093] Step three, a constraint condition network is established by using a physical information neural network, and the network is trained by using the accurate solution data set. Due to the simplicity of the system, the network expression effect is good, and the training method based on the genetic algorithm of the physical information neural network is not used for training. The traditional training method is directly used.

[0094] Since A is reversible and the eigenvalues are all negative, the constraint condition system is stable. With the increase of the dimension n of the equation, the dimension of the dependent variable and the parameter will also increase accordingly. To achieve high-precision expression, the training set also needs to be adjusted with the dimension. Specifically, the initial condition training data set N I =500·n, the accurate solution training data set N B =1000·(n-1), the domain training data set N F =20000, and the target function network random sample point is 500. Further, the constraint condition network is set to a DNN network with 6 layers in the middle layer and a size of 128.

[0095] Step four, a deep learning network is used to establish a target function network.

[0096] The target function network is a DNN network with 1 layer in the middle layer and a size of 8 connected in series with a parameter normalization equation.

[0097] Step five, combining Figure 2 , the target function is used as a loss function, and the constraint condition network is embedded into the target function network to establish a double network coupling architecture.

[0098] Specifically, the loss function is

[0099] Step six, the double network coupling architecture is trained by using a random training set. The specific process is:

[0100] The training process of the double network coupling architecture of step S6 is:

[0101] S601: a set of N m-dimensional sample points Y is randomly generated;

[0102] S602: input the sample point set Y into the deep learning network of the objective function network to obtain a predicted value wherein, a is the to-be-trained parameter of the deep learning network of the objective function network .

[0103] S603: utilize the parameter normalization equation to process the predicted value , and obtain a parameter value p.

[0104] S604: input the parameter value p and the independent variable t into the constraint condition network to obtain a dependent variable predicted value wherein, θ is the to-be-trained parameter of the constraint condition network .

[0105] S605: randomly select a slack variable in the interval [-γ, γ] to relax the dependent variable and obtain a dependent variable value wherein, in order to facilitate calculation, all elements in γ can take a, and a = 1e-4.

[0106] S606: input the independent variable t, the parameter value p and the dependent variable value x as inputs, and calculate the loss value Loss = J(p) of the loss function as the loss function;

[0107] S607: utilize the gradient descent of the loss value to adjust the parameter a of the deep learning network of the objective function network.

[0108] S608: repeat steps S602 to S607 until the loss value meets the threshold value to complete the training.

[0109] Step seven, utilize the trained objective function network to output a predicted optimal solution result.

[0110] Step eight, utilize the Newton iteration method to correct the predicted optimal solution result to obtain a high-precision optimal solution result.

[0111] The comparative experiment was performed on a desktop computer with CPU of Intel(R) Core(TM) i7-14700KF, 3.40GHz and memory of 32GB DDR4, and the time consumption average value (unit: seconds) of 10 repeated experiments was taken, and the specific experimental results are shown in Table 4.

[0112] Table 4 Performance comparison experimental results of example 2

[0113]

[0114]

[0115] [1] Bajaj, Ishan, Faruque HM. Global dynamic optimization using edge-concave under estimator [J / OL]. Journal of Global Optimization, 2020, 77(3): 487-512. https: / / doi.org / 10.1007 / s10898-020-00883-2.

[0116] Performance analysis: The computation time of branch-and-bound method grows exponentially with the problem dimension, and its response speed cannot meet the real-time optimization requirement when the dimension exceeds 6. In sharp contrast, the double-coupled network architecture proposed in the present invention shows significant advantages: although the offline training stage requires high computational cost, the total time consumption (prediction time consumption + bias correction time consumption) of the online prediction stage is approximately linear with the dimension, and the computational efficiency is at least one order of magnitude higher than that of the branch-and-bound method. In particular, in the multi-objective optimization scenario (such as rich personalized product demand), with the increase of the number of objective functions, the average total time consumption (training time consumption + online time consumption) of a single solution of the present architecture shows a decreasing trend, and eventually achieves a cross-order leap in overall efficiency.

[0117] Precision analysis: Experimental verification shows that the average deviation rate of the predicted solution of the present architecture from the theoretical optimal solution is basically less than 1e-3. By increasing the random sample points and the training precision of the constraint conditions, this distance can be further reduced. It is worth noting that the predicted solution without bias correction basically meets the actual precision requirement, which provides an effective precision-efficiency trade-off scheme for scenarios with strict real-time requirements.

