Control method and system for motor controller

Abstract

The present invention provides a control method and system for a motor controller, which relates to the technical field of motor control, including: collecting multi-dimensional parameters and environmental parameters, combining them to obtain an operation data stream, performing feature extraction to obtain a spatiotemporal feature representation, learning state features and determining probability distribution, constructing a state transfer equation and obtaining a cognitive mapping model; obtaining motor operation parameters to calculate a multi-granularity feature representation, predicting a state trajectory through an autoregressive model, performing multi-step decision solving and generating a motor control strategy sequence, determining the similarity between a current strategy and a reference strategy, constructing a hybrid strategy evaluation mechanism and performing evaluation, determining an optimal control strategy and performing motor control; obtaining real-time operation data and calculating deviation data and performing data analysis, constructing a multi-dimensional evaluation system, determining abnormal points and abnormal intervals, generating an abnormal analysis report, updating a cognitive mapping model and a strategy evaluation model, and outputting an optimal control strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of motor control, and in particular to a control method and system for a motor controller. Background Art

[0002] Motor control technology is widely used in industrial automation, robotics, electric vehicles and other fields. It is of great significance to improve production efficiency, save energy and reduce consumption, and ensure system reliability and safety. Traditional motor control methods are mainly based on classical control theories such as PID control, fuzzy control, and sliding mode control. By establishing a motor mathematical model and designing controller parameters, the motor's speed, torque, position and other control targets can be achieved.

[0003] However, traditional control methods are not only difficult to adapt to changes in motor parameters and external disturbances, lack self-learning and self-adaptation capabilities, and are difficult to cope with complex working conditions and dynamically changing control requirements, but also have the problem of being unable to fully mine and utilize the large amount of data generated during the operation of the motor, and unable to achieve data-driven intelligent control and optimization;

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a control method and system for a motor controller, which can at least solve some of the problems existing in the prior art.

[0006] According to a first aspect of an embodiment of the present invention, there is provided a control method for a motor controller, comprising:

[0007] The multi-dimensional parameters corresponding to the current motor are collected by a motor data collection unit, and the environmental parameters are collected in combination with a pre-set environmental sensor array. The environmental parameters are combined with the multi-dimensional parameters by an adaptive data fusion method based on active learning to obtain an operation data stream, and the operation data stream is adaptively extracted with spatiotemporal features to obtain a spatiotemporal feature representation and model the motor state data distribution based on a generative adversarial network. The state features corresponding to the spatiotemporal feature representation are learned by a generator, and the probability distribution of the state features is determined by a variational autoencoder and a maximum marginal likelihood estimation algorithm, and a state transfer equation is constructed to obtain a cognitive mapping model.

[0008] The motor operation parameters are obtained through the motor data acquisition unit and added to the cognitive mapping model, the multi-granularity feature representation corresponding to the current operation state is calculated through forward reasoning, the state trajectory is predicted by the autoregressive model in combination with the historical state evolution trend obtained in advance, based on the state trajectory, a multi-step decision solution is performed in combination with an adaptive dynamic programming algorithm and a motor control strategy sequence is generated, a reference strategy set is constructed, a strategy evaluation model is constructed based on contrastive learning and the similarity between the current strategy and the reference strategy is determined, a hybrid strategy evaluation mechanism is constructed in combination with long-term cumulative rewards and short-term instantaneous effects, and the control strategy sequence is evaluated, the optimal control strategy is determined and parsed into a control instruction diagram for motor control;

[0009] During the control process, real-time operation data is acquired and deviation data is calculated, and data analysis is performed in combination with a distributed active inference algorithm to obtain deviation analysis results. The real-time operation data is mapped into performance indicators based on a knowledge graph algorithm and a multi-dimensional evaluation system is constructed. According to the deviation analysis results, anomaly points and abnormal intervals are determined in combination with an anomaly detection method based on confidence learning, and an anomaly analysis report is generated in combination with a pre-set expert knowledge base. Based on the anomaly analysis report, the cognitive mapping model and the strategy evaluation model are updated in combination with a meta-learning algorithm. The update is repeated until the preset maximum number of iterations is reached and the optimal control strategy generated by the last iteration is output.

[0010] In an optional embodiment,

[0011] The multi-dimensional parameters corresponding to the current motor are collected by the motor data collection unit, and the environmental parameters are collected in combination with the pre-set environmental sensor array. The environmental parameters are combined with the multi-dimensional parameters by an adaptive data fusion method based on active learning to obtain an operation data stream, and the operation data stream is adaptively extracted with spatiotemporal features to obtain a spatiotemporal feature representation and model the motor state data distribution based on a generative adversarial network. The state features corresponding to the spatiotemporal feature representation are learned by a generator, and the probability distribution of the state features is determined by a variational autoencoder and a maximum marginal likelihood estimation algorithm. The state transfer equation is constructed and the cognitive mapping model is obtained, including:

[0012] The multi-dimensional parameters corresponding to the current motor are collected by a pre-set motor data collection unit, wherein the multi-dimensional parameters include speed, current, voltage, vibration frequency, power factor and load torque; the ambient temperature, humidity and air pressure are collected as environmental parameters by a pre-set environmental sensor array; the credibility and relevance of the motor data collection unit and the environmental sensor array are evaluated by an adaptive data fusion method based on active learning; a data source evaluation result is generated based on the credibility and relevance; the weight coefficients corresponding to the motor data collection unit and the environmental sensor array are dynamically adjusted based on the data source evaluation result; the multi-dimensional parameters and the environmental parameters are adaptively fused by weighted fusion to obtain the operation data stream;

[0013] For the running data stream, in the time dimension, the time series is modeled by a gated recurrent unit, the dependency of the time series is determined by combining the update gate and the reset gate, and the attention coefficient corresponding to different time segments is adjusted in combination with the attention mechanism to extract the dynamic change pattern of the time series feature. In the spatial dimension, feature extraction is performed by a graph convolutional neural network, the data corresponding to the current motor is used as a node in the graph structure and the relationship between different data is used as an edge to construct a motor parameter association graph, the neighborhood information of each node is aggregated through a graph convolution operation to obtain the structural feature corresponding to each node, and the spatial association pattern of the data corresponding to the current motor is obtained at the same time, the spatial association pattern is dynamically updated through a time-varying graph convolutional network to obtain the spatial feature, the time series feature and the spatial feature are mapped to a common semantic space and semantically aligned, and then feature fusion is performed to obtain a spatiotemporal feature representation;

[0014] The spatiotemporal feature representation is added as input to the generator network of the generative adversarial network. The generator network stacks multiple transposed convolutional layers and upsampling layers through a deep convolutional neural network structure to map the low-dimensional random noise vector to a high-dimensional state feature space. The motor operating conditions are input as additional conditions in combination with the conditional generation mechanism to generate state features corresponding to the spatiotemporal feature representation. The discriminator in the generative adversarial network maps the state features to binary classification probabilities by stacking multiple convolutional layers and downsampling layers. In combination with a multi-scale discrimination strategy, multiple discriminators with different receptive fields are set to evaluate the authenticity of the state features, and adversarial training is performed and a variational autoencoder is set. The state feature is mapped to the latent variable space through the variational autoencoder, the posterior distribution of the latent variables in the latent variable space is learned, the state feature is reconstructed from the latent variable space through the decoder network and the prior distribution under the state feature is learned, based on the maximum marginal likelihood estimation algorithm, the posterior distribution is used as the conditional probability distribution of the latent variable, the prior distribution is used as the marginal probability distribution of the latent variable, the latent variable is represented as a combination of deterministic variables and noise in combination with the reparameterization technique to complete low variance sampling, and the geometric distance of the latent variable in the manifold space is minimized in combination with the manifold regularization method to obtain the probability distribution corresponding to the state feature;

[0015] The relationship equation between state variables is determined by a gray box modeling method based on physical laws, and the state transfer equation is determined by a deep neural network in combination with a data-driven black box modeling method. The cognitive mapping model is constructed based on the state transfer equation.

[0016] In an optional embodiment,

[0017] The objective function of the variational autoencoder is shown in the following formula:

[0018] ;

[0019] in, represents the weight of the generator network, represents the weight of the variational encoder, Indicates state characteristics, Indicates i hidden variable samples, For K The expectation of importance sampling samples, which means calculating the The average value of Express expectations, represents the number of samples for importance sampling, Indicates that in a given state feature In the case of Samples The joint distribution of Indicates state characteristics and latent variable samples The joint probability distribution of is a hidden variable The posterior distribution of The hidden variable The probability distribution of .

[0020] In an optional embodiment,

[0021] The motor operation parameters are obtained by the motor data acquisition unit and added to the cognitive mapping model, the multi-granularity feature representation corresponding to the current operation state is calculated by forward reasoning, the state trajectory is predicted by the autoregressive model in combination with the historical state evolution trend obtained in advance, based on the state trajectory, a multi-step decision solution is performed in combination with an adaptive dynamic programming algorithm and a motor control strategy sequence is generated, a reference strategy set is constructed, a strategy evaluation model is constructed according to contrastive learning and the similarity between the current strategy and the reference strategy is determined, a hybrid strategy evaluation mechanism is constructed in combination with long-term cumulative rewards and short-term instantaneous effects, and the control strategy sequence is evaluated, and the optimal control strategy is determined and parsed into a control instruction diagram for motor control, including:

[0022] The motor operating parameters corresponding to the current motor are obtained through the motor data acquisition unit and added to the cognitive mapping model, the cognitive mapping model learns the semantic features corresponding to the operating state of the current motor and obtains the low-level features, intermediate features and high-level features corresponding to the motor operating parameters through forward propagation calculation, and the calculated features are combined to generate the multi-granularity feature representation;

[0023] Obtain historical operating data of the current motor and construct a historical data set, set a time window and a sliding step size of the autoregressive model, divide the historical data set into a training set and a test set, use the training set to train the autoregressive model, optimize the hyperparameters of the autoregressive model in combination with the real labels of the training set until the model converges, add the elements in the test set as input to the autoregressive model, and obtain the historical state evolution trend;

