Method for estimating high-fidelity torque of collimated direct-drive actuator based on incremental width learning

Abstract

The invention discloses a collimation direct-drive actuator high-fidelity torque estimation method based on increment width learning, and belongs to the technical field of robot actuator torque estimation, and the method comprises the following steps: S1, collecting multi-dimensional state parameters, and constructing a data set; s2, obtaining weights of a feature layer and a top layer according to the data set in the S1, and constructing a basic width learning model; s3, based on the basic width learning model in the S2, when a new data set is obtained, using an incremental learning algorithm to update weights of a feature layer and a top layer, and obtaining an updated model; and S4, deploying the model to a torque control system, and constructing closed-loop control. By adopting the high-fidelity torque estimation method for the collimated direct-drive actuator based on increment width learning, the representation capability can be improved, the estimation precision can be ensured, the training overhead can be reduced, the model can be updated in real time to adapt to dynamic working conditions, high-precision estimation can be realized without an additional torque sensor, and actuator control can be effectively supported.

Description

Technical Field

[0001] This invention belongs to the field of robot actuator torque estimation technology, specifically involving a high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning. Background Technology

[0002] Quasi-direct drive actuators are core drive components in fields such as legged robots and exoskeleton robots. To achieve high-precision torque closed-loop control, it is necessary to sense the real torque at the actuator output in real time. However, there are currently technical challenges that urgently need to be overcome in high-fidelity torque sensing.

[0003] Existing technologies have the following shortcomings: installing torque sensors significantly increases system cost, weight, and volume, making it difficult to meet lightweight design requirements; deep neural network-based methods are time-consuming to train, computationally complex, and resource-intensive, making them difficult to adapt to embedded systems with limited memory, and new data requires retraining the entire model, failing to meet the real-time update requirements under dynamic operating conditions; dynamic model-based methods are highly dependent on the accurate understanding of motor models and parameters, requiring complex offline parameter identification, and are difficult to model accurately.

[0004] Therefore, there is an urgent need for a new method that can reduce training overhead, update in real time, and achieve high-fidelity torque estimation. Summary of the Invention

[0005] The purpose of this invention is to provide a high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning. This method can improve representation capabilities, ensure estimation accuracy, and reduce training overhead. It can update the model in real time to adapt to dynamic working conditions and achieve high-precision estimation without the need for additional torque sensors, effectively supporting actuator control.

[0006] To achieve the above objectives, this invention provides a high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning, comprising the following steps:

[0007] S1. Collect multidimensional state parameters and construct a dataset;

[0008] S2. Use the dataset constructed in S1 to train the basic width learning model and obtain the feature layer and top layer weights;

[0009] S3. Based on the basic width learning model trained by S2, when a new dataset is acquired, the feature layer and top-level weights are dynamically updated through an incremental learning algorithm to obtain the updated model.

[0010] S4. Deploy the basic width learning model trained in S2 or the updated model in S3 to the torque control system of the quasi-direct drive actuator to construct closed-loop control.

[0011] Preferably, the multidimensional state parameters in S1 include: rotor position, angular velocity, motor current, and output torque.

[0012] Preferably, multi-dimensional state parameters are collected in two scenarios: Scenario 1 is a humanoid robot bench test scenario; Scenario 2 is an exoskeleton robot application scenario.

[0013] Preferably, the feature layer is obtained in S2 as follows:

[0014] The initial training set is represented as:

[0015] {(X,Y)|X∈R N×d ,Y∈R N×c};

[0016] Where N is the number of samples; d is the feature space dimension of the samples; X is the input sample; Y is the target output corresponding to the input sample X; and c is the dimension of the target output.

[0017] The first group of mapped feature nodes Z1 is represented as follows:

[0018] Z1=φ1(W e1 X+β e1 );

[0019] Among them, W e1 β represents the weights of the first set of mapped feature nodes. e1 φ1 is the bias of the first group of mapped feature nodes, which are all randomly generated; φ1 is the activation function of the first group of mapped feature nodes.

