Motor stator vibration suppression method and system based on electromagnetic pad and deep learning

Abstract

The invention provides a motor stator vibration suppression method and system based on an electromagnetic pad and deep learning, and relates to the field of electrical technology, and the method comprises the steps: collecting a vibration signal of a motor stator, and carrying out the preprocessing of the vibration signal; inputting the preprocessed signal into a deep learning model, and training and optimizing the deep learning model; fusing the features through a full connection layer to obtain a comprehensive feature vector; predicting the future vibration state of the motor stator by using the comprehensive feature vector; establishing an electromagnetic force model, and designing an adaptive control law based on the predicted future vibration state and the electromagnetic force model; electromagnetic force required by vibration suppression is calculated by using a self-adaptive control law and is converted into a control signal to adjust the electromagnetic pad; a multi-objective optimization algorithm is used to consider the vibration suppression effect and energy consumption at the same time, an adaptive control law is adjusted, and closed-loop feedback control is formed to continuously optimize the vibration suppression effect.

Description

Technical Field

[0001] The present invention relates to electrical technology, and in particular to a method and system for suppressing motor stator vibration based on electromagnetic pads and deep learning. Background Art

[0002] Existing motor stator vibration suppression technologies have many shortcomings. Although structural optimization methods can reduce vibration sources to a certain extent, they often require substantial modifications to the motor, increasing manufacturing costs and complexity, and are difficult to adapt to motors that have already been put into use. Passive damping technology is simple to implement, but its effect is limited. It is difficult to cope with changes in vibration characteristics under different working conditions and cannot meet the needs of high-precision vibration suppression. Although traditional active control methods have certain adaptability, they usually rely on accurate system models, which are often difficult to obtain in practical applications, especially for complex motor systems. In addition, the control algorithms of these methods are generally complex, computationally intensive, and difficult to achieve real-time control.

[0003] Although the existing electromagnetic pad-based vibration suppression methods have made breakthroughs in execution performance, they still have deficiencies in the design of control strategies. Most methods still use traditional control theories, such as PID control or adaptive control, which often fail to achieve ideal control effects when faced with the complex nonlinear dynamic characteristics of the motor stator. At the same time, these methods also have difficulty in effectively handling the rapid changes in the motor's working state and external interference.

[0004] In addition, existing technologies also have limitations in vibration characteristic modeling. Most methods rely on simplified mathematical models or empirical formulas, which make it difficult to accurately describe the complex vibration behavior of the motor stator under different operating conditions. This inaccurate modeling directly affects the performance of the control system, resulting in unsatisfactory vibration suppression effects. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for suppressing motor stator vibration based on electromagnetic pads and deep learning, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention, Provides a motor stator vibration suppression method based on electromagnetic pad and deep learning, including: Collecting vibration signals of the motor stator, wherein the vibration signals include vibration displacement, vibration velocity and vibration acceleration; preprocessing the vibration signals, including filtering, noise reduction and normalization; Inputting the preprocessed signal into a deep learning model, the deep learning model comprising a one-dimensional convolutional neural network, a long short-term memory network and a temporal convolutional network; training and optimizing the deep learning model; fusing features through a fully connected layer to obtain a comprehensive feature vector; and predicting the future vibration state of the motor stator using the comprehensive feature vector; An electromagnetic force model is established, and an adaptive control law is designed based on the predicted future vibration state and the electromagnetic force model. The electromagnetic force required to suppress vibration is calculated using the adaptive control law, and converted into a control signal to adjust the electromagnetic pad. The vibration suppression effect is monitored in real time to obtain actual vibration data. The actual vibration data is compared with the predicted future vibration state to obtain error information. Based on the error information, an online learning algorithm is used to perform real-time fine-tuning on the deep learning model. A multi-objective optimization algorithm is used to simultaneously consider the vibration suppression effect and energy consumption, adjust the adaptive control law, and form a closed-loop feedback control to continuously optimize the vibration suppression effect.