[0024] Based on the historical state evolution trend, the multi-granularity feature representation is added to the autoregressive model and the multi-granularity feature representation corresponding to each time step is predicted, and the state trajectory is generated by combining the prediction results. Based on the state trajectory, a Markov decision process is constructed in combination with the current motor control objectives and constraints. A multi-step decision problem is constructed by defining the state space, action space and reward function. A value function approximator is set based on an adaptive dynamic programming algorithm and the value function estimate is calculated using the state and action corresponding to the current motor as input. The value function is updated through a value iteration algorithm until convergence. A policy function approximator is set to represent the policy function. number, taking the state corresponding to the current motor as input, calculating the action probability distribution corresponding to the output and updating the policy function by maximizing the expected cumulative reward, initializing the value function approximator and the policy function approximator and generating trajectory data according to the current policy function, updating the parameters of the value function approximator based on the trajectory data, updating the parameters of the policy function approximator according to the updated value function combined with the policy gradient algorithm, repeatedly updating to obtain the optimal value function and the optimal policy function after convergence, based on the optimal policy function, performing forward reasoning according to the state trajectory corresponding to the current motor, generating candidate control strategies, and repeatedly generating to obtain the motor control strategy sequence;

[0025] A strategy evaluation model is constructed based on contrastive learning and a reference strategy set is set. For the current strategy in the motor control strategy sequence, a reference strategy is selected and the similarity is calculated. The strategy representation is learned by maximizing the similarity of positive strategy pairs and minimizing the similarity of negative strategy pairs. A hybrid strategy evaluation mechanism is constructed by combining long-term cumulative rewards and short-term instantaneous effects. A comprehensive evaluation result is obtained by weighted fusion of the value function corresponding to the long-term cumulative rewards and the instant reward calculation. Based on the comprehensive evaluation result, an optimal control strategy is selected and the optimal control strategy is parsed into a control instruction. Based on the control instruction, a control instruction graph is generated and motor control is performed.

[0026] In an optional embodiment,

[0027] The parameters of the policy function approximator are updated according to the updated value function combined with the policy gradient algorithm as shown in the following formula:

[0028] ;

[0029] in, Indicates the parameters of the current strategy, Express expectations, Indicates that the current policy is in state Next select action The probability of Indicates that the old policy is in state Next select action The probability of represents the clipping function, A hyperparameter representing the clipping range, Represents the advantage function, which is used to measure the state Next Action Advantages, Represents the time step.

[0030] In an optional embodiment,

[0031] During the control process, real-time operation data is acquired and deviation data is calculated, and data analysis is performed in combination with a distributed active inference algorithm to obtain deviation analysis results. The real-time operation data is mapped to performance indicators based on a knowledge graph algorithm and a multi-dimensional evaluation system is constructed. According to the deviation analysis results, anomaly points and abnormal intervals are determined in combination with an anomaly detection method based on confidence learning, and an anomaly analysis report is generated in combination with a pre-set expert knowledge base. Based on the anomaly analysis report, the cognitive mapping model and the strategy evaluation model are updated in combination with a meta-learning algorithm, and the update is repeated until the preset maximum number of iterations is reached and the optimal control strategy generated by the last iteration is output, including:

[0032] Acquire real-time operation data corresponding to the current motor during the execution of motor control by the motor data acquisition unit, wherein the real-time operation data includes current, voltage, speed and temperature, compare the real-time operation data with pre-set standard operation data, calculate deviation data and construct an inference rule library through a distributed active inference algorithm, wherein the inference rule library includes abnormal pattern recognition rules and data correlation analysis rules, select an inference path based on the inference rule library and perform data analysis to obtain a deviation analysis result corresponding to the deviation data;

[0033] Based on the knowledge graph algorithm, the real-time operation data is mapped into multiple dimensions of motor performance and a multi-dimensional evaluation index system is constructed, a comprehensive evaluation function corresponding to the motor performance is constructed based on the multi-dimensional evaluation index system, a judgment matrix is ​​constructed by the hierarchical analysis method, the eigenvector corresponding to the maximum eigenvalue of the judgment matrix is ​​solved, and the weight coefficient corresponding to each dimension is obtained, and the real-time operation data is substituted into the comprehensive evaluation function to obtain the comprehensive evaluation value corresponding to the current motor;

[0034] Through the anomaly detection method based on confidence learning, an anomaly point determination function is constructed and the anomaly threshold is calculated in combination with the confidence interval estimation algorithm. For the real-time operation data of each time step, the anomaly value is calculated through the anomaly point determination function and screened through the anomaly threshold. If the anomaly value is greater than the anomaly threshold, the real-time operation data of the current time step is considered to be an anomaly point. The screening is repeated until all time steps are screened. For the anomaly point, the distance between each anomaly point and the point corresponding to other real-time operation data is calculated, a distance matrix is ​​generated, and the anomaly points are clustered to obtain the anomaly interval. The cause of the anomaly is analyzed and the impact is evaluated in combination with a preset expert knowledge base, and a corresponding anomaly analysis report is generated;

[0035] The abnormal analysis report, the deviation analysis result and the comprehensive evaluation value are used as inputs of a meta-learning algorithm. An optimization-based meta-learning algorithm is used, with the cognitive mapping model and the strategy evaluation model as base learners and adaptively updated. A maximum number of iterations is set and the cognitive mapping model and the strategy evaluation model are repeatedly updated until the maximum number of iterations is reached, and the optimal control strategy generated in the last iteration is used as the control instruction of the current motor.

[0036] In an optional embodiment,

[0037] The comprehensive evaluation function corresponding to the motor performance is constructed based on the multi-dimensional evaluation index system as shown in the following formula:

[0038] ;

[0039] in, represents the evaluation value, represents the product operation, Represents the total number of dimensions, Indicates m The nonlinear evaluation function in dimensions, Indicates m The weight coefficient of each dimension, Indicates m Dimensions and n The weight coefficient of the interaction between dimensions, represents the adjustment parameter, Indicates n The nonlinear evaluation function in dimensions, Indicates m Specific indicators related to each dimension, Indicates n Specific indicators related to each dimension, Indicates m Dimensions and n The interaction function between the dimensions.

[0040] According to a second aspect of an embodiment of the present invention, a control system for a motor controller is provided, comprising:

[0041] The first unit is used to collect multi-dimensional parameters corresponding to the current motor through a motor data collection unit, collect environmental parameters in combination with a preset environmental sensor array, combine the environmental parameters with the multi-dimensional parameters through an adaptive data fusion method based on active learning to obtain an operation data stream, perform adaptive spatiotemporal feature extraction on the operation data stream, obtain a spatiotemporal feature representation, and model the motor state data distribution based on a generative adversarial network, learn the state features corresponding to the spatiotemporal feature representation through a generator, and determine the probability distribution of the state features through a variational autoencoder and a maximum marginal likelihood estimation algorithm, construct a state transfer equation, and obtain a cognitive mapping model;

[0042] The second unit is used to obtain the motor operation parameters through the motor data acquisition unit and add them to the cognitive mapping model, calculate the multi-granularity feature representation corresponding to the current operation state through forward reasoning, and predict the state trajectory through the autoregressive model in combination with the historical state evolution trend obtained in advance, based on the state trajectory, combine the adaptive dynamic programming algorithm to perform multi-step decision solving and generate a motor control strategy sequence, build a reference strategy set, build a strategy evaluation model based on contrastive learning and determine the similarity between the current strategy and the reference strategy, build a hybrid strategy evaluation mechanism in combination with long-term cumulative rewards and short-term instantaneous effects, and evaluate the control strategy sequence, determine the optimal control strategy and parse it into a control instruction diagram for motor control;

[0043] The third unit is used to obtain real-time operation data and calculate deviation data during the control process and perform data analysis in combination with a distributed active inference algorithm to obtain deviation analysis results, map the real-time operation data into performance indicators based on a knowledge graph algorithm and build a multi-dimensional evaluation system, determine abnormal points and abnormal intervals based on the deviation analysis results in combination with an anomaly detection method based on confidence learning, generate an anomaly analysis report in combination with a pre-set expert knowledge base, update the cognitive mapping model and the strategy evaluation model based on the anomaly analysis report in combination with a meta-learning algorithm, repeat the update until the preset maximum number of iterations is reached and output the optimal control strategy generated by the last iteration.

[0044] According to a third aspect of the embodiments of the present invention,

[0045] An electronic device is provided, comprising:

[0046] processor;

[0047] a memory for storing processor-executable instructions;

[0048] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0049] A fourth aspect of the embodiments of the present invention is:

[0050] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0051] In the present invention, the multi-dimensional motor parameters and environmental parameters are combined into an operation data stream through an active learning adaptive data fusion method, so as to realize the effective fusion of heterogeneous data, improve data quality and integrity, and provide a high-quality data foundation for subsequent analysis and modeling. The motor state data distribution is modeled based on a generative adversarial network, and the probability distribution of state characteristics is learned through a variational autoencoder and a maximum marginal likelihood estimation algorithm. The state transition equation is constructed to form a cognitive mapping model, which can effectively characterize the dynamic evolution law of the motor state and provide support for state prediction and decision optimization. The real-time operation parameters are added to the cognitive mapping model, and the multi-granular feature representation of the current operation state is calculated through forward reasoning. Combined with the historical state evolution trend, the cognitive mapping model is formed. By predicting the state trajectory through the autoregressive model, the future evolution of the motor state can be predicted at different granularities, and multi-scale and multi-step prediction information can be provided for control decisions, thereby improving the foresight and robustness of the decisions, comprehensively considering the long-term and short-term effects of the strategy, balancing local optimization and global optimization, selecting the optimal control strategy, and improving the rationality and effectiveness of the control decision. In summary, the present invention makes full use of the motor operation data, adaptively learns the dynamic evolution law of the motor state, generates the optimal control strategy, realizes the intelligent, adaptive and continuous optimization of the motor control, can effectively deal with the nonlinearity, time-varying and uncertainty of the motor system, improves the efficiency, reliability and safety of the motor operation, and provides an innovative solution for the field of motor control. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of a control method for a motor controller according to an embodiment of the present invention;

