Servo motor control parameter self-supervision optimization method based on physical embedding

Abstract

The invention discloses a servo motor control parameter self-supervision optimization method based on physical embedding. A self-supervision optimization model of a cascade structure of a bidirectional hierarchical encoder and a real physical system decoder is constructed. The two-way hierarchical encoder combines artificial decoding and neural network decoding, so that the network scale in the encoding process is light; a real physical system decoder is embedded into a servo motor system, a potential feature space output by an encoder is forced to be limited in a servo motor control parameter space, potential features are directly optimized, and therefore optimization of servo motor control parameters is achieved. In the self-supervised optimization model training process, by minimizing a loss function between an ideal position track and an actual position track, optimal potential features are learned, and after training is completed, potential features output by an encoder are used as optimized control parameters. The method has the characteristics that the control structure is non-intrusive, and a plurality of control parameters with strong coupling characteristics can be efficiently optimized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of servo motor control, and particularly relates to a servo motor control parameter self-supervision optimization method based on physical embedding. BACKGROUND

[0002] In the field of intelligent manufacturing, service robots and other high-tech fields, servo motors are important components, and their high-precision servo control directly determines the completion level of related tasks. Position servo control is an important scenario, which usually adopts a hierarchical control strategy, involves multiple coupled control parameters including position, speed and current, and urgently needs to solve the problem of multi-parameter coordination and improve the performance of the servo system.

[0003] Trial-and-error method is a traditional parameter adjustment method, but the process is time-consuming and the control precision is limited, so a more systematic and structured parameter adjustment method is needed. Common model-based tuning methods mainly include Ziegler-Nichols method and dominant pole placement method. However, these methods rely on low-order approximate modeling analysis of the motor system and the selection of empirical coefficients, which on the one hand still requires rich professional knowledge and manual experience, and on the other hand performs poorly when dealing with actual high-order nonlinear models.

[0004] Another type of method is intelligent parameter tuning method, including heuristic parameter tuning and neural network parameter tuning methods, heuristic tuning method includes particle swarm optimization algorithm (such as literature [C. Du, Z. Yin, Y. Zhang, J. Liu, X. Sun and Y. Zhong“Research on active disturbance rejection control with parameter autotune mechanism for induction motors based on adaptive particle swarm optimization algorithm with dynamic inertia weight”IEEE Trans. Power Electron, vol. 34, no. 3, pp. 2841-2855, Mar. 2019]) and ant colony optimization algorithm (such as literature [Z. Yin, C. Du, J. Liu, X. Sun and Y. Zhong“Research on auto disturbance-rejection control of induction motors based on an ant colony optimization algorithm”IEEE Trans. Ind. Electron, vol. 65, no. 4, pp. 3077-3094, Apr. 2018] and the like, when facing optimization problems containing more strongly coupled parameters, a larger population size usually needs to be set to improve the optimization ability, which leads to lower optimization efficiency, and the optimization result is more sensitive to the initial parameter configuration.The neural network parameter tuning method realizes parameter optimization by learning to establish a linear or nonlinear mapping relationship between system input and output. For example, the document [Y. Zeng, A. I. Maswood, J. Pou, X. Zhang, Z. Li and C. Sun“Active disturbance rejection control using artificial neural network for dual-active-bridge-based energy storage system”IEEE J. Emerg. Sel. Topics Power Electron, vol. 11, no. 1, pp. 301-311, Feb. 2023] uses pre-collected experimental or simulation data to train a neural network model, and then uses the trained model to generate actual control parameters; the document [J. P. de Moura, J. V. d. F. Neto and P. H. M. “Aneuro-fuzzy model for online optimal tuning of PID controllers in industrial system applications to the mining sector”IEEE Trans. Fuzzy Syst, vol. 28, no. 8, pp. 1864-1877, Aug. 2020] combines the characteristics of a specific control structure, fits its input-output relationship, and adjusts the network parameters in real time according to the control error, thereby obtaining optimal control parameters online. However, such neural network methods usually require a large amount of data, need to be deeply embedded in the control structure, and are more suitable for fast local optimization of a small number of parameters, and the computational overhead is relatively large.