[0007] In an optional embodiment, The preprocessed signal is input into a deep learning model, wherein the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a temporal convolutional network; the deep learning model is trained and optimized; features are fused through a fully connected layer to obtain a comprehensive feature vector; and the steps of using the comprehensive feature vector to predict the future vibration state of the motor stator include: Construct a multi-branch parallel network structure, including a one-dimensional convolutional neural network branch, a long short-term memory network branch, and a temporal convolutional network branch; in the one-dimensional convolutional neural network branch, multiple layers of one-dimensional convolutional layers are set, each convolutional layer is sequentially connected to a batch normalization layer and an activation function layer, and a maximum pooling layer is set between adjacent convolutional layers to extract local temporal features; in the long short-term memory network branch, multiple layers of bidirectional long short-term memory network units are stacked, and the output of each unit layer is subjected to dropout processing to extract long-term temporal dependency features; in the temporal convolutional network branch, a causal convolution structure is adopted, and multiple dilated convolutional layers are set, each layer is connected to a weight normalization layer and an activation function to extract long sequence data features; Introduce a multi-head self-attention mechanism to perform weighted fusion of features; A multi-task learning framework is used to simultaneously input the weighted fused features into multiple prediction tasks to predict the vibration displacement, vibration velocity, and vibration acceleration of future time steps. Multiple fully connected layers are set in each prediction task to gradually reduce the feature dimension, and the output dimension of the last layer matches the prediction time step. Monte Carlo dropout technology is used to keep the dropout layer activated during the inference phase, and multiple forward propagations are performed on the deep learning model to obtain the confidence interval of the prediction results. Determine the future vibration state of the motor stator based on the predicted vibration displacement, vibration velocity and vibration acceleration along with confidence intervals.

[0008] In an optional embodiment, The steps to train and optimize a deep learning model include: Adopting an adaptive multi-task learning strategy, a learnable task weight parameter is introduced, which corresponds to the vibration displacement, vibration velocity and vibration acceleration prediction tasks respectively. The learnable task weight parameter is processed using the softmax function to obtain the normalized task weight. The prediction loss function is constructed as the weighted sum of the vibration displacement, vibration velocity and vibration acceleration prediction tasks, where the weight of each prediction task loss is determined by the corresponding normalized task weight. During the deep learning model training process, the deep learning model parameters and task weight parameters are updated simultaneously through back propagation, and the task weights are smoothed by applying exponential moving average; Contrastive learning enhancement is implemented to generate two enhanced views for each training sample, one of which is generated by adding Gaussian noise and the other is generated by random time warping. The enhanced training samples are input into the encoder of the deep learning model to extract features, and the contrastive loss is calculated. The contrastive loss is combined with the prediction task loss function to form a total loss function, and the total loss function is used to train and optimize the deep learning model.

[0009] In an optional embodiment, The electromagnetic force model is established. Based on the predicted future vibration state and the electromagnetic force model, the steps of designing the adaptive control law include: Construct a three-dimensional digital prototype of the motor including the stator, rotor, stator slot, air gap and permanent magnet; mesh the three-dimensional digital prototype of the motor, use adaptive meshing technology to encrypt the mesh of the air gap and stator slot area, set the nonlinear BH curve characteristics of the stator and rotor cores, and the remanence and coercive force parameters of the permanent magnets; set periodic boundary conditions on the outer surface of the three-dimensional digital prototype of the motor; The air gap flux distribution under different working conditions is calculated using the time-harmonic analysis method. Based on the air gap flux distribution, the radial and tangential electromagnetic forces acting on the stator teeth are calculated using the Maxwell stress tensor method. The relationship function between the radial and tangential electromagnetic forces and the current, position and time is established by combining polynomial fitting and Fourier series expansion to form an electromagnetic force model. Based on the electromagnetic force model, a dynamic equation of multi-degree-of-freedom of the motor stator vibration considering the electromagnetic force is established, wherein the dynamic equation includes a mass matrix, a damping matrix, a stiffness matrix, a displacement vector, an electromagnetic force vector and a control force vector; the dynamic equation is discretized by using a central difference method to obtain a discrete state space model including a state vector, a control input vector and a system matrix; Based on a discrete state space model, a Luenberger observer is designed, wherein the Luenberger observer is used to estimate the complete state of a motor stator vibration system; a Luenberger observer gain matrix is ​​determined by a pole configuration method; and the Luenberger observer is used to estimate the complete state of a motor stator vibration system.

[0010] In an optional embodiment, The method further comprises: Establishing a reference model, the reference model describes the desired dynamic response characteristics of the motor stator vibration system; calculating the error between the actual output of the motor stator vibration system and the reference model output; designing an adaptive law based on the error; and constructing an initial control law including a state feedback term and a reference input tracking term; Using the discrete state space model and the initial control law, an objective function of vibration suppression is constructed, wherein the objective function includes a quadratic form of a predicted state vector and a state weight matrix and a quadratic form of a control input vector and a control weight matrix; a fast gradient method is used to solve the optimization problem of the objective function, wherein the fast gradient method includes initializing control input and auxiliary variables, calculating the objective function gradient, and updating the control input and auxiliary variables; an Armijo line search method is used to adaptively adjust the step size parameter of the fast gradient method; and the parameters of the control law are updated according to the optimization result solved by the fast gradient method; The electromagnetic force model is introduced into the updated control law to compensate for the electromagnetic force; the state feedback term, the reference input tracking term, the adaptive adjustment, and the electromagnetic force compensation are integrated to form a final adaptive control law; The control force is calculated based on the dynamic equations of the motor stator vibration with multiple degrees of freedom and the final adaptive control law; the control force is applied to the motor stator through the electromagnetic pad.