[0053] Figure 2 The figure is a schematic diagram of the structure of a control system for a motor controller according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0056] Figure 1 FIG. 1 is a flow chart of a control method for a motor controller according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0057] S1. Collect multi-dimensional parameters corresponding to the current motor through the motor data acquisition unit, collect environmental parameters in combination with the pre-set environmental sensor array, combine the environmental parameters with the multi-dimensional parameters through an adaptive data fusion method based on active learning to obtain an operation data stream, perform adaptive spatiotemporal feature extraction on the operation data stream, obtain a spatiotemporal feature representation, and model the motor state data distribution based on a generative adversarial network, learn the state features corresponding to the spatiotemporal feature representation through a generator, and determine the probability distribution of the state features through a variational autoencoder and a maximum marginal likelihood estimation algorithm, construct a state transfer equation, and obtain a cognitive mapping model;

[0058] The environmental sensor array is a group of sensors deployed in a specific area for monitoring and collecting physical or chemical parameters related to the environment, such as temperature, humidity, air pressure, gas concentration, light intensity, etc. The motor data acquisition unit is a hardware system for monitoring and recording the running status of the motor. The adaptive data fusion method based on active learning combines active learning and data fusion technology, and adjusts the data fusion strategy adaptively to improve the accuracy and efficiency of data fusion. The operating data stream refers to the continuous data stream generated in real time during the operation of the system. The adaptive spatiotemporal feature extraction is a method for automatically extracting important features from spatiotemporal data according to the temporal and spatial characteristics of the data. Adaptively select appropriate feature extraction strategies to capture spatiotemporal patterns and changing trends in the data. The generative adversarial network is a deep learning model composed of two adversarial neural networks, a generator and a discriminator. The variational autoencoder is a generative model that generates new samples by learning the latent variable distribution of the data. The maximum marginal likelihood estimation algorithm aims to maximize the marginal likelihood by optimizing the model parameters, that is, the total probability of the observed data under the model. The state transfer equation describes the rule or probability of the system transferring from one state to another. The cognitive mapping model is a model that represents cognitive processes or cognitive structures, which is used to describe and simulate human decision-making processes, cognitive paths or knowledge structures.

[0059] In an optional embodiment,

[0060] The multi-dimensional parameters corresponding to the current motor are collected by the motor data collection unit, and the environmental parameters are collected in combination with the pre-set environmental sensor array. The environmental parameters are combined with the multi-dimensional parameters by an adaptive data fusion method based on active learning to obtain an operation data stream, and the operation data stream is adaptively extracted with spatiotemporal features to obtain a spatiotemporal feature representation and model the motor state data distribution based on a generative adversarial network. The state features corresponding to the spatiotemporal feature representation are learned by a generator, and the probability distribution of the state features is determined by a variational autoencoder and a maximum marginal likelihood estimation algorithm. The state transfer equation is constructed and the cognitive mapping model is obtained, including:

[0061] The multi-dimensional parameters corresponding to the current motor are collected by a pre-set motor data collection unit, wherein the multi-dimensional parameters include speed, current, voltage, vibration frequency, power factor and load torque; the ambient temperature, humidity and air pressure are collected as environmental parameters by a pre-set environmental sensor array; the credibility and relevance of the motor data collection unit and the environmental sensor array are evaluated by an adaptive data fusion method based on active learning; a data source evaluation result is generated based on the credibility and relevance; the weight coefficients corresponding to the motor data collection unit and the environmental sensor array are dynamically adjusted based on the data source evaluation result; the multi-dimensional parameters and the environmental parameters are adaptively fused by weighted fusion to obtain the operation data stream;

[0062] For the running data stream, in the time dimension, the time series is modeled by a gated recurrent unit, the dependency of the time series is determined by combining the update gate and the reset gate, and the attention coefficient corresponding to different time segments is adjusted in combination with the attention mechanism to extract the dynamic change pattern of the time series feature. In the spatial dimension, feature extraction is performed by a graph convolutional neural network, the data corresponding to the current motor is used as a node in the graph structure and the relationship between different data is used as an edge to construct a motor parameter association graph, the neighborhood information of each node is aggregated through a graph convolution operation to obtain the structural feature corresponding to each node, and the spatial association pattern of the data corresponding to the current motor is obtained at the same time, the spatial association pattern is dynamically updated through a time-varying graph convolutional network to obtain the spatial feature, the time series feature and the spatial feature are mapped to a common semantic space and semantically aligned, and then feature fusion is performed to obtain a spatiotemporal feature representation;

[0063] The spatiotemporal feature representation is added as input to the generator network of the generative adversarial network. The generator network stacks multiple transposed convolutional layers and upsampling layers through a deep convolutional neural network structure to map the low-dimensional random noise vector to a high-dimensional state feature space. The motor operating conditions are input as additional conditions in combination with the conditional generation mechanism to generate state features corresponding to the spatiotemporal feature representation. The discriminator in the generative adversarial network maps the state features to binary classification probabilities by stacking multiple convolutional layers and downsampling layers. In combination with a multi-scale discrimination strategy, multiple discriminators with different receptive fields are set to evaluate the authenticity of the state features, and adversarial training is performed and a variational autoencoder is set. The state feature is mapped to the latent variable space through the variational autoencoder, the posterior distribution of the latent variables in the latent variable space is learned, the state feature is reconstructed from the latent variable space through the decoder network and the prior distribution under the state feature is learned, based on the maximum marginal likelihood estimation algorithm, the posterior distribution is used as the conditional probability distribution of the latent variable, the prior distribution is used as the marginal probability distribution of the latent variable, the latent variable is represented as a combination of deterministic variables and noise in combination with the reparameterization technique to complete low variance sampling, and the geometric distance of the latent variable in the manifold space is minimized in combination with the manifold regularization method to obtain the probability distribution corresponding to the state feature;

[0064] The relationship equation between state variables is determined by a gray box modeling method based on physical laws, and the state transfer equation is determined by a deep neural network in combination with a data-driven black box modeling method. The cognitive mapping model is constructed based on the state transfer equation.

[0065] The gated recurrent unit is a variant of a recurrent neural network, designed to solve the gradient vanishing problem of recurrent neural networks in long sequence data. The graph convolutional neural network is a neural network model for processing graph structure data. The spatial association pattern refers to the association relationship or common change trend between multiple data points or variables in spatial data. The time-varying graph convolutional network is a neural network model that combines time series data and graph structure data. The common semantic space is a representation of cross-modal or multi-domain data in the same space. The transposed convolution layer (also called deconvolution layer) is a convolution operation for upsampling. The low-dimensional random noise vector is the input in the generative model for generating high-dimensional data. The conditional generation mechanism refers to the generation model introducing additional conditional information (such as labels, categories) when generating data, so that the generation process is controlled by the condition. The binary classification probability refers to the probability estimate of a sample belonging to a certain category in the binary classification problem. The multi-scale discrimination strategy refers to the simultaneous consideration of different scales (such as space, time) in the discrimination task. The invention discloses a method for optimizing the features of a model under the time or frequency scale to improve the discrimination ability of the model. The latent variable space refers to the space composed of potential variables that are not directly observed in the model. The reparameterization technique is a method used in variational autoencoders (VAE), which enables the model to directly optimize the distribution parameters of latent variables through gradient descent. The low variance sampling is a sampling method used in algorithms such as Markov chain Monte Carlo or particle filtering, which aims to reduce the variance in the sampling process, thereby improving the stability and accuracy of the estimation. The manifold regularization method is a regularization technique that improves the generalization ability of the model by assuming that the data is distributed on a low-dimensional manifold and learning on the manifold. The gray box modeling combines a model based on physical laws (white box) and a data-driven model (black box). Part of the model structure or parameters are determined by physical laws, while the rest are estimated through data. The black box modeling relies entirely on data to build a system model without considering the internal physical mechanism. The model structure and parameters are usually learned from input and output data through machine learning or statistical methods.

[0066] The multi-dimensional parameters corresponding to the current motor, including speed, current, voltage, vibration frequency, power factor and load torque, are collected through a pre-set motor data acquisition unit. The ambient temperature, humidity and air pressure are collected as environmental parameters through a pre-set environmental sensor array. The credibility and relevance of the motor data acquisition unit and the environmental sensor array are evaluated through an adaptive data fusion method based on active learning, and a data source evaluation result is generated. The weight coefficients corresponding to the motor data acquisition unit and the environmental sensor array are dynamically adjusted based on the data source evaluation result. The multi-dimensional parameters and the environmental parameters are adaptively fused through weighted fusion to obtain an operation data flow. For example, it is assumed that the motor data acquisition unit The collected data are: speed 1500rpm, current 10A, voltage 220V, vibration frequency 50Hz, power factor 0.8, load torque 20N·m. The data collected by the environmental sensor array are: temperature 25°C, humidity 60%, air pressure 101kPa. The credibility of the motor data acquisition unit is evaluated to be 0.9 and the correlation is 0.8 through the adaptive data fusion method. The credibility of the environmental sensor array is 0.7 and the correlation is 0.6. According to the evaluation results, the weight coefficient of the motor data acquisition unit is adjusted to 0.6, and the weight coefficient of the environmental sensor array is adjusted to 0.4. The multi-dimensional parameters and environmental parameters are weighted and fused according to the weights to obtain the operation data flow.

[0067] In the time dimension of the running data flow, the time series is modeled through the gated recurrent unit, the dependency of the time series is determined by combining the update gate and the reset gate, and the attention coefficient corresponding to different time segments is adjusted in combination with the attention mechanism to extract the dynamic change pattern of the time series features. In the spatial dimension, feature extraction is performed through the graph convolutional neural network, the data corresponding to the current motor is used as the node in the graph structure, and the relationship between different data is used as the edge to construct the motor parameter association graph. The neighborhood information of each node is aggregated through the graph convolution operation to obtain the structural features corresponding to each node, and the spatial association pattern of the data corresponding to the current motor is obtained. The spatial association pattern is dynamically updated through the time-varying graph convolutional network to obtain the spatial features. The temporal features and spatial features are mapped to the common semantic space and semantically aligned before feature fusion. Obtaining spatiotemporal feature representation. For example, assuming that the running data stream contains 10 time segments, each time segment contains 6 multidimensional parameters and 3 environmental parameters, the time series is modeled by a gated recurrent unit to obtain the hidden state of each time segment, and the attention mechanism is combined to calculate the attention coefficient of each time segment to extract the dynamic change pattern of the time series feature. In the spatial dimension, the 9 parameters are used as nodes in the graph structure, and the edges are constructed according to the physical relationship between the parameters to obtain the motor parameter association graph. The neighborhood information of each node is aggregated through the graph convolution operation to obtain the structural characteristics and spatial association pattern of each node. The spatial association pattern is dynamically updated through the time-varying graph convolution network to obtain the spatial feature. The time series feature and the spatial feature are mapped to the common semantic space and semantically aligned, and the spatiotemporal feature representation is obtained through feature fusion.