[0118] Example 3: for the scenario of Example 1, Figure 3 A schematic diagram of an electronic device 10 that can be used to implement the embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers.

[0119] The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smartphones, wearable devices (such as headsets, glasses, watches, etc.), and other similar computing devices. The components shown in the present invention, their connections, and relationships, and their functions, are merely examples, and are not intended to limit the implementation of the present invention described and / or claimed in the present invention.

[0120] As Figure 3As shown, the electronic device 10 includes at least one processor 11, and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., communicatively connected to the at least one processor 11, where the memory stores a computer program executable by the at least one processor. The processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or loaded into the random access memory (RAM) 13 from the storage unit 18. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0121] Various components in the electronic device 10 are connected to the I / O interface 15, including an input unit 16, such as a keyboard, a mouse, etc., an output unit 17, such as various types of displays, a speaker, etc., a storage unit 18, such as a magnetic disk, an optical disk, etc., and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0122] The processor 11 can be various general and / or special-purpose processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 11 performs various methods and processes described above, such as the coupled physical neural network based robotic arm optimization control method.

[0123] In some embodiments, the coupled physical neural network based robotic arm optimization control method can be implemented as a computer program tangibly embodied in a computer readable storage medium, such as the storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the coupled physical neural network based robotic arm optimization control method described above can be performed. Alternatively, in other embodiments, the processor 11 can be configured to perform the coupled physical neural network based robotic arm optimization control method by any other appropriate means, such as by means of firmware.

[0124] The various embodiments of the systems and techniques described above can be implemented in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a load programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0125] Computer programs used to implement the processes of the present application can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the computer program, when executed, can cause instructions defined in the flow charts and / or block diagrams to be implemented on the computer or other programmable apparatus. The computer programs can be executed entirely on a machine, partially on a machine, partially on a machine as a stand-alone software package, partially on a machine and partially on a remote machine or entirely on a remote machine or server.

[0126] In the context of the present application, a computer-readable storage medium can be a tangible medium that can contain or store computer programs for use by or in connection with an instruction execution system, apparatus, or device. Computer-readable storage media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium can be a machine-readable signal medium. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0127] To provide for interaction with a user, the systems and techniques described here can be implemented on an electronic device having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the electronic device. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0128] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0129] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. A server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system, to solve the defects of large management difficulty and weak business scalability in traditional physical host and VPS service.

[0130] Finally, it is to be understood that the above-described embodiments are merely exemplary of how to practice the application, and that the application should not be limited to the specifically enumerated benefits and advantages set forth above but should be given a full range of equivalents of equivalents of the broadest possible interpretation of the appended claims to take into account the conveyances of substitutes, alternatives, and modifications as comprised within the present application.

Claims

1. A robotic arm optimization control method based on coupled physical neural networks, characterized in that: The method includes the following steps: S1: Perform dynamic analysis on the robotic arm to obtain the constraint equations, and establish the objective function based on the control strategy of the robotic arm, thereby extracting the optimization problem of the robotic arm; S2: Based on the dynamic characteristics of the robotic arm, construct the exact solution dataset corresponding to the constraint equations; S3: Build a constraint network using a physical information neural network and train it using an exact solution dataset; S4: Utilize deep learning networks to establish the objective function network; S5: Use the objective function as the loss function and embed the constraint network into the objective function network to establish a dual-network coupled architecture; S6: Train the dual-network coupled architecture using a random training set; S7: Utilize the trained objective function network to output the predicted optimal solution; S8: Use correction methods to adjust the predicted optimal solution results to obtain high-precision optimal solution results; S9: The robot arm is set according to the high-precision optimal solution results to adjust the parameters and realize the control strategy; The optimization problem described is in the form of: Among them, the independent variable parameter The dependent variable x = x(t,p) = (x1(t,p),…,x n (t,p)), dynamic equation Robotic arm interval process control strategy robotic arm end-effector control strategy obj corresponds to the objective function equation, st corresponds to the constraint equation; m is the parameter dimension, a positive integer; n is the dependent variable dimension, a positive integer; t0, t f p represents the initial and final times. L p U These are the lower and upper bounds of the parameter p, respectively; The constraint network is a deep learning network connected in series with constraint equations; the objective function network is a deep learning network connected in series with a parameter normalization equation; the dual-network coupling architecture is composed of an objective function network connected in series with a constraint network, and then connected in series with an objective function equation. The exact solution dataset corresponding to the constraint equation described in step S2 is obtained by solving the dependent variable x of the dynamic equation Φ using a randomly fixed parameter p.