[0020] The i-th group of mapping feature nodes Z i Represented as:

[0021] Z i =φ i (W Zi Z i-1 +W ei X+β ei ), i∈[2,3…,n];

[0022] Among them, W Zi and W ei β represents the weight of the i-th group of mapped feature nodes; ei The bias of the i-th group of mapped feature nodes is randomly generated; φ i Let be the activation function for the i-th group of mapped feature nodes; n is the number of groups of mapped feature nodes;

[0023] Mapping feature node layer Z n Represented as:

[0024] Z n =[Z1,Z2,…,Z n ];

[0025] Generate enhanced feature nodes. The first group of enhanced feature nodes H1 is represented as follows:

[0026] H1=ξ1(W k1 Z n +β k1 );

[0027] Among them, W k1 The weights for the first group of enhanced feature nodes; β k1 ξ1 is the bias of the first group of enhanced feature nodes, which are all randomly generated; ξ1 is the activation function of the first group of enhanced feature nodes.

[0028] The j-th group of enhanced feature nodes H j Represented as:

[0029] H j =ξ j (W Hj H j-1 +W kj Z n +β kj ),j∈[2,3...m];

[0030] Among them, W Hj and W kj The weights of the enhanced feature nodes in the j-th group; β kj The bias of the j-th group of enhanced feature nodes is randomly generated; ξ j Let be the activation function for the j-th enhanced feature node; m is the number of mapped feature node groups;

[0031] Enhanced Feature Node Layer H m Represented as:

[0032] H m =[H1,H2,…,H m ];

[0033] The integrated feature layer is represented as:

[0034]

[0035] in, This is the feature layer matrix.

[0036] Preferably, the top-level weights are obtained in S2 as follows:

[0037] Based on the generated feature layer, the output of the width learning model is represented as follows:

[0038]

[0039] in, These are the top-level weights connecting the feature layer and the output layer;

[0040] The loss function with L2 regularization is minimized as follows:

[0041]

[0042] Where λ is the regularization parameter; The norm operator for the loss term; The norm operator for the regularization term;

[0043] By differentiating the loss function and setting the derivative to zero, the analytical solution for the top-level weights is obtained, expressed as:

[0044]

[0045] Where I is the identity matrix, for The pseudo-inverse matrix;

[0046] The torque estimation model is constructed by combining the feature layer and the top-level weights, and is expressed as follows:

[0047]

[0048] Among them, T out To estimate the torque value.

[0049] Preferably, the incremental update mechanism in S3 is as follows:

[0050] First, the newly acquired dataset {(X I ,Y I )|X∈R N×d ,Y∈R N×c}, generate the corresponding feature layer

[0051] Next, the feature layer matrix is ​​updated, as follows:

[0052]

[0053] in, For A I The transpose of the matrix; This is the updated feature layer matrix;

[0054] The updated pseudo-inverse matrix is ​​represented as:

[0055]

[0056] in,

[0057]

[0058] in,

[0059] Update the top-level weights, represented as:

[0060]

[0061] in, The corresponding output for incremental training data; This is the updated top-level weight matrix;

[0062] The width learning model updated based on the incremental learning algorithm is represented as:

[0063]

[0064] in, This is an estimate of the output torque; ψ * This represents the input-output mapping relationship of a width learning model updated based on an incremental learning algorithm.

[0065] Preferably, in S4, the closed-loop control includes the following steps:

[0066] S401, Real-time acquisition of motor three-phase current (i A i B i C ), which is converted into the current component in the rotating coordinate system through Clarke-Park forward transformation (i d i q );

[0067] S402, transform i d i q The model learns the width of input multi-dimensional state parameters such as rotor position and angular velocity, and outputs the estimated torque value T in real time. out Simultaneously, the controller sends out the torque reference value T. * ;

[0068] The S403 PI torque controller uses the difference between the torque reference value and the estimated torque value as the error signal; it outputs the q-axis current reference value through proportional and integral regulation.