[0068] The spatiotemporal feature representation is added as input to the generator network of the generative adversarial network. The generator network stacks multiple transposed convolutional layers and upsampling layers through a deep convolutional neural network structure to map the low-dimensional random noise vector to a high-dimensional state feature space. The motor operating conditions are input as additional conditions in combination with the conditional generation mechanism to generate state features corresponding to the spatiotemporal feature representation. The discriminator in the generative adversarial network maps the state features to binary classification probabilities by stacking multiple convolutional layers and downsampling layers. Combined with the multi-scale discrimination strategy, multiple discriminators with different receptive fields are set to evaluate the true value of the state features. Reality, adversarial training, setting up variational autoencoders, mapping state features to latent variable space through variational autoencoders, learning the posterior distribution of latent variables in latent variable space, reconstructing state features from latent variable space through decoder network and learning the prior distribution under state features, based on the maximum marginal likelihood estimation algorithm, using the posterior distribution as the conditional probability distribution of latent variables, using the prior distribution as the marginal probability distribution of latent variables, combining reparameterization techniques to represent latent variables as a combination of deterministic variables and noise to complete low variance sampling, and combining manifold regularization methods to minimize the latent variables on the manifold. The geometric distance in space is used to obtain the probability distribution corresponding to the state feature. For example, assuming that the dimension of the spatiotemporal feature representation is 128, the input of the generator network is a 100-dimensional random noise vector and a 10-dimensional motor operating condition vector. The random noise vector is mapped to the 128-dimensional state feature space through 5 transposed convolution layers and upsampling layers. At the same time, the motor operating condition is input as an additional condition to generate the state feature corresponding to the spatiotemporal feature representation. The discriminator network maps the state feature to the binary classification probability through 4 convolution layers and downsampling layers, and sets 3 discriminators with different receptive fields to evaluate the state. To ensure the authenticity of the features, adversarial training is performed. The state features are mapped to a 20-dimensional latent variable space through a variational autoencoder, and the posterior distribution of the latent variables is learned. The state features are reconstructed from the latent variable space through a decoder network and the prior distribution under the state features is learned. The posterior distribution is used as the conditional probability distribution of the latent variables, and the prior distribution is used as the marginal probability distribution of the latent variables. The latent variables are represented as a combination of deterministic variables and noise by combining the reparameterization technique to complete low-variance sampling. The geometric distance of the latent variables in the manifold space is minimized by combining the manifold regularization method to obtain the probability distribution corresponding to the state features.

[0069] The relationship equation between state variables is determined by a gray box modeling method based on physical laws, and the state transfer equation is determined by a deep neural network in combination with a data-driven black box modeling method. A cognitive mapping model is constructed based on the state transfer equation. By way of example, it is assumed that the state variables of the motor system include speed, current, voltage, vibration frequency, power factor and load torque. The relationship equation between state variables is determined by physical laws, such as the relationship between voltage and current satisfies Ohm's law, and the relationship between power factor and voltage and current satisfies the power triangle law. In combination with a data-driven method, a deep neural network is used to learn the state transfer equation, such as the relationship between state variables such as speed, current, voltage at the current moment and the state variables at the next moment. The relationship equation between state variables and the state transfer equation are combined to construct a cognitive mapping model of the motor system, which is used to describe the dynamic behavior and evolution law of the motor system.

[0070] In this embodiment, multi-source heterogeneous data are collected through the motor data acquisition unit and the environmental sensor array, and the credibility and correlation are evaluated through the adaptive data fusion method, and the weight coefficient is dynamically adjusted to realize the adaptive fusion of data, which can effectively deal with the uncertainty and redundancy of the data, improve the data quality and reliability, provide high-quality data support for subsequent analysis, fully explore the spatiotemporal dependency of the data, capture the dynamic evolution law of the motor system, and provide rich feature representation for state evaluation and fault diagnosis. The state characteristics corresponding to the spatiotemporal feature representation are generated by the generative adversarial network, and the motor operating condition information is introduced in combination with the conditional generation mechanism. The multi-scale discrimination strategy is used to evaluate The authenticity of state features, at the same time, the introduction of variational autoencoders to achieve low variance sampling and geometric constraints can generate high-quality and diversified state features, enhance the representation and generalization capabilities of features, take into account physical knowledge and data information, establish causal relationships and dynamic mappings between state variables, and form a comprehensive cognition and understanding of the behavior of the motor system, providing an explainable decision-making basis for fault diagnosis, health assessment and predictive maintenance. In summary, this embodiment realizes comprehensive monitoring and modeling of the motor system through multi-source data fusion, spatiotemporal feature extraction, generative modeling and cognitive mapping technologies, providing a novel and effective technical route for the monitoring, diagnosis and optimization of the motor system.

[0071] In an optional embodiment,

[0072] The objective function of the variational autoencoder is shown in the following formula:

[0073] ;

[0074] in, represents the weight of the generator network, represents the weight of the variational encoder, Indicates state characteristics, Indicates i hidden variable samples, For K The expectation of importance sampling samples, which means calculating the The average value of Express expectations, represents the number of samples for importance sampling, Indicates that in a given state feature In the case of Samples The joint distribution of Indicates state characteristics and latent variable samples The joint probability distribution of is a hidden variable The posterior distribution of The hidden variable The probability distribution of .

[0075] In this embodiment, high-dimensional state features are mapped to low-dimensional latent variable space through variational autoencoders, and learning the probability distribution of latent variables can extract the intrinsic structure and semantic information of state features, obtain a more compact and effective feature representation, reduce feature redundancy and noise, and learn the generation process of state features, capture the intrinsic distribution and statistical laws of features, and provide probabilistic model support for subsequent feature generation and anomaly detection. Given the state features, the conditional probability distribution of latent variables can be inferred to achieve conditional generation and inference of latent variables. In summary, this embodiment can obtain a more compact and effective state feature representation, capture the intrinsic distribution and statistical laws of features, achieve feature generation and inference, provide a probabilistic interpretation framework, and provide new technical means and theoretical support for tasks such as state monitoring, fault diagnosis, and anomaly detection.

[0076] S2. Obtain motor operating parameters through the motor data acquisition unit and add them to the cognitive mapping model, calculate the multi-granularity feature representation corresponding to the current operating state through forward reasoning, and predict the state trajectory through the autoregressive model in combination with the historical state evolution trend acquired in advance. Based on the state trajectory, combine the adaptive dynamic programming algorithm to perform multi-step decision solving and generate a motor control strategy sequence, build a reference strategy set, build a strategy evaluation model based on contrastive learning and determine the similarity between the current strategy and the reference strategy, build a hybrid strategy evaluation mechanism in combination with long-term cumulative rewards and short-term instantaneous effects, and evaluate the control strategy sequence, determine the optimal control strategy and parse it into a control instruction diagram for motor control;

[0077] The multi-granularity feature representation refers to the extraction and representation of data features at different scales or levels. The historical state evolution trend refers to the trend of change of the system or variable over time, which is usually used to analyze and predict future states. The autoregressive model is a time series model that assumes that the current observation value can be represented by a linear combination of several previous observation values. The state trajectory refers to the path or trajectory formed by the system state in the process of changing over time. The adaptive dynamic programming is an algorithm for solving optimal control problems. It optimizes the long-term cumulative return by adaptively learning the dynamic model and strategy of the system. The multi-step decision solution involves making a series of interrelated decisions within a given time range to optimize the final goal. The motor control strategy sequence refers to A series of strategies or command sequences formulated to control the movement and behavior of the motor. The contrastive learning is a self-supervised learning method that learns effective feature representation by maximizing the similarity between similar samples and minimizing the similarity between non-similar samples. The long-term cumulative reward is a concept in reinforcement learning, which refers to the sum of the rewards obtained by an agent in multiple time steps in the future. The short-term transient effect refers to the effect or influence of an action or event in a short period of time in decision-making or control. The hybrid strategy evaluation mechanism combines the evaluation methods of multiple strategies or models to comprehensively consider the advantages and disadvantages of different strategies and finally select the optimal strategy. The control instruction graph is a graph structure used to represent the instruction sequence in the control system and the dependency relationship between different instructions.

[0078] In an optional embodiment,

[0079] The motor operation parameters are obtained by the motor data acquisition unit and added to the cognitive mapping model, the multi-granularity feature representation corresponding to the current operation state is calculated by forward reasoning, the state trajectory is predicted by the autoregressive model in combination with the historical state evolution trend obtained in advance, based on the state trajectory, a multi-step decision solution is performed in combination with an adaptive dynamic programming algorithm and a motor control strategy sequence is generated, a reference strategy set is constructed, a strategy evaluation model is constructed according to contrastive learning and the similarity between the current strategy and the reference strategy is determined, a hybrid strategy evaluation mechanism is constructed in combination with long-term cumulative rewards and short-term instantaneous effects, and the control strategy sequence is evaluated, and the optimal control strategy is determined and parsed into a control instruction diagram for motor control, including:

[0080] The motor operating parameters corresponding to the current motor are obtained through the motor data acquisition unit and added to the cognitive mapping model, the cognitive mapping model learns the semantic features corresponding to the operating state of the current motor and obtains the low-level features, intermediate features and high-level features corresponding to the motor operating parameters through forward propagation calculation, and the calculated features are combined to generate the multi-granularity feature representation;

[0081] Obtain historical operating data of the current motor and construct a historical data set, set a time window and a sliding step size of the autoregressive model, divide the historical data set into a training set and a test set, use the training set to train the autoregressive model, optimize the hyperparameters of the autoregressive model in combination with the real labels of the training set until the model converges, add the elements in the test set as input to the autoregressive model, and obtain the historical state evolution trend;