[0005] Therefore, when facing the optimization task of multi-variable strong coupling, the existing control parameter optimization algorithm is often difficult to meet the following requirements at the same time: ① moderate model complexity; ② non-intrusive architecture of control structure; ③ efficient optimization of multiple parameters in parallel. Therefore, a new method is needed to solve the problem of multi-parameter collaborative setting. SUMMARY

[0006] In view of the above, the present application provides a servo motor control parameter self-supervised optimization method based on physical embedding, which constructs a self-supervised optimization model in the form of an encoder-decoder cascade structure, has the characteristics of non-intrusive control structure, and can efficiently optimize multiple control parameters with strong coupling characteristics.

[0007] A servo motor control parameter self-supervised optimization method based on physical embedding, comprising the following steps:

[0008] (1) building a self-supervised optimization model with encoder and decoder cascaded structure;

[0009] (2) designing specific implementation forms of the encoder and the decoder;

[0010] (3) training the self-supervised optimization model;

[0011] (4) after the training is completed, generating latent features by the encoder in the self-supervised optimization model as optimized servo motor control parameters.

[0012] Further, in the step (1), the input of the self-supervised optimization model is a given ideal rotor position trajectory, and the output is a reconstructed actual rotor position trajectory; the encoder encodes the input trajectory and outputs learned latent features, which are a compressed form of the input trajectory and represent key characteristics or structural information learned by the encoder from the input trajectory, and which contain key control parameters in the servo motor controller; the decoder decodes the latent features to obtain the actual rotor position trajectory; the model makes the encoder self-supervised learn the latent features of the input trajectory by minimizing the reconstruction error between the input trajectory and the output trajectory.

[0013] Further, in the step (2), the encoder adopts a bidirectional hierarchical encoder, which adopts a two-stage encoding structure design, first passes the input trajectory through a fixed-point encoder in the artificial encoding stage to obtain a low-dimensional fixed point, and then passes the low-dimensional fixed point through a neural network encoder in the machine encoding stage to output the latent features.

[0014] Further, the fixed-point encoder adopts non-explicit modeling, and the output low-dimensional fixed point is a vector determined by human, the dimension of the vector is determined according to the data dimension of the latent features, and the element in the vector takes a value range of (0, 1).

[0015] Further, the neural network encoder adopts a feedforward neural network modeling, which is composed of an input layer, a first hidden layer, a second hidden layer and an output layer, and the specific expression is:

[0016] z1=W1x+b1

[0017] a1=σ(z1)

[0018] z2=W2a1+b2

[0019] a2=σ(z2)

[0020] z=W3a2+b3

[0021] Where: x is a low-dimensional fixed point, z1 and z2 are intermediate variables, z is the potential feature, a1 and a2 are the outputs of the first hidden layer and the second hidden layer respectively, σ() represents the nonlinear activation function, W1 is the weight parameter vector between the input layer and the first hidden layer, b1 is the bias parameter vector between the input layer and the first hidden layer, W2 is the weight parameter vector between the first hidden layer and the second hidden layer, b2 is the bias parameter vector between the first hidden layer and the second hidden layer, W3 is the weight parameter vector between the second hidden layer and the output layer, and b3 is the bias parameter vector between the second hidden layer and the output layer.

[0022] Furthermore, the nonlinear activation function adopts the Softsign activation function, which is expressed as follows:

[0023]

[0024] Where: a represents the independent variable of the function.

[0025] Furthermore, in step (2), the decoder uses the real physical system of the servo motor to directly replace the usual neural network modeling. The real physical system includes a servo motor controller, a servo motor and its inverter. Under a given set of control parameters, the potential features are input to the decoder, and the system will respond with a certain motor state, thereby obtaining the actual rotor position trajectory.