[0069] S404, with This indicates the deviation between the d-axis current reference value and the actual value; (in words) This indicates the deviation between the q-axis current reference value and the actual value. The PI current controller outputs a control voltage in the rotating coordinate system based on the error signals between the d-axis and q-axis. Achieve rapid current tracking of reference value;

[0070] S405. Through inverse Park transformation, the voltage in the rotating coordinate system is... Converted to a two-phase stationary coordinate system voltage SVPWM module based on A PWM pulse signal is generated to drive the motor inverter output voltage, thereby achieving torque output.

[0071] This invention also provides a high-fidelity torque estimation system for quasi-direct drive actuators based on incremental width learning, comprising:

[0072] The data acquisition and preprocessing module is used to collect multidimensional state parameters and construct a dataset;

[0073] The width learning model training module is connected to the data acquisition and preprocessing module. It trains a basic width learning model based on the dataset to obtain the feature layer and top-level weights.

[0074] The incremental learning model update module is connected to the width learning model training module. Based on the basic width learning model, when a new dataset is acquired, the feature layer and top-level weights are dynamically updated through the incremental learning algorithm to obtain the updated model.

[0075] The closed-loop control deployment module, connected to the incremental learning model update module, is used to deploy the basic width learning model or the updated model to the torque control system of the collimator actuator to construct closed-loop control.

[0076] Therefore, the present invention employs the above-mentioned high-fidelity torque estimation method for collimated drive actuators based on incremental width learning. Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0077] (1) This invention can achieve high-precision torque estimation without the need for an additional torque sensor, providing support for high-precision, high-dynamic-response torque control of quasi-direct drive actuators;

[0078] (2) The present invention uses the ridge regression algorithm to solve the optimal top-level weights, which greatly reduces the training complexity and time cost while ensuring high estimation accuracy.

[0079] (3) This invention introduces an incremental learning algorithm, which realizes real-time model updates through an efficient weight update mechanism. New data can be integrated without retraining, adapting to dynamic working conditions.

[0080] (4) This invention improves the model's representation ability by constructing a wide learning model with a cyclic cascade mechanism, fully extracting the dynamic features of multidimensional state parameters.

[0081] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0082] Figure 1This is a flowchart of an embodiment of the high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning according to the present invention;

[0083] Figure 2 This is a diagram illustrating the architecture and algorithm structure of an embodiment of the high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning according to the present invention.

[0084] Figure 3 This is a schematic diagram of a humanoid robot bench data acquisition platform, representing an embodiment of the high-fidelity torque estimation method for collimated drive actuators based on incremental width learning according to the present invention.

[0085] Figure 4 This is a schematic diagram of the integrated torque monitoring interface of a knee exoskeleton device, representing an embodiment of the high-fidelity torque estimation method for collimated drive actuators based on incremental width learning of the present invention.

[0086] Figure 5 This figure shows the experimental results of multi-model comparison of torque estimation for quasi-direct drive actuators, based on the high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning, according to an embodiment of the present invention. Figure 5 (a) in the figure is a comparison chart of torque estimation throughout the bench test cycle; Figure 5 (b) in the figure is a magnified comparison of the local area during the dynamic decay stage;

[0087] Figure 6 The figure shows the experimental results of the subject torque estimation in an embodiment of the high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning of the present invention. Figure 6 (a) represents 45% to 55% of the gait cycle, which corresponds to the transition of the exoskeleton knee joint from "supporting body weight" to "swinging the lower leg forward"; Figure 6 (b) represents 60% to 70% of the gait cycle, which corresponds to the transition of the exoskeleton knee joint from "lower leg swinging forward" to "supporting the body again". Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0089] Example 1

[0090] like Figures 1-4As shown, the high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning of the present invention includes the following steps:

[0091] S1: Collect multi-dimensional state parameters of the quasi-direct drive actuator under different scenarios and working conditions, including rotor position, angular velocity, motor current and output torque, and establish a dataset; the data acquisition scenarios include humanoid robot bench test scenarios and exoskeleton robot application scenarios;

[0092] The humanoid robot bench test scenario includes: quasi-direct drive actuators, torque sensors, brakes, couplings, data acquisition boards, and power systems.