[0082] Based on the historical state evolution trend, the multi-granularity feature representation is added to the autoregressive model and the multi-granularity feature representation corresponding to each time step is predicted, and the state trajectory is generated by combining the prediction results. Based on the state trajectory, a Markov decision process is constructed in combination with the current motor control objectives and constraints. A multi-step decision problem is constructed by defining the state space, action space and reward function. A value function approximator is set based on an adaptive dynamic programming algorithm and the value function estimate is calculated using the state and action corresponding to the current motor as input. The value function is updated through a value iteration algorithm until convergence. A policy function approximator is set to represent the policy function. number, taking the state corresponding to the current motor as input, calculating the action probability distribution corresponding to the output and updating the policy function by maximizing the expected cumulative reward, initializing the value function approximator and the policy function approximator and generating trajectory data according to the current policy function, updating the parameters of the value function approximator based on the trajectory data, updating the parameters of the policy function approximator according to the updated value function combined with the policy gradient algorithm, repeatedly updating to obtain the optimal value function and the optimal policy function after convergence, based on the optimal policy function, performing forward reasoning according to the state trajectory corresponding to the current motor, generating candidate control strategies, and repeatedly generating to obtain the motor control strategy sequence;

[0083] A strategy evaluation model is constructed based on contrastive learning and a reference strategy set is set. For the current strategy in the motor control strategy sequence, a reference strategy is selected and the similarity is calculated. The strategy representation is learned by maximizing the similarity of positive strategy pairs and minimizing the similarity of negative strategy pairs. A hybrid strategy evaluation mechanism is constructed by combining long-term cumulative rewards and short-term instantaneous effects. A comprehensive evaluation result is obtained by weighted fusion of the value function corresponding to the long-term cumulative rewards and the instant reward calculation. Based on the comprehensive evaluation result, an optimal control strategy is selected and the optimal control strategy is parsed into a control instruction. Based on the control instruction, a control instruction graph is generated and motor control is performed.

[0084] The time window refers to a time range selected when analyzing or processing time series data. The sliding step size refers to the distance or step size of each sliding of the window position when the time window moves in time series analysis. The Markov decision process is a mathematical framework for modeling decision processes, in which the future state depends only on the current state and the action currently taken, and is independent of the past state. The multi-step decision problem involves a continuous decision process, in which each decision affects future decisions and the final result. The value function approximator is a model or method for estimating the value function. The value iteration algorithm is a dynamic programming algorithm for solving the optimal strategy in the Markov decision process. The strategy function approximator is used to approximate or estimate the strategy function of the optimal action under a given state. The candidate control strategy refers to a variety of alternative strategies proposed in the control problem. The positive strategy pair refers to a pair of strategy combinations with a positive relationship in contrastive learning or strategy evaluation. The negative strategy pair refers to a pair of strategy combinations with a negative relationship. The negative strategy pair is used to help the model identify and avoid erroneous behavior patterns and plays a "contrast" role in contrastive learning.

[0085] The motor data acquisition unit obtains the current motor operating parameters, such as speed, current, voltage, temperature, etc., and adds these parameters to the cognitive mapping model. The cognitive mapping model learns the semantic features corresponding to the current motor operating state, such as normal operation, minor fault, etc. The low-level features (such as time domain statistical features), intermediate features (such as frequency domain features) and high-level features (such as time-frequency domain features) corresponding to the motor operating parameters are calculated through forward propagation. The multi-level features calculated are combined to generate multi-granularity feature representations to comprehensively characterize the motor's operating state.

[0086] Obtain the historical operation data of the current motor, build a historical data set, set the time window (such as 30 days) and sliding step size (such as 1 day) of the autoregressive model, divide the historical data set into a training set and a test set, use the training set to train the autoregressive model, optimize the model parameters by minimizing the prediction error, and tune the hyperparameters of the autoregressive model (such as the number of hidden layers, learning rate, etc.) in combination with the real labels of the training set until the model converges. Add the data in the test set as input to the trained autoregressive model, recursively predict the state of future time steps, and obtain the historical state evolution trend. Based on the historical state evolution trend, add multi-granularity feature representation to the autoregressive model, predict the features of each future time step, combine the predicted multi-granularity features, and generate a state trajectory to represent the changes in the running state of the motor in the future period of time.

[0087] Based on the state trajectory, combined with the current motor control objectives (such as efficiency maximization) and constraints (such as temperature limit), a Markov decision process is constructed to define the state space (the operating state of the motor), the action space (optional control operations) and the reward function (quantified control effect). The motor control is modeled as a multi-step decision problem. A value function approximator (such as a neural network) is set up. The current state and action are used as inputs to calculate the estimated value of the state-action value function. The parameters of the value function approximator are iteratively updated through an adaptive dynamic programming algorithm until the optimal value function is obtained through convergence. A policy function approximator (such as another neural network) is set up. The current state is used as input and the probability distribution of the action is output. The policy function approximator parameters are updated based on the trajectory data through a policy gradient algorithm so that it can maximize the expected cumulative reward. The value function and the policy function are optimized alternately until the optimal policy function is obtained through convergence. Based on the optimal policy function, reasoning is performed according to the current motor state to generate control candidates and form a motor control strategy sequence.

[0088] Construct a strategy evaluation model, learn strategy representation through contrastive learning, set a reference strategy set, including multiple known good strategies, select positive strategy and negative strategy in the reference set for the current strategy in the control strategy sequence, calculate the similarity between the current strategy and the positive and negative strategy, optimize strategy representation by maximizing the positive similarity and minimizing the negative similarity, design a hybrid strategy evaluation mechanism, comprehensively consider long-term cumulative returns (estimated by value function) and short-term immediate rewards, weightedly integrate long-term value and immediate rewards, and obtain a comprehensive evaluation result of the current strategy. Based on the comprehensive evaluation result, select the optimal control strategy from the candidate strategies, parse the optimal control strategy into specific control instructions (such as adjusting the speed, current, etc.), generate a control instruction graph, and send the control instructions to the motor execution unit to achieve optimal control of the motor;

[0089] For example, taking a wind turbine as an example, assuming that the data acquisition unit collects the motor's operating parameters such as speed, current, voltage, temperature, etc. once per second, the cognitive mapping model learns the generator's operating status, such as normal power generation, low wind speed, cabin vibration, etc., and extracts multi-granularity features (such as mean, peak, spectrum, etc.). Based on the historical data of the past 30 days, the autoregressive model predicts the generator's state trajectory for the next 24 hours. Based on the predicted state, the Markov decision process generates a series of control strategies (such as "increase the cut-in wind speed", "increase the yaw angle", etc.), and the strategy evaluation model selects similar high-quality strategies (such as "historical optimal operation strategy") and bad strategies (such as "frequent start-stop strategy") from the reference set. By comparing and learning the optimization strategy representation, the optimal control strategy (such as "dynamically adjusting the pitch angle") is selected by comprehensively considering the long-term power generation and short-term unit loss. The optimal strategy is converted into a specific control instruction (such as: adjusting the pitch angle to 12 degrees), and an instruction chart is generated and sent to the execution unit of the generator to adjust the operating parameters in real time to keep the wind turbine in the best working state and achieve a balance between power generation efficiency and reliability.

[0090] In this embodiment, the semantic features of the motor's operating state are learned through the cognitive mapping model, and multi-granular feature representations are extracted to characterize the motor's state from different scales and angles, providing comprehensive and accurate state perception capabilities, providing reliable input for subsequent predictions and decisions, and generating state trajectories through recursive predictions to capture the long-term evolution trend of the motor's state. At the same time, combined with multi-granular feature representations, the state changes at different time scales in the future are predicted, providing long-term and short-term combined prediction capabilities, providing forward-looking guidance for control decisions, and realizing adaptive optimization of value functions and policy functions through value iteration and policy gradient algorithms. The optimal control strategy is autonomously generated according to the current state and long-term returns, which has adaptive The method has good robustness and can cope with the dynamic changes of motor working conditions. By comparing the similarity with the reference strategy, the advantages and disadvantages of the learning strategy are expressed, and a quantitative evaluation of the strategy quality is provided. A hybrid evaluation mechanism is designed to take into account both long-term cumulative returns and short-term immediate benefits, and to select the control strategy with the best comprehensive effect to ensure the effectiveness and reliability of the control. In summary, this embodiment provides a data-driven adaptive motor control optimization method, which realizes intelligent state perception, prediction, and decision-making through multi-stage modeling and learning, generates a dynamically optimized control strategy, can adaptively cope with changes in motor working conditions, continuously optimize operating performance, improve the effectiveness and reliability of control, and reduce the need for manual intervention.

[0091] In an optional embodiment,

[0092] The parameters of the policy function approximator are updated according to the updated value function combined with the policy gradient algorithm as shown in the following formula:

[0093] ;

[0094] in, Indicates the parameters of the current strategy, Express expectations, Indicates that the current policy is in state Next select action The probability of Indicates that the old policy is in state Next select action The probability of represents the clipping function, A hyperparameter representing the clipping range, Represents the advantage function, which is used to measure the state Next Action Advantages, Represents the time step.

[0095] In this embodiment, the parameters of the policy function are updated by gradient ascent so that the expected cumulative reward can be maximized. When calculating the policy gradient, the importance sampling weight is introduced, and the gradient estimate is adjusted by the probability ratio of the current policy and the old policy taking a specific action under a given state. This can correct the distribution shift caused by the policy update and improve the accuracy and stability of the gradient estimate. The importance sampling weight is clipped using a clipping function to prevent the gradient estimate from being too large or too small, avoid the instability of the policy update, and improve the robustness of the training. The advantage function reflects the relative advantages and disadvantages of the action and can more effectively guide the policy to update in a better direction. In summary, this embodiment realizes the adaptive optimization of the policy function, can learn and improve the control strategy online, adapt to the dynamic changes of the motor working conditions, and continuously optimize the operating performance of the motor. Compared with traditional strategies, it has higher sample efficiency, more stable training process and faster convergence speed.