[0026] Furthermore, the servo motor controller is based on a hierarchical control framework and adopts a three-loop structure of current loop, speed loop and position loop, wherein the position loop converts a given rotor position angle θ r The speed command ω0 is obtained through the feedforward link, and θ r and the actual rotor position angle θ a The difference between the two is used to generate the reference speed ω1 through the proportional link, and then the speed reference instruction ω is obtained by superimposing ω0 and ω1. r ; The speed loop will ω r The current command i0 is obtained through the feedforward link, and ω r and the actual speed ω a The difference between the two is used to generate the reference current i1 through the PI (proportional integral) link, and then i0 and i1 are superimposed to obtain the q-axis current reference instruction i qr ; The current loop will i qr and the actual q-axis stator current i q The difference between the two generates the q-axis reference voltage u through the PI link q * , the d-axis current reference instruction i dr =0 and the actual d-axis stator current i d The difference between the two generates the d-axis reference voltage u through the PI link d * , and then ud * and u q * The three-phase PWM signal is generated by a SVPWM (space vector pulse width modulation) module to drive the on-off of the power switching devices in the servo motor inverter.

[0027] Further, the self-supervised optimization model is trained in the step (3) by using the method of running while learning, and the specific process is as follows:

[0028] 3.1, initializing model parameters, including bias parameters and weight parameters in the neural network, learning rate and selection of the optimizer;

[0029] 3.2, inputting the given ideal rotor position trajectory into the model, so that the low-dimensional fixed-point input generated by the fixed-point encoder is input into the neural network encoder to obtain the latent feature, and the corresponding actual rotor position trajectory is obtained through the decoder response, and the reconstruction loss between the ideal rotor position trajectory and the actual rotor position trajectory is calculated;

[0030] 3.3, according to the reconstruction loss, the bias parameters and the weight parameters are updated by the gradient descent method through the optimizer, until the reconstruction loss converges or reaches the maximum iteration number, and the training is completed.

[0031] Further, the function expression of the reconstruction loss is as follows:

[0032] E=(θ r -θ a ) T (θ r -θ a ) / 2

[0033] Wherein: E is the function value of the reconstruction loss, θ r is the given ideal rotor position trajectory, θ a is the actual rotor position trajectory, and the superscript T represents transposition.

[0034] Based on the above technical scheme, the present application has the following beneficial technical effects:

[0035] 1. The present application adopts a feature extraction mechanism to optimize parameters, and constrains the feature extraction direction with the characteristics of the real motor system, avoiding the generation of features without control significance.

[0036] 2. The decoder of the self-supervised optimization model of the present application only cares about the real position trajectory of the motor system in response to the input control parameters; the network parameter update only depends on the position reconstruction error, without the need to analyze or modify the internal structure of the system, and has the non-invasive characteristic.

[0037] 3. The application introduces artificial feature coding in the feature coding stage, significantly reducing the size of the encoder; the network design avoids trigonometric functions and exponential functions, while maintaining strong nonlinear representation capabilities and improving computational efficiency, adapting to the control parameter optimization of low-power platforms.

[0038] 4. The latent feature dimension of the encoder output in the application is consistent with the dimension of the control parameter to be optimized, and the algorithm has the ability to simultaneously optimize multiple coupled control parameters. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The structure diagram of the servo motor controller in the embodiment of the application.

[0040] Figure 2 The flowchart of the self-supervised optimization method of the servo motor control parameter in the embodiment of the application.

[0041] Figure 3 The overall framework diagram of the self-supervised optimization model in the embodiment of the application.

[0042] Figure 4 The position tracking effect diagram before optimization in the embodiment of the application.

[0043] Figure 5 The position tracking effect diagram after optimization in the embodiment of the application.

[0044] Figure 6 The loss function change diagram of the model training process in the embodiment of the application. DETAILED DESCRIPTION

[0045] In order to more specifically describe the application, the technical solutions of the application will be described in detail below in combination with the drawings and specific embodiments.

[0046] This embodiment takes a small servo motor with a rated power of 2.3kW, a rated torque of 14.6Nm, and a rated speed of 1500r / min as an example to optimize multiple control parameters in the control structure; the servo motor controller uses a hierarchical control framework to achieve accurate tracking of a given ideal position trajectory.