[0093] In speed control mode, different speed reference values ​​are set, and the brake current is adjusted to simulate different loads. Simultaneously, rotor position, angular velocity, motor current, and output torque measured by torque sensors are acquired and transmitted via a CAN network. The dataset is divided into training, testing, and incremental datasets at 60%, 30%, and 10% to cover different speed and load conditions.

[0094] In this embodiment, the quasi-direct drive actuator integrates a brushless motor, a 7:1 planetary reducer, a 14-bit magnetic encoder, and a current sensor; the sampling frequency for rotor position and angular velocity is 50Hz.

[0095] Application scenarios for exoskeleton robots include: constructing bilateral knee exoskeleton devices and generating auxiliary torque using a bio-inspired torque control strategy. Seven healthy adult subjects were selected (4 males and 3 females, with a mean age of 23.0 ± 1.0 years).

[0096] Subjects walked on a treadmill at a comfortable speed, and data were simultaneously collected from rotor position, angular velocity, motor current, and output torque measured by torque sensors. Personalized datasets were constructed from six subjects, using leave-one-out cross-validation: data from one subject was used as the test set, the rest as the training set, and data from the seventh subject was used as the incremental dataset.

[0097] In this embodiment, the bilateral knee exoskeleton device includes a quasi-direct drive actuator that integrates a high torque density brushless DC motor, a 9:1 planetary gearbox, a 14-bit magnetic encoder, and a current sensor.

[0098] S2: Use the torque dataset constructed in S1 to train a basic width learning model to achieve nonlinear modeling of actuator dynamics; including generating feature layers based on a cyclic cascade mechanism and using the ridge regression algorithm to solve for the optimal top-level weight matrix;

[0099] In this step, generating the feature layer specifically involves:

[0100] The initial training set is represented as:

[0101] {(X,Y)|X∈R N×d ,Y∈R N×c};

[0102] Where N is the number of samples; d is the feature space dimension of the samples; X is the input sample; Y is the target output corresponding to the input sample X; and c is the dimension of the target output.

[0103] The first group of mapped feature nodes Z1 is represented as follows:

[0104] Z1=φ1(W e1 X+β e1 );

[0105] Among them, W e1 β represents the weights of the first set of mapped feature nodes. e1 φ1 is the bias of the first group of mapped feature nodes, which are all randomly generated; φ1 is the activation function of the first group of mapped feature nodes.

[0106] The i-th group of mapping feature nodes Z i Represented as:

[0107] Z i =φ i (W Zi Z i-1 +W ei X+β ei ), i∈[2,3…,n];

[0108] Among them, W Zi and W ei β represents the weight of the i-th group of mapped feature nodes; ei The bias of the i-th group of mapped feature nodes is randomly generated; φ i Let be the activation function for the i-th group of mapped feature nodes; n is the number of groups of mapped feature nodes;

[0109] Mapping feature node layer Z n Represented as:

[0110] Z n =[Z1,Z2,…,Z n ];

[0111] The generated mapping feature nodes are used to generate enhanced feature nodes;

[0112] Generate enhanced feature nodes. The first group of enhanced feature nodes H1 is represented as follows:

[0113] H1=ξ1(W k1 Z n +β k1 );

[0114] Among them, W k1The weights for the first group of enhanced feature nodes; β k1 ξ1 is the bias of the first group of enhanced feature nodes, which are all randomly generated; ξ1 is the activation function of the first group of enhanced feature nodes.