2. The robotic arm optimization control method based on coupled physical neural networks according to claim 1, characterized in that, The method for training the constraint network in step S3 is a physical information neural network training method based on genetic algorithm. Specifically, the physical information neural network training method based on genetic algorithm is as follows: S301: Initialize the population: Randomly select from the parameter range [p L ,p U Generate an initial population containing M individuals. S302: Traversal from Select parameters Replace the parameter p in Φ and use numerical methods to solve for the dependent variable value x at time t. * ; S303: Will Store it in the exact solution training dataset; S304: Training the PINN network using the exact solution training dataset Conduct training; S305: If Then, new parameter individuals are generated using selection, crossover, and mutation operations. And repeat steps S302 to S305 until... And output the trained network Or it may reach the maximum number of iterations and report an error; where α is the absolute error threshold.

3. The robotic arm optimization control method based on coupled physical neural networks according to claim 2, characterized in that, The selection operation is to... Chinese correspondence All individuals selected are directly used as new parameter individuals for the next generation; these parameter individuals do not need to be recalculated numerically for the dependent variable x. * After the selection operation is completed, if the number of individuals does not reach M, then crossover and mutation operations are performed proportionally. The crossover operation involves pairing the selected individuals and performing a crossover operation with a certain probability to generate new individuals. The mutation operation involves performing a mutation operation with a certain probability within the parameter range [p]. L ,p U New individuals are generated within the [ ].

4. The robotic arm optimization control method based on coupled physical neural networks according to claim 1, characterized in that, The training process of the dual-network coupled architecture described in step S6 is as follows: S601: A set Y of N randomly generated m-dimensional sample points; S602: Input the sample point set Y into the deep learning network of the objective function network to obtain the predicted value. Where α is the deep learning network of the objective function network. The parameters to be trained; S603: Using parameter normalization equations For predicted values The parameter value p is obtained by processing it. S604: Input the parameter value p and the independent variable t into the constraint network to obtain the predicted value of the dependent variable. Where θ is the constraint network The parameters to be trained; S605: Randomly select slack variables within the interval [-γ, γ]. The dependent variable value is obtained by relaxing the dependent variable. in, γ = max{|E(∈)-3·D(∈)|,|E(∈)+3·D(∈)|}, where ε is the constraint network. The residual; S606: Using the independent variable t, parameter value p, and dependent variable value x as inputs, calculate the loss value Loss = J(p) with the objective function being the loss function; S607: Use gradient descent of the loss value to adjust the parameter α of the deep learning network of the objective function network; S608: Repeat steps S602 to S607 until the loss value meets the threshold setting to complete the training.

5. The robotic arm optimization control method based on coupled physical neural networks according to claim 4, characterized in that, For ease of calculation, all elements in γ are taken as α, where α is likely greater than γ in a meaningful sense, which can guarantee the equivalence of the solutions.

6. The robotic arm optimization control method based on coupled physical neural networks according to claim 1, characterized in that, Steps S1 to S4 are performed offline to first build and train the constraint network and the objective function network; when online, only the operations of steps S5 to S9 need to be adjusted according to the objective function; this can significantly improve computational efficiency.

7. The robotic arm optimization control method based on coupled physical neural networks according to claim 1, characterized in that, The aforementioned correction method is either Newton's iteration method or random walk method.

8. The robotic arm optimization control method based on coupled physical neural networks according to claim 1, characterized in that, In step S2, the exact solution dataset corresponding to the constraint equation is constructed by randomly fixing the parameter p and measuring the dependent variable x of the robotic arm.

9. An apparatus for use in the robotic arm optimization control method based on coupled physical neural networks as described in any one of claims 1 to 8, characterized in that, The device includes: An electronic device includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the above-described robotic arm optimization control method based on coupled physical neural networks; A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that are used to cause the processor to execute the above-described robotic arm optimization control method based on coupled physical neural networks.

Images