[0047] In this embodiment, the given ideal position trajectory After S-curve planning, there are L=1000 discrete sampling points; the S-curve is controlled by five parameters of time length T, maximum position P max , maximum speed V max , maximum acceleration A max , and maximum jerk J max , where T=1s, P max =28.4rad, V max=52.4rad / s, A max =228.6rad / s 2 , J max =1000rad / s 3 .like Figure 1 As shown in the figure, the hierarchical control framework is divided into three loop structures: current loop, speed loop and position loop. The current loop uses a proportional-integral controller to control the d-axis and q-axis currents respectively, the speed loop uses a proportional-integral controller to control the speed, and the position loop uses a proportional controller to control the position. At the same time, in order to reduce tracking delay and error and improve response speed and accuracy, a feedforward link is added to the position loop and speed loop to compensate the control quantity in advance by predicting the dynamic situation of the system. The main parameters involved in this control framework include: the position loop proportional coefficient k pp , speed loop proportional coefficient k pw , current loop proportional coefficient k pi , speed loop integral coefficient k iw , current loop integral coefficient k ii , Position loop feedforward coefficient k wf , speed loop feedforward coefficient k af .

[0048] exist Figure 1 In the control structure shown, the actual position θ output by the position sensor C0 on the motor shaft is a and actual speed ω a , given position θ in the position loop r The speed command ω0 is obtained through the feedforward link C2 and the actual position θ a Comparison is performed, and the reference speed ω1 is generated through the proportional link C1, and the speed reference instruction ω is finally obtained through superposition. r ; Given position ω in the speed loop r The current command i0 is obtained through the feedforward link C4 and the actual speed ω a By comparison, the reference current i1 is generated through the PI link C3, and the q-axis current reference instruction i is finally obtained through superposition. qr ; Given q-axis current reference instruction i in the current loop qr and the actual q-axis current i q Compare, given the d-axis current reference instruction i dr and the actual d-axis current i d By comparison, the voltage reference vector is obtained after passing through the proportional integral links C5 and C6 respectively, and then the motor is controlled to track the given S-shaped position curve.

[0049] In order to achieve control parameter optimization, this embodiment establishes a self-supervised optimization model of the encoder and decoder cascade structure. Input into the self-supervised optimization model and output the reconstructed actual position trajectory The encoder learns the latent features of the input data by minimizing the reconstruction error between the given position trajectory and the actual position trajectory, thereby optimizing the control parameters. The encoder learns the latent features of the input data by minimizing the reconstruction error between the given position trajectory and the actual position trajectory, thereby optimizing the control parameters.

[0050] As shown in Figure 2 , the physical embedded servo motor control parameter self-supervised optimization method based on the present embodiment includes the following steps:

[0051] (1) Determine the framework composition of the self-supervised optimization model.

[0052] As shown in Figure 3 , the overall self-supervised optimization model in the present embodiment adopts a cascaded structure of an encoder and a decoder, and the encoder is designed as a bidirectional hierarchical neural network, with the given ideal position trajectory L representing the number of sampling points of the trajectory curve, and the output being the latent feature The latent feature z is a compressed form of the input trajectory, representing the key characteristics or structural information learned by the encoder from the input trajectory, where K represents the number of control parameters to be optimized. In the above hierarchical control framework, the parameters of the speed loop and the position loop need additional attention, and for this purpose, the present embodiment optimizes five control parameters, including the position loop proportional coefficient k pp , the speed loop proportional coefficient k pw , the speed loop integral coefficient k iw , the position loop feedforward coefficient k wf , and the speed loop feedforward coefficient k af , so K = 5.