[0115] The j-th group of enhanced feature nodes H j Represented as:

[0116] H j =ξ j (W Hj H j-1 +W kj Z n +β kj ),j∈[2,3...m];

[0117] Among them, W Hj and W kj The weights of the enhanced feature nodes in the j-th group; β kj The bias of the j-th group of enhanced feature nodes is randomly generated; ξ j Let be the activation function for the j-th enhanced feature node; m is the number of mapped feature node groups;

[0118] Enhanced Feature Node Layer H m Represented as:

[0119] H m =[H1,H2,…,H m ];

[0120] The integrated feature layer is represented as:

[0121]

[0122] in, The feature layer matrix;

[0123] By introducing preceding node information, the model can effectively capture data dependencies that change over time and extract dynamic features from the input data.

[0124] Ridge regression is used to solve for the top-level weights. By establishing a mapping relationship between the inputs (motor rotor position, angular velocity, and motor current) and the output torque, the output torque is estimated, specifically as follows:

[0125] Based on the generated feature layer, the output of the width learning model is represented as follows:

[0126]

[0127] in, These are the top-level weights connecting the feature layer and the output layer;

[0128] The loss function with L2 regularization is minimized as follows:

[0129]

[0130] Where λ is the regularization parameter; The norm operator for the loss term; The norm operator for the regularization term;

[0131] By differentiating the loss function and setting the derivative to zero, the analytical solution for the top-level weights is obtained, expressed as:

[0132]

[0133] Where I is the identity matrix, for The pseudo-inverse matrix;

[0134] Ridge regression is used to solve the top-level weights, avoiding iterative optimization and improving training efficiency;

[0135] The torque estimation model is constructed by combining the feature layer and the top-level weights, and is expressed as follows:

[0136]

[0137] Among them, T out To estimate the torque value;

[0138] The torque estimation model was trained and evaluated on the test set. The estimation accuracy was quantified using normalized root mean square error. The time from feature layer generation to weight calculation was statistically analyzed to verify the efficiency advantage.

[0139] S3: Based on the basic width learning model trained by S2, when a new torque dataset is acquired, the feature layer and top-level weights are dynamically updated through an incremental learning algorithm; this includes updating the feature layer matrix, deriving the pseudo-inverse matrix update formula, and defining the top-level weight update method, so that the torque estimation model can be updated without retraining.

[0140] In this step, the incremental update mechanism first updates the newly acquired dataset {(X)}. I ,Y I )|X∈R N×d ,Y∈R N×c}, generate the corresponding feature layer

[0141] Next, the feature layer matrix is ​​updated, as follows:

[0142]

[0143] in, This is the initial network feature layer. For A IThe transpose of the matrix; This is the updated feature layer matrix;

[0144] The updated pseudo-inverse matrix is ​​represented as:

[0145]

[0146] in,

[0147]

[0148] in,

[0149] Update the top-level weights, represented as:

[0150]

[0151] in, The corresponding output for incremental training data; This is the updated top-level weight matrix;

[0152] By using a width-based learning model updated using an incremental learning algorithm, the torque estimation model can be updated without retraining, as shown below:

[0153]

[0154] in, This is an estimate of the output torque; ψ * This represents the input-output mapping relationship as updated by the width learning model based on the incremental learning algorithm.

[0155] S4: Deploy the basic width learning model trained in S2 or the updated model in S3 to the torque control system of the quasi-direct drive actuator to build closed-loop control; acquire rotor position, angular velocity and motor current in real time as input, estimate the output torque through the torque estimation model, and provide feedback for torque closed-loop control;

[0156] This step involves the following steps to achieve closed-loop control:

[0157] S401, Real-time acquisition of motor three-phase current (i A i B i C ), which is converted into the current component in the rotating coordinate system through Clarke-Park forward transformation (i d i q );

[0158] S402, transform i d i qThe model learns the width of input multi-dimensional state parameters such as rotor position and angular velocity, and outputs the estimated torque value T in real time. out Simultaneously, the controller (such as the motion planning layer of a humanoid robot) sends out the torque reference value T. * ;

[0159] The S403 PI torque controller uses the difference between the torque reference value and the estimated torque value as the error signal; it outputs the q-axis current reference value through proportional and integral regulation.