[0096] S3. During the control process, real-time operation data is acquired and deviation data is calculated, and data analysis is performed in combination with a distributed active inference algorithm to obtain deviation analysis results. Based on the knowledge graph algorithm, the real-time operation data is mapped into performance indicators and a multi-dimensional evaluation system is constructed. According to the deviation analysis results, anomaly points and abnormal intervals are determined in combination with an anomaly detection method based on confidence learning, and an anomaly analysis report is generated in combination with a pre-set expert knowledge base. Based on the anomaly analysis report, the cognitive mapping model and the strategy evaluation model are updated in combination with a meta-learning algorithm. The update is repeated until the preset maximum number of iterations is reached and the optimal control strategy generated by the last iteration is output.

[0097] The deviation data refers to the phenomenon that the statistical characteristics of the data in the data set deviate from the true value or expected value due to some reason. The distributed active inference algorithm is an algorithm that performs reasoning and decision-making in a distributed manner among multiple nodes or agents. The deviation analysis result refers to the conclusion obtained by analyzing the deviation in the data or model. The confidence space refers to a multidimensional space containing confidence intervals in probability or statistical inference. The abnormal interval refers to the time period or data interval where abnormal behavior occurs in a time series or data stream.

[0098] In an optional embodiment,

[0099] During the control process, real-time operation data is acquired and deviation data is calculated, and data analysis is performed in combination with a distributed active inference algorithm to obtain deviation analysis results. The real-time operation data is mapped to performance indicators based on a knowledge graph algorithm and a multi-dimensional evaluation system is constructed. According to the deviation analysis results, anomaly points and abnormal intervals are determined in combination with an anomaly detection method based on confidence learning, and an anomaly analysis report is generated in combination with a pre-set expert knowledge base. Based on the anomaly analysis report, the cognitive mapping model and the strategy evaluation model are updated in combination with a meta-learning algorithm, and the update is repeated until the preset maximum number of iterations is reached and the optimal control strategy generated by the last iteration is output, including:

[0100] Acquire real-time operation data corresponding to the current motor during the execution of motor control by the motor data acquisition unit, wherein the real-time operation data includes current, voltage, speed and temperature, compare the real-time operation data with pre-set standard operation data, calculate deviation data and construct an inference rule library through a distributed active inference algorithm, wherein the inference rule library includes abnormal pattern recognition rules and data correlation analysis rules, select an inference path based on the inference rule library and perform data analysis to obtain a deviation analysis result corresponding to the deviation data;

[0101] Based on the knowledge graph algorithm, the real-time operation data is mapped into multiple dimensions of motor performance and a multi-dimensional evaluation index system is constructed, a comprehensive evaluation function corresponding to the motor performance is constructed based on the multi-dimensional evaluation index system, a judgment matrix is ​​constructed by the hierarchical analysis method, the eigenvector corresponding to the maximum eigenvalue of the judgment matrix is ​​solved, and the weight coefficient corresponding to each dimension is obtained, and the real-time operation data is substituted into the comprehensive evaluation function to obtain the comprehensive evaluation value corresponding to the current motor;

[0102] Through the anomaly detection method based on confidence learning, an anomaly point determination function is constructed and the anomaly threshold is calculated in combination with the confidence interval estimation algorithm. For the real-time operation data of each time step, the anomaly value is calculated through the anomaly point determination function and screened through the anomaly threshold. If the anomaly value is greater than the anomaly threshold, the real-time operation data of the current time step is considered to be an anomaly point. The screening is repeated until all time steps are screened. For the anomaly point, the distance between each anomaly point and the point corresponding to other real-time operation data is calculated, a distance matrix is ​​generated, and the anomaly points are clustered to obtain the anomaly interval. The cause of the anomaly is analyzed and the impact is evaluated in combination with a preset expert knowledge base, and a corresponding anomaly analysis report is generated;

[0103] The abnormal analysis report, the deviation analysis result and the comprehensive evaluation value are used as inputs of a meta-learning algorithm. An optimization-based meta-learning algorithm is used, with the cognitive mapping model and the strategy evaluation model as base learners and adaptively updated. A maximum number of iterations is set and the cognitive mapping model and the strategy evaluation model are repeatedly updated until the maximum number of iterations is reached, and the optimal control strategy generated in the last iteration is used as the control instruction of the current motor.

[0104] The inference rule base is a set of rules for deriving conclusions, the abnormal pattern recognition rules are rules for detecting abnormal patterns or behaviors in data, the data association analysis rules are used to evaluate the relationship or degree of association between data variables, the hierarchical analysis method is a multi-criteria decision analysis method, which is used to make decisions by decomposing complex problems into a hierarchical structure, the judgment matrix is ​​a matrix used in the hierarchical analysis method to compare the relative importance of different decision factors, the outlier determination function is a function used to identify outliers in a data set, usually based on statistical methods or machine learning algorithms, and determines outliers by calculating the degree of deviation of data points from normal data distribution, and the base learner is the basic component unit in the integrated learning method, which refers to a single learning algorithm or model.

[0105] In the process of executing motor control, the real-time operation data of the current motor, including current, voltage, speed and temperature, is obtained through the motor data acquisition unit. The real-time operation data is compared with the pre-set standard operation data, and the deviation data is calculated to reflect the abnormal degree of the motor operation state. The inference rule library is constructed through the distributed active inference algorithm, which includes abnormal pattern recognition rules (such as current mutation, high temperature, etc.) and data correlation analysis rules (such as the relationship between current and speed). Based on the inference rule library, the appropriate inference path is selected (such as first judging the current abnormality and then analyzing the temperature), and the deviation data is analyzed to obtain the deviation analysis results, explaining the abnormal causes and effects;

[0106] Use the knowledge graph algorithm to map the real-time operation data into multiple dimensions of motor performance, such as efficiency, reliability, stability, etc., and build a multi-dimensional evaluation index system. Based on the evaluation index system, build a comprehensive evaluation function of motor performance, comprehensively consider the index values ​​of each dimension, and use the hierarchical analysis method to build a judgment matrix. Compare the importance of each dimension in pairs to form a judgment matrix. Solve the eigenvector corresponding to the maximum eigenvalue of the judgment matrix to obtain the weight coefficient of each dimension, reflecting its contribution to the comprehensive performance. Substitute the real-time operation data into the comprehensive evaluation function, combine the dimension weights, calculate the comprehensive evaluation value of the current motor, and quantify the overall performance level of the motor.

[0107] Construct an outlier determination function to determine whether the real-time operation data of each time step is abnormal. Combined with the confidence interval estimation algorithm, calculate the anomaly threshold as the basis for determining the anomaly. For the real-time operation data of each time step, calculate the outlier value through the outlier determination function and compare it with the anomaly threshold. If the outlier value is greater than the anomaly threshold, the data of this time step is marked as an outlier, otherwise it is regarded as a normal point. Repeat the screening process until the data of all time steps are detected and a complete set of outliers is obtained. For the set of outliers, calculate the distance between each outlier and other data points to generate a distance matrix. Based on the distance matrix, use a clustering algorithm (such as DBSCAN) to cluster the outliers to obtain an abnormal interval, which represents the duration of the abnormal behavior. Combined with the pre-set expert knowledge base, perform cause analysis and impact assessment on the abnormal interval, and generate an abnormal analysis report to explain the cause of the abnormality and the possible consequences.

[0108] The abnormal analysis report, deviation analysis results and comprehensive evaluation values ​​are used as the input of the meta-learning algorithm to reflect the multi-dimensional information of the motor's operating status. The optimization-based meta-learning algorithm is used, with the cognitive mapping model and the strategy evaluation model as the base learner. Its generalization ability is improved through adaptive updates, and the maximum number of iterations is set, and the meta-learning optimization process is repeated. In each iteration, the base learner is adaptively updated according to the new input data to generate a better control strategy, and the iteration is repeated until the maximum number of iterations is reached or the strategy converges. The optimal control strategy generated in the last iteration is used as the control instruction of the current motor to achieve adaptive optimization control.

[0109] For example, taking the servo motor of an industrial robot as an example, the current, voltage, speed and temperature data of the motor are collected in real time through the data acquisition unit. The real-time data is compared with the standard data, and it is found that the current deviation is large. Through the analysis of the inference rule base, it is judged that the motor may be overloaded. The real-time data is mapped into three dimensions of efficiency, reliability and stability using the knowledge graph algorithm. A comprehensive evaluation function is constructed, and the weight of each dimension is determined by the hierarchical analysis method. The calculated comprehensive evaluation value is 0.75, indicating that the motor performance needs to be improved. Based on the anomaly detection method of confidence learning, the anomaly threshold is set to 0.1, and the real-time data is gradually checked. It is found that within a certain period of time, the abnormal value exceeds the threshold continuously and is marked as an abnormal interval. Combined with the expert knowledge base, it is analyzed that the abnormality may be caused by the long-term high-load operation of the motor, resulting in temperature increase. The control strategy needs to be adjusted in time. The abnormal analysis report, deviation analysis result (overload) and comprehensive evaluation value (0.75) are input into the meta-learning algorithm. The cognitive mapping model and strategy evaluation model are used as the base learner, and the maximum number of iterations is set to 10. Through adaptive updating, the control strategy is continuously optimized. Finally, an optimal control strategy for reducing load and adjusting speed is generated, which is sent to the motor execution unit to realize the adaptive control optimization of the motor.

[0110] In this embodiment, the real-time operation data of the motor is obtained through the motor data acquisition unit, including key parameters such as current, voltage, speed and temperature, so as to realize real-time monitoring of the motor operation status, provide a data basis for subsequent analysis and optimization, build an inference rule library through the distributed active reasoning algorithm, select the appropriate reasoning path to analyze the deviation data, generate deviation analysis results, and quickly find abnormal conditions in the operation of the motor, and give possible explanations of the causes, to assist fault diagnosis and maintenance decisions, determine the weights of each dimension through the hierarchical analysis method, and calculate the comprehensive evaluation value of the motor, so as to comprehensively evaluate the performance of the motor, quantify the working status of the motor in different aspects, and provide a reference for control optimization. Based on the above, the deviation analysis results and the comprehensive evaluation values ​​are used as the input of the meta-learning algorithm. The optimization-based meta-learning algorithm is used, with the cognitive mapping model and the strategy evaluation model as the base learner. The optimal control strategy is generated through adaptive updating. The control strategy can be dynamically adjusted according to the actual operating state of the motor to adapt to the performance requirements under different working conditions and realize adaptive optimization of motor control. In summary, this embodiment realizes the intelligence and automation of motor control by real-time monitoring of the motor state, analyzing and evaluating performance, detecting and diagnosing abnormal behavior, and adaptively optimizing the control strategy, thereby improving the efficiency, reliability and safety of motor operation, and providing a novel and effective solution for motor control in industrial production.