[0053] The bidirectional hierarchical encoder adopts a two-stage encoding structure design, i.e., sequentially passing through the fixed-point encoder φ in the artificial encoding stage and the encoder g in the neural network encoding stage. The fixed-point encoder φ is not explicitly modeled, and its output is determined by artificial means. For servo motor systems, the number of sampling points of the position trajectory curve is large, and the dimension L of the input data will be much larger than the dimension K of the latent feature, which requires a large-scale neural network for position trajectory feature encoding, which is not conducive to some low-computing-resource applications. However, the position trajectory has determinacy, and there are corresponding fixed-point data in the input data space. Therefore, the ideal position trajectory θ r is artificially encoded to convert the input data into low-dimensional fixed-point The dimension D depends on the data dimension of the subsequent neural network encoder, and in the present embodiment, D = 5; to reduce the encoding difficulty of the subsequent neural network encoder, the element value range of the low-dimensional fixed-point x is (0, 1), and in the present embodiment, the value is 0.5.

[0054] The low-dimensional fixed point x is outputted as latent feature z by the encoder g of the neural network encoding stage. In this embodiment, the encoder g is modeled by four layers of feedforward neural network, including an input layer, a first hidden layer, a second hidden layer and an output layer. The input layer is composed of D neurons, the first hidden layer is composed of M neurons, which is used to linearly weight the neurons of the input layer and perform nonlinear activation; the second hidden layer is composed of N neurons, which is used to linearly weight the neurons of the first hidden layer and perform nonlinear activation; the output layer is composed of K neurons, which is used to linearly weight the neurons of the second hidden layer and output the latent feature z, that is, the control parameters of the real physical system. The specific expression is as follows:

[0055] z1 = W1x + b1

[0056] a1 = σ(z1)

[0057] z2 = W2a1 + b2

[0058] a2 = σ(z2)

[0059] z = W3a2 + b3

[0060] In the formula: is the intermediate output, are the outputs of the first hidden layer and the second hidden layer respectively, and are the weight parameters of the encoder g, and are the bias parameters of the encoder, and σ represents a nonlinear activation function; the weight parameters and the bias parameters are trainable parameters, which need to be updated according to the reconstruction error.

[0061] In this embodiment, the nonlinear activation function σ adopts a Softsign activation function, and its expression is as follows:

[0062]

[0063] The derivative of the nonlinear activation function σ is σ', and its expression is as follows:

[0064]

[0065] The conventional selection scheme of the activation function is sigmoid, tanh or ReLu function, sigmoid and tanh have strong nonlinear activation ability, but face the calculation of trigonometric function or exponential function, and a certain execution time is consumed when a large amount of data is involved in the calculation; ReLu is a piecewise linear function, but for shallow network, its nonlinear ability is weak and cannot be used in the design of the proposed encoder g. The embodiment adopts the Softsign function which has strong nonlinear activation ability, and the self and corresponding derivative function only involves basic operation, which can well balance the network size, nonlinear activation ability and calculation time.

[0066] In the embodiment, the decoder h directly replaces the general neural network modeling by using the real physical system related to the servo motor system, which includes a unified motor system from the controller to the inverter and the controlled motor. By using the above-mentioned hierarchical control framework, the encoder input is the latent feature z representing the control parameters, and the output is the actual position trajectory reconstructed by the decoder

[0067] The encoder learns the latent feature of the input data by minimizing the reconstruction error between the given position trajectory and the actual position trajectory. That is, the model realizes self-supervised learning only through the input-output relationship without any label guidance; due to the cascade design of the encoder and the decoder, the decoder embedded with the real physical system forces the latent feature to become the control parameters of the servo motor system, thereby avoiding non-physical learning results. Moreover, the reconstruction error is directly directed to the position tracking performance of the servo motor system, and after minimizing the reconstruction error, the self-supervised optimization model learns the control parameters corresponding to the high-precision position tracking performance, thereby realizing the optimization of the control parameters of the servo motor system.

[0068] (2) Training the self-supervised optimization model.