[0160] S404, with This indicates the deviation between the d-axis current reference value and the actual value; (in words) This indicates the deviation between the q-axis current reference value and the actual value. The PI current controller outputs a control voltage in the rotating coordinate system based on the error signals between the d-axis and q-axis. Achieve rapid current tracking of reference value;

[0161] S405. Through inverse Park transformation, the voltage in the rotating coordinate system is... Converted to a two-phase stationary coordinate system voltage SVPWM module based on A PWM pulse signal is generated to drive the motor inverter output voltage, ultimately achieving precise torque output and completing closed-loop control.

[0162] like Figures 5-6 As shown, the Wide Learning System (BLS) demonstrates higher estimation accuracy, better dynamic tracking capability, and better generalization performance than traditional methods in torque estimation tasks, making it more suitable for real-time and accurate torque estimation.

[0163] Therefore, the present invention adopts the above-mentioned high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning. This method can improve the representation capability, ensure the estimation accuracy and reduce the training overhead. It can update the model in real time to adapt to dynamic working conditions and achieve high-precision estimation without additional torque sensors, effectively supporting actuator control.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning, characterized in that, Includes the following steps: S1. Collect multidimensional state parameters and construct a dataset; S2. Use the dataset constructed in S1 to train the basic width learning model and obtain the feature layer and top layer weights; S3. Based on the basic width learning model trained by S2, when a new dataset is acquired, the feature layer and top-level weights are dynamically updated through an incremental learning algorithm to obtain the updated model. S4. Deploy the basic width learning model trained in S2 or the updated model in S3 to the torque control system of the quasi-direct drive actuator to construct closed-loop control.

2. The high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning according to claim 1, characterized in that, The multidimensional state parameters in S1 include: rotor position, angular velocity, motor current, and output torque.

3. The high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning according to claim 2, characterized in that, Multidimensional state parameters were collected in two scenarios: Scenario 1 was a humanoid robot bench test scenario; Scenario 2 was an exoskeleton robot application scenario.

4. The high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning according to claim 3, characterized in that, The specific steps for obtaining the feature layer in S2 are as follows: The initial training set is represented as: {(X,Y)|X∈R N×d ,Y∈R N×c }; Where N is the number of samples; d is the feature space dimension of the samples; X is the input sample; Y is the target output corresponding to the input sample X; and c is the dimension of the target output. The first group of mapped feature nodes Z1 is represented as follows: Z1=φ1(W e1 X+β e1 ); Among them, W e1 β represents the weights of the first set of mapped feature nodes. e1 φ1 is the bias of the first group of mapped feature nodes, which are all randomly generated; φ1 is the activation function of the first group of mapped feature nodes. The i-th group of mapping feature nodes Z i Represented as: WITH i =φ i (IN Zi WITH i-1 +W ei X+β ei ),i∈[2,3…,n]; Among them, W Zi and W ei β represents the weight of the i-th group of mapped feature nodes; ei The bias of the i-th group of mapped feature nodes is randomly generated; φ i Let be the activation function for the i-th group of mapped feature nodes; n is the number of groups of mapped feature nodes; Mapping feature node layer Z n Represented as: WITH n =[Z1,Z2,…,Z n ]; Generate enhanced feature nodes. The first group of enhanced feature nodes H1 is represented as follows: H1=ξ1(W k1 WITH n +β k1 ); Among them, W k1 The weights for the first group of enhanced feature nodes; β k1 ξ1 is the bias of the first group of enhanced feature nodes, which are all randomly generated; ξ1 is the activation function of the first group of enhanced feature nodes. The j-th group of enhanced feature nodes H j Represented as: H j =ξ j (W Hj H j-1 +W kj Z n +β kj ),j∈[2,3...m]; Among them, W Hj and W kj The weights of the enhanced feature nodes in the j-th group; β kj The bias of the j-th group of enhanced feature nodes is randomly generated; ξ j Let be the activation function for the j-th enhanced feature node; m is the number of mapped feature node groups; Enhanced Feature Node Layer H m Represented as: H m =[H1,H2,…,H m ]; The integrated feature layer is represented as: in, This is the feature layer matrix.