[0111] In an optional embodiment,

[0112] The comprehensive evaluation function corresponding to the motor performance is constructed based on the multi-dimensional evaluation index system as shown in the following formula:

[0113] ;

[0114] in, represents the evaluation value, represents the product operation, Represents the total number of dimensions, Indicates m The nonlinear evaluation function in dimensions, Indicates m The weight coefficient of each dimension, Indicates m Dimensions and n The weight coefficient of the interaction between dimensions, represents the adjustment parameter, Indicates n The nonlinear evaluation function in dimensions, Indicates m Specific indicators related to each dimension, Indicates n Specific indicators related to each dimension, Indicates m Dimensions and n The interaction function between the dimensions.

[0115] In this embodiment, the multi-dimensional modeling method can comprehensively characterize the performance characteristics of the motor and provide a more accurate and fine-grained performance evaluation. The dimension weights are determined by methods such as hierarchical analysis method, and the relative importance of each dimension is quantified to make the evaluation results more reasonable and explainable. The nonlinear evaluation function can better capture the changing trends and sensitive intervals of performance indicators, and provide a more accurate and detailed performance evaluation. The introduction of dimensional interaction terms to model the interaction effects between different dimensions takes into account the mutual influence and coupling relationship between dimensions, and can capture the comprehensive impact of dimensional combinations on performance, providing a more comprehensive and systematic performance evaluation. In summary, this embodiment realizes the quantitative evaluation of motor performance through methods such as nonlinear evaluation, dimensional weight quantification, and interaction effect modeling. It can comprehensively consider multiple aspects of motor performance, provide more accurate, fine-grained and systematic performance evaluation results, and provide a quantitative basis for motor performance optimization and control decisions.

[0116] Figure 2 FIG. 1 is a schematic diagram of a control system for a motor controller according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0117] The first unit is used to collect multi-dimensional parameters corresponding to the current motor through a motor data collection unit, collect environmental parameters in combination with a preset environmental sensor array, combine the environmental parameters with the multi-dimensional parameters through an adaptive data fusion method based on active learning to obtain an operation data stream, perform adaptive spatiotemporal feature extraction on the operation data stream, obtain a spatiotemporal feature representation, and model the motor state data distribution based on a generative adversarial network, learn the state features corresponding to the spatiotemporal feature representation through a generator, and determine the probability distribution of the state features through a variational autoencoder and a maximum marginal likelihood estimation algorithm, construct a state transfer equation, and obtain a cognitive mapping model;

[0118] The second unit is used to obtain the motor operation parameters through the motor data acquisition unit and add them to the cognitive mapping model, calculate the multi-granularity feature representation corresponding to the current operation state through forward reasoning, and predict the state trajectory through the autoregressive model in combination with the historical state evolution trend obtained in advance, based on the state trajectory, combine the adaptive dynamic programming algorithm to perform multi-step decision solving and generate a motor control strategy sequence, build a reference strategy set, build a strategy evaluation model based on contrastive learning and determine the similarity between the current strategy and the reference strategy, build a hybrid strategy evaluation mechanism in combination with long-term cumulative rewards and short-term instantaneous effects, and evaluate the control strategy sequence, determine the optimal control strategy and parse it into a control instruction diagram for motor control;

[0119] The third unit is used to obtain real-time operation data and calculate deviation data during the control process and perform data analysis in combination with a distributed active inference algorithm to obtain deviation analysis results, map the real-time operation data into performance indicators based on a knowledge graph algorithm and build a multi-dimensional evaluation system, determine abnormal points and abnormal intervals based on the deviation analysis results in combination with an anomaly detection method based on confidence learning, generate an anomaly analysis report in combination with a pre-set expert knowledge base, update the cognitive mapping model and the strategy evaluation model based on the anomaly analysis report in combination with a meta-learning algorithm, repeat the update until the preset maximum number of iterations is reached and output the optimal control strategy generated by the last iteration.

[0120] According to a third aspect of the embodiments of the present invention,

[0121] An electronic device is provided, comprising:

[0122] processor;

[0123] a memory for storing processor-executable instructions;

[0124] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0125] A fourth aspect of the embodiments of the present invention is:

[0126] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0127] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for a motor controller, characterized in that: include: The multi-dimensional parameters corresponding to the current motor are collected by a motor data collection unit, and the environmental parameters are collected in combination with a pre-set environmental sensor array. The environmental parameters are combined with the multi-dimensional parameters by an adaptive data fusion method based on active learning to obtain an operation data stream, and the operation data stream is adaptively extracted with spatiotemporal features to obtain a spatiotemporal feature representation and model the motor state data distribution based on a generative adversarial network. The state features corresponding to the spatiotemporal feature representation are learned by a generator, and the probability distribution of the state features is determined by a variational autoencoder and a maximum marginal likelihood estimation algorithm, and a state transfer equation is constructed to obtain a cognitive mapping model. The motor operation parameters are obtained through the motor data acquisition unit and added to the cognitive mapping model, the multi-granularity feature representation corresponding to the current operation state is calculated through forward reasoning, the state trajectory is predicted by the autoregressive model in combination with the historical state evolution trend obtained in advance, based on the state trajectory, a multi-step decision solution is performed in combination with an adaptive dynamic programming algorithm and a motor control strategy sequence is generated, a reference strategy set is constructed, a strategy evaluation model is constructed based on contrastive learning and the similarity between the current strategy and the reference strategy is determined, a hybrid strategy evaluation mechanism is constructed in combination with long-term cumulative rewards and short-term instantaneous effects, and the control strategy sequence is evaluated, the optimal control strategy is determined and parsed into a control instruction diagram for motor control; During the control process, real-time operation data is acquired and deviation data is calculated, and data analysis is performed in combination with a distributed active inference algorithm to obtain deviation analysis results. Based on a knowledge graph algorithm, the real-time operation data is mapped into performance indicators and a multi-dimensional evaluation system is constructed. According to the deviation analysis results, anomaly points and abnormal intervals are determined in combination with an anomaly detection method based on confidence learning, and an anomaly analysis report is generated in combination with a pre-set expert knowledge base. Based on the anomaly analysis report, the cognitive mapping model and the strategy evaluation model are updated in combination with a meta-learning algorithm, and the update is repeated until a preset maximum number of iterations is reached and the optimal control strategy generated by the last iteration is output; The multi-dimensional parameters corresponding to the current motor are collected by the motor data collection unit, and the environmental parameters are collected in combination with the pre-set environmental sensor array. The environmental parameters are combined with the multi-dimensional parameters by an adaptive data fusion method based on active learning to obtain an operation data stream, and the operation data stream is adaptively extracted with spatiotemporal features to obtain a spatiotemporal feature representation and model the motor state data distribution based on a generative adversarial network. The state features corresponding to the spatiotemporal feature representation are learned by a generator, and the probability distribution of the state features is determined by a variational autoencoder and a maximum marginal likelihood estimation algorithm. The state transfer equation is constructed and the cognitive mapping model is obtained, including: The multi-dimensional parameters corresponding to the current motor are collected by a pre-set motor data collection unit, wherein the multi-dimensional parameters include speed, current, voltage, vibration frequency, power factor and load torque; the ambient temperature, humidity and air pressure are collected as environmental parameters by a pre-set environmental sensor array; the credibility and relevance of the motor data collection unit and the environmental sensor array are evaluated by an adaptive data fusion method based on active learning; a data source evaluation result is generated based on the credibility and relevance; the weight coefficients corresponding to the motor data collection unit and the environmental sensor array are dynamically adjusted based on the data source evaluation result; the multi-dimensional parameters and the environmental parameters are adaptively fused by weighted fusion to obtain the operation data stream; For the running data stream, in the time dimension, the time series is modeled by a gated recurrent unit, the dependency of the time series is determined by combining the update gate and the reset gate, and the attention coefficient corresponding to different time segments is adjusted in combination with the attention mechanism to extract the dynamic change pattern of the time series feature. In the spatial dimension, feature extraction is performed by a graph convolutional neural network, the data corresponding to the current motor is used as a node in the graph structure and the relationship between different data is used as an edge to construct a motor parameter association graph, the neighborhood information of each node is aggregated through a graph convolution operation to obtain the structural feature corresponding to each node, and the spatial association pattern of the data corresponding to the current motor is obtained at the same time, the spatial association pattern is dynamically updated through a time-varying graph convolutional network to obtain the spatial feature, the time series feature and the spatial feature are mapped to a common semantic space and semantically aligned, and then feature fusion is performed to obtain a spatiotemporal feature representation; The spatiotemporal feature representation is added as input to the generator network of the generative adversarial network. The generator network stacks multiple transposed convolutional layers and upsampling layers through a deep convolutional neural network structure to map the low-dimensional random noise vector to a high-dimensional state feature space. The motor operating conditions are input as additional conditions in combination with the conditional generation mechanism to generate state features corresponding to the spatiotemporal feature representation. The discriminator in the generative adversarial network maps the state features to binary classification probabilities by stacking multiple convolutional layers and downsampling layers. In combination with a multi-scale discrimination strategy, multiple discriminators with different receptive fields are set to evaluate the authenticity of the state features, and adversarial training is performed and a variational autoencoder is set. The state feature is mapped to the latent variable space through the variational autoencoder, the posterior distribution of the latent variables in the latent variable space is learned, the state feature is reconstructed from the latent variable space through the decoder network and the prior distribution under the state feature is learned, based on the maximum marginal likelihood estimation algorithm, the posterior distribution is used as the conditional probability distribution of the latent variable, the prior distribution is used as the marginal probability distribution of the latent variable, the latent variable is represented as a combination of deterministic variables and noise in combination with the reparameterization technique to complete low variance sampling, and the geometric distance of the latent variable in the manifold space is minimized in combination with the manifold regularization method to obtain the probability distribution corresponding to the state feature; Determine the relationship equation between state variables through a gray box modeling method based on physical laws, determine the state transfer equation through a deep neural network in combination with a data-driven black box modeling method, and construct the cognitive mapping model based on the state transfer equation; The motor operation parameters are obtained by the motor data acquisition unit and added to the cognitive mapping model, the multi-granularity feature representation corresponding to the current operation state is calculated by forward reasoning, the state trajectory is predicted by the autoregressive model in combination with the historical state evolution trend obtained in advance, based on the state trajectory, a multi-step decision solution is performed in combination with an adaptive dynamic programming algorithm and a motor control strategy sequence is generated, a reference strategy set is constructed, a strategy evaluation model is constructed according to contrastive learning and the similarity between the current strategy and the reference strategy is determined, a hybrid strategy evaluation mechanism is constructed in combination with long-term cumulative rewards and short-term instantaneous effects and the control strategy sequence is evaluated, and the optimal control strategy is determined and parsed into a control instruction diagram for motor control, including: The motor operating parameters corresponding to the current motor are obtained through the motor data acquisition unit and added to the cognitive mapping model, the cognitive mapping model learns the semantic features corresponding to the operating state of the current motor and obtains the low-level features, intermediate features and high-level features corresponding to the motor operating parameters through forward propagation calculation, and the calculated features are combined to generate the multi-granularity feature representation; Obtain historical operating data of the current motor and construct a historical data set, set a time window and a sliding step size of the autoregressive model, divide the historical data set into a training set and a test set, use the training set to train the autoregressive model, optimize the hyperparameters of the autoregressive model in combination with the real labels of the training set until the model converges, add the elements in the test set as input to the autoregressive model, and obtain the historical state evolution trend; Based on the historical state evolution trend, the multi-granularity feature representation is added to the autoregressive model and the multi-granularity feature representation corresponding to each time step is predicted, and the state trajectory is generated by combining the prediction results. Based on the state trajectory, a Markov decision process is constructed in combination with the current motor control objectives and constraints. A multi-step decision problem is constructed by defining the state space, action space and reward function. A value function approximator is set based on an adaptive dynamic programming algorithm and the value function estimate is calculated using the state and action corresponding to the current motor as input. The value function is updated through a value iteration algorithm until convergence. A policy function approximator is set to represent the policy function. number, taking the state corresponding to the current motor as input, calculating the action probability distribution corresponding to the output and updating the policy function by maximizing the expected cumulative reward, initializing the value function approximator and the policy function approximator and generating trajectory data according to the current policy function, updating the parameters of the value function approximator based on the trajectory data, updating the parameters of the policy function approximator according to the updated value function combined with the policy gradient algorithm, repeatedly updating to obtain the optimal value function and the optimal policy function after convergence, based on the optimal policy function, performing forward reasoning according to the state trajectory corresponding to the current motor, generating candidate control strategies, and repeatedly generating to obtain the motor control strategy sequence; Based on contrastive learning, a strategy evaluation model is constructed and a reference strategy set is set. For the current strategy in the motor control strategy sequence, a reference strategy is selected and similarity is calculated. The strategy representation is learned by maximizing the similarity of positive strategy pairs and minimizing the similarity of negative strategy pairs. A hybrid strategy evaluation mechanism is constructed by combining long-term cumulative rewards and short-term instantaneous effects. A comprehensive evaluation result is obtained by weighted fusion of the value function corresponding to the long-term cumulative rewards and the instant reward calculation. Based on the comprehensive evaluation result, an optimal control strategy is selected and the optimal control strategy is parsed into a control instruction. A control instruction diagram is generated based on the control instruction and motor control is performed. During the control process, real-time operation data is acquired and deviation data is calculated, and data analysis is performed in combination with a distributed active inference algorithm to obtain deviation analysis results. The real-time operation data is mapped to performance indicators based on a knowledge graph algorithm and a multi-dimensional evaluation system is constructed. According to the deviation analysis results, anomaly points and abnormal intervals are determined in combination with an anomaly detection method based on confidence learning, and an anomaly analysis report is generated in combination with a pre-set expert knowledge base. Based on the anomaly analysis report, the cognitive mapping model and the strategy evaluation model are updated in combination with a meta-learning algorithm, and the update is repeated until the preset maximum number of iterations is reached and the optimal control strategy generated by the last iteration is output, including: Acquire real-time operation data corresponding to the current motor during the execution of motor control by the motor data acquisition unit, wherein the real-time operation data includes current, voltage, speed and temperature, compare the real-time operation data with pre-set standard operation data, calculate deviation data and construct an inference rule library through a distributed active inference algorithm, wherein the inference rule library includes abnormal pattern recognition rules and data correlation analysis rules, select an inference path based on the inference rule library and perform data analysis to obtain a deviation analysis result corresponding to the deviation data; Based on the knowledge graph algorithm, the real-time operation data is mapped into multiple dimensions of motor performance and a multi-dimensional evaluation index system is constructed, a comprehensive evaluation function corresponding to the motor performance is constructed based on the multi-dimensional evaluation index system, a judgment matrix is ​​constructed by the hierarchical analysis method, the eigenvector corresponding to the maximum eigenvalue of the judgment matrix is ​​solved, and the weight coefficient corresponding to each dimension is obtained, and the real-time operation data is substituted into the comprehensive evaluation function to obtain the comprehensive evaluation value corresponding to the current motor; Through the anomaly detection method based on confidence learning, an anomaly point determination function is constructed and the anomaly threshold is calculated in combination with the confidence interval estimation algorithm. For the real-time operation data of each time step, the anomaly value is calculated through the anomaly point determination function and screened through the anomaly threshold. If the anomaly value is greater than the anomaly threshold, the real-time operation data of the current time step is considered to be an anomaly point. The screening is repeated until all time steps are screened. For the anomaly point, the distance between each anomaly point and the point corresponding to other real-time operation data is calculated, a distance matrix is ​​generated, and the anomaly points are clustered to obtain the anomaly interval. The cause of the anomaly is analyzed and the impact is evaluated in combination with a preset expert knowledge base, and a corresponding anomaly analysis report is generated; The abnormal analysis report, the deviation analysis result and the comprehensive evaluation value are used as inputs of a meta-learning algorithm. An optimization-based meta-learning algorithm is used, with the cognitive mapping model and the strategy evaluation model as base learners and adaptively updated. A maximum number of iterations is set and the cognitive mapping model and the strategy evaluation model are repeatedly updated until the maximum number of iterations is reached, and the optimal control strategy generated in the last iteration is used as the control instruction of the current motor.

2. The method according to claim 1, characterized in that The objective function of the variational autoencoder is shown in the following formula: ； in, represents the weight of the generator network, represents the weight of the variational encoder, Indicates state characteristics, represents the i-th latent variable sample, For The expectation of importance sampling samples, which means calculating the The average value of Express expectations, represents the number of samples for importance sampling, Indicates that in a given state feature In the case of Samples The joint distribution of Indicates state characteristics and latent variable samples The joint probability distribution of is a hidden variable The posterior distribution of The hidden variable The probability distribution of .

3. The method according to claim 1, characterized in that: The parameters of the policy function approximator are updated according to the updated value function combined with the policy gradient algorithm as shown in the following formula: ； in, Indicates the parameters of the current strategy, Express expectations, Indicates that the current policy is in state Next select action The probability of Indicates that the old policy is in state Next select action The probability of represents the clipping function, A hyperparameter representing the clipping range, Represents the advantage function, which is used to measure the state Next Action Advantages, Represents the time step.

4. The method according to claim 1, characterized in that The comprehensive evaluation function corresponding to the motor performance is constructed based on the multi-dimensional evaluation index system as shown in the following formula: ； in, represents the evaluation value, represents the product operation, Represents the total number of dimensions, represents the nonlinear evaluation function of the mth dimension, represents the weight coefficient of the mth dimension, Represents the weight coefficient of the interaction between the mth dimension and the nth dimension, represents the adjustment parameter, represents the nonlinear evaluation function of the nth dimension, Represents the specific indicators related to the mth dimension, Represents the specific indicators related to the nth dimension, Represents the interaction function between the m-th dimension and the n-th dimension.

5. A control system for a motor controller, used to implement the method according to any one of claims 1 to 4, characterized in that: include: The first unit is used to collect multi-dimensional parameters corresponding to the current motor through a motor data collection unit, collect environmental parameters in combination with a preset environmental sensor array, combine the environmental parameters with the multi-dimensional parameters through an adaptive data fusion method based on active learning to obtain an operation data stream, perform adaptive spatiotemporal feature extraction on the operation data stream, obtain a spatiotemporal feature representation, and model the motor state data distribution based on a generative adversarial network, learn the state features corresponding to the spatiotemporal feature representation through a generator, and determine the probability distribution of the state features through a variational autoencoder and a maximum marginal likelihood estimation algorithm, construct a state transfer equation, and obtain a cognitive mapping model; The second unit is used to obtain the motor operation parameters through the motor data acquisition unit and add them to the cognitive mapping model, calculate the multi-granularity feature representation corresponding to the current operation state through forward reasoning, and predict the state trajectory through the autoregressive model in combination with the historical state evolution trend obtained in advance, based on the state trajectory, combine the adaptive dynamic programming algorithm to perform multi-step decision solving and generate a motor control strategy sequence, build a reference strategy set, build a strategy evaluation model based on contrastive learning and determine the similarity between the current strategy and the reference strategy, build a hybrid strategy evaluation mechanism in combination with long-term cumulative rewards and short-term instantaneous effects, and evaluate the control strategy sequence, determine the optimal control strategy and parse it into a control instruction diagram for motor control; The third unit is used to obtain real-time operation data and calculate deviation data during the control process and perform data analysis in combination with a distributed active inference algorithm to obtain deviation analysis results, map the real-time operation data into performance indicators based on a knowledge graph algorithm and build a multi-dimensional evaluation system, determine abnormal points and abnormal intervals based on the deviation analysis results in combination with an anomaly detection method based on confidence learning, generate an anomaly analysis report in combination with a pre-set expert knowledge base, update the cognitive mapping model and the strategy evaluation model based on the anomaly analysis report in combination with a meta-learning algorithm, repeat the update until the preset maximum number of iterations is reached and output the optimal control strategy generated by the last iteration.

6. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images