[0069] Step 1: initializing the network parameters and learning rate parameters of the encoder g in the self-supervised optimization model;

[0070] Step 2: inputting the artificially determined low-dimensional fixed point x into the encoder g, outputting the latent feature z representing the control parameters, inputting the latent feature z into the decoder built with the real motor physical system, and obtaining the actual position trajectory θ a ;

[0071] Step 3: calculating the reconstruction loss function between the ideal position trajectory θ r and the actual position trajectory θ a , and the expression form is:

[0072] E=(θ r -θ a ) T (θr -θ a ) / 2

[0073] Step 4: according to the reconstruction loss between the ideal position trajectory θ r and the actual position trajectory θ a , the network parameters of the encoder g in the self-supervised optimization model are updated using the gradient descent method, and the update form is:

[0074]

[0075] In the formula: η is the optimization learning rate, and sgn is the sign function.

[0076] Step 5: repeat steps 2-4 until the reconstruction loss function is lower than the threshold value or the iteration reaches the maximum number of iterations, and the training of the self-supervised optimization model is completed.

[0077] (4) output the optimized servo motor system control parameters.

[0078] Take out the trained self-supervised optimization model encoder g, input the low-dimensional fixed point x, obtain the output latent feature z and use it as the optimized servo motor system control parameters.

[0079] The embodiment applies the self-supervised optimization model to optimize the control parameters. Before optimization, the position trajectory tracking effect under the action of the initial control parameters is as shown in Figure 4 , there are obvious tracking errors at the beginning and end of the trajectory, and the position loss function E=3970.96 at this time; after optimization, the position tracking effect under the action of the optimized control parameters is as shown in Figure 5 , it can be seen that the servo motor system can accurately track the given position curve at this time, and the corresponding loss function E=1.55, and the tracking accuracy is improved by nearly 2562 times. The loss function value change curve of the iteration process is as shown in Figure 6 , it can be seen from the figure that the optimization algorithm of the present application gradually converges after the error is rapidly reduced at the beginning, and the simultaneous optimization of multiple control parameters is completed in 25 rounds, which has a significant efficiency advantage.

[0080] The above description of the embodiments is to facilitate those of ordinary skill in the art to understand and apply the present application. Those skilled in the art can easily make various modifications to the above embodiments, and apply the general principles described herein to other embodiments without creative labor. Therefore, the present application is not limited to the above embodiments, and the improvements and modifications of the present application made by those skilled in the art according to the disclosure of the present application should be within the scope of protection of the present application.

Claims

1. A self-supervised optimization method for servo motor control parameters based on physical embedding, characterized in that: The steps include: (1) Build a self-supervised optimization model for the encoder and decoder cascade structure; (2) Design the specific implementation of the encoder and decoder; (3) training the self-supervised optimization model; (4) After training, the encoder in the self-supervised optimization model is used to generate latent features as the optimized servo motor control parameters.

2. The method for self-supervised optimization of servo motor control parameters based on physical embedding according to claim 1, characterized in that: The input of the self-supervised optimization model in step (1) is a given ideal rotor position trajectory, and the output is a reconstructed actual rotor position trajectory; the encoder encodes the input trajectory and outputs the learned latent features, which are a compressed form of the input trajectory and represent the key characteristics or structural information learned by the encoder from the input trajectory, which include various key control parameters in the servo motor controller; the decoder decodes the latent features and reconstructs them to obtain the actual rotor position trajectory; the model enables the encoder to self-supervise the learning of the latent features of the input trajectory by minimizing the reconstruction error between the input trajectory and the output trajectory.

3. The method for self-supervised optimization of servo motor control parameters based on physical embedding according to claim 2, characterized in that: In step (2), the encoder adopts a bidirectional hierarchical encoder, which adopts a two-stage encoding structure design. First, the input trajectory is passed through a fixed-point encoder in the manual encoding stage to obtain a low-dimensional fixed point, and then the low-dimensional fixed point is passed through a neural network encoder in the machine encoding stage to output the potential feature.

4. The method for self-supervised optimization of servo motor control parameters based on physical embedding according to claim 3, characterized in that: The fixed-point encoder adopts non-explicit modeling, and the low-dimensional fixed point outputted by it is represented by a vector determined manually. The dimension of the vector is determined according to the data dimension of the potential feature, and the value range of the elements in the vector is (0, 1).