5. The high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning according to claim 4, characterized in that, The specific steps to obtain the top-level weights in S2 are as follows: Based on the generated feature layer, the output of the width learning model is represented as follows: in, These are the top-level weights connecting the feature layer and the output layer; The loss function with L2 regularization is minimized as follows: Where λ is the regularization parameter; The norm operator for the loss term; The norm operator for the regularization term; By differentiating the loss function and setting the derivative to zero, the analytical solution for the top-level weights is obtained, expressed as: Where I is the identity matrix, for The pseudo-inverse matrix; The torque estimation model is constructed by combining the feature layer and the top-level weights, and is expressed as follows: Among them, T out To estimate the torque value.

6. The high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning according to claim 5, characterized in that, The incremental update mechanism in S3 is as follows: First, the newly acquired dataset {(X I ,Y I )|X∈R N×d ,Y∈R N×c }, generate the corresponding feature layer Next, the feature layer matrix is ​​updated, as follows: in, For A I The transpose of the matrix; This is the updated feature layer matrix; The updated pseudo-inverse matrix is ​​represented as: in, in, Update the top-level weights, represented as: Among them, Y I T The corresponding output for incremental training data; This is the updated top-level weight matrix; The width learning model updated based on the incremental learning algorithm is represented as: in, This is an estimate of the output torque; ψ * This represents the input-output mapping relationship of a width learning model updated based on an incremental learning algorithm.

7. The high-fidelity torque estimation method for quasi-direct drive actuators based on incremental width learning according to claim 6, characterized in that, In S4, closed-loop control includes the following steps: S401, Real-time acquisition of motor three-phase current (i A i B i C ), which is converted into the current component in the rotating coordinate system through Clarke-Park forward transformation (i d i q ); S402, transform i d i q The model learns the width of input multi-dimensional state parameters such as rotor position and angular velocity, and outputs the estimated torque value T in real time. out Simultaneously, the controller sends out the torque reference value T. * ; The S403 PI torque controller uses the difference between the torque reference value and the estimated torque value as the error signal; it outputs the q-axis current reference value through proportional and integral regulation. S404, with This indicates the deviation between the d-axis current reference value and the actual value; (in words) This indicates the deviation between the q-axis current reference value and the actual value. The PI current controller outputs a control voltage in the rotating coordinate system based on the error signals between the d-axis and q-axis. Achieve rapid current tracking of reference value; S405. Through inverse Park transformation, the voltage in the rotating coordinate system is... Converted to a two-phase stationary coordinate system voltage SVPWM module based on A PWM pulse signal is generated to drive the motor inverter output voltage, thereby achieving torque output.

8. A high-fidelity torque estimation system for collimated drive actuators based on incremental width learning, applied to the high-fidelity torque estimation method for collimated drive actuators based on incremental width learning as described in any one of claims 1-7, characterized in that, include: The data acquisition and preprocessing module is used to collect multidimensional state parameters and construct a dataset; The width learning model training module is connected to the data acquisition and preprocessing module. It trains a basic width learning model based on the dataset to obtain the feature layer and top-level weights. The incremental learning model update module is connected to the width learning model training module. Based on the basic width learning model, when a new dataset is acquired, the feature layer and top-level weights are dynamically updated through the incremental learning algorithm to obtain the updated model. The closed-loop control deployment module, connected to the incremental learning model update module, is used to deploy the basic width learning model or the updated model to the torque control system of the collimator actuator to construct closed-loop control.

Images