5. The method for self-supervised optimization of servo motor control parameters based on physical embedding according to claim 3, characterized in that: The neural network encoder is modeled using a feedforward neural network and consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The specific expression is: z1=W1x+b1 a1=σ(z1) z2=W2a1+b2 a2=σ(z2) z=W3a2+b3 Where: x is a low-dimensional fixed point, z1 and z2 are intermediate variables, z is the potential feature, a1 and a2 are the outputs of the first hidden layer and the second hidden layer respectively, σ() represents the nonlinear activation function, W1 is the weight parameter vector between the input layer and the first hidden layer, b1 is the bias parameter vector between the input layer and the first hidden layer, W2 is the weight parameter vector between the first hidden layer and the second hidden layer, b2 is the bias parameter vector between the first hidden layer and the second hidden layer, W3 is the weight parameter vector between the second hidden layer and the output layer, and b3 is the bias parameter vector between the second hidden layer and the output layer.

6. The method for self-supervised optimization of servo motor control parameters based on physical embedding according to claim 5, characterized in that: The nonlinear activation function adopts the Softsign activation function, which is expressed as follows: Where: a represents the independent variable of the function.

7. The method for self-supervised optimization of servo motor control parameters based on physical embedding according to claim 1, characterized in that: In step (2), the decoder uses the real physical system of the servo motor to directly replace the usual neural network modeling. The real physical system includes a servo motor controller, a servo motor and its inverter. Under a given set of control parameters, the potential features are input to the decoder, and the system will respond with a certain motor state to obtain the actual rotor position trajectory.

8. The method for self-supervised optimization of servo motor control parameters based on physical embedding according to claim 7, characterized in that: The servo motor controller is based on a hierarchical control framework and adopts a three-loop structure of current loop, speed loop and position loop, in which the position loop converts the given rotor position angle θ r The speed command ω0 is obtained through the feedforward link, and θ r and the actual rotor position angle θ a The difference between the two is used to generate the reference speed ω1 through the proportional link, and then the speed reference instruction ω is obtained by superimposing ω0 and ω1. r ; The speed loop will ω r The current command i0 is obtained through the feedforward link, and ω r and the actual speed ω a The difference between the two is used to generate the reference current i1 through the PI link, and then i0 and i1 are superimposed to obtain the q-axis current reference instruction i qr ; The current loop will i qr and the actual q-axis stator current i q The difference between the two generates the q-axis reference voltage u through the PI link q * , the d-axis current reference instruction i dr =0 and the actual d-axis stator current i d The difference between the two generates the d-axis reference voltage u through the PI link d * , and then u d * and u q * The SVPWM module generates three-phase PWM signals to drive the on and off of the power switching devices in the servo motor inverter.

9. The method for self-supervised optimization of servo motor control parameters based on physical embedding according to claim 3, characterized in that: In step (3), the self-supervised optimization model is trained by learning while running. The specific process is as follows: 3.1 Initialize model parameters, including bias parameters and weight parameters in the neural network, learning rate, and optimizer selection; 3.2 Input the given ideal rotor position trajectory into the model, input the low-dimensional fixed-point generated by the fixed-point encoder into the neural network encoder to obtain the latent features, and then use the latent features to obtain the corresponding actual rotor position trajectory through the decoder response, and calculate the reconstruction loss between the ideal rotor position trajectory and the actual rotor position trajectory; 3.3 Based on the reconstruction loss, the optimizer is used to iteratively update the bias parameters and weight parameters through the gradient descent method until the reconstruction loss converges or the maximum number of iterations is reached, and the training is completed.

10. The method for self-supervised optimization of servo motor control parameters based on physical embedding according to claim 9, characterized in that: The functional expression of the reconstruction loss is as follows: E=(θ r -θ a ) T (i r -θ a ) / 2 Where: E is the function value of reconstruction loss, θ r is a given ideal rotor position trajectory, θ a is the actual rotor position trajectory, and the superscript T represents transposition.

Images