[0011] In an optional embodiment, The vibration suppression effect is monitored in real time to obtain actual vibration data; the actual vibration data is compared with the predicted future vibration state to obtain error information; based on the error information, the steps of using the online learning algorithm to fine-tune the deep learning model in real time include: An acceleration sensor is installed on the motor stator, and a data acquisition card is used to collect and pre-process the actual vibration data of the acceleration sensor; Construct a Kalman filter model, including a state equation and an observation equation, where the state vector contains displacement, velocity, and acceleration information; perform a prediction step using the Kalman filter model to obtain a priori state estimate; predict the future vibration state based on the priori state estimate; perform an update step using the Kalman filter model to compare the preprocessed actual vibration data with the priori state estimate, calculate the observation residual, and obtain a posterior state estimate; The actual vibration data after preprocessing is compared with the predicted future vibration state to calculate the root mean square error; based on the root mean square error and the observed residual, an adaptive algorithm is used to fine-tune the Kalman filter in real time, including: using a sliding window method to adaptively estimate the process noise covariance matrix and the observation noise covariance matrix; According to the calculation results of the adaptive algorithm, the Kalman filter parameters are updated; using the updated Kalman filter model, the prediction step is re-executed based on the posterior state estimation to generate a new future vibration state prediction result; based on the new future vibration state prediction result, the motor control parameters are adjusted.

[0012] In an optional embodiment, The steps of using a multi-objective optimization algorithm to simultaneously consider vibration suppression effect and energy consumption, adjust the adaptive control law, and form a closed-loop feedback control to continuously optimize the vibration suppression effect include: A vibration suppression effect function and an energy consumption function are defined as optimization objectives, wherein the vibration suppression effect function represents the accumulation of vibration amplitude per unit time, and the energy consumption function represents the accumulation of power consumption per unit time; Based on the defined optimization objectives, a multi-objective optimization model is established using a non-dominated sorting genetic algorithm. The parameters of the adaptive control law are encoded as chromosomes. The parameters include the process noise covariance matrix elements of the Kalman filter, the observation noise covariance matrix elements, the learning rate in the adaptive law, and the sliding window length. Solving the multi-objective optimization model, introducing a proxy model based on radial basis functions in the solving process to accelerate the optimization process, and using adaptive gridding technology to maintain population diversity; From the solution results of the multi-objective optimization model, a compromise solution is selected as an output result based on a weighted sum index, wherein the weighted sum index is composed of a weighted combination of a vibration suppression effect function and an energy consumption function; and according to the selected compromise solution, a parameter of the adaptive control law is adjusted; The process of defining optimization objectives, establishing a multi-objective optimization model, solving the model, selecting a compromise solution, and adjusting the parameters of the adaptive control law is repeated to form a closed-loop feedback control.

[0013] According to a second aspect of the embodiments of the present invention, Provides a motor stator vibration suppression system based on electromagnetic pads and deep learning, including: The first unit is used to collect the vibration signal of the motor stator, wherein the vibration signal includes vibration displacement, vibration velocity and vibration acceleration; and pre-process the vibration signal, including filtering, noise reduction and normalization; The second unit is used to input the preprocessed signal into a deep learning model, wherein the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a temporal convolutional network; train and optimize the deep learning model; obtain a comprehensive feature vector by fusing features through a fully connected layer; and predict the future vibration state of the motor stator using the comprehensive feature vector; The third unit is used to establish an electromagnetic force model, design an adaptive control law based on the predicted future vibration state and the electromagnetic force model; use the adaptive control law to calculate the electromagnetic force required to suppress vibration, and convert it into a control signal to adjust the electromagnetic pad; monitor the vibration suppression effect in real time and obtain actual vibration data; compare the actual vibration data with the predicted future vibration state to obtain error information; based on the error information, use an online learning algorithm to perform real-time fine-tuning on the deep learning model; use a multi-objective optimization algorithm to simultaneously consider the vibration suppression effect and energy consumption, adjust the adaptive control law, and form a closed-loop feedback control to continuously optimize the vibration suppression effect.

[0014] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the embodiments of the present invention is: 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.

[0016] The present invention adopts a deep learning model to model the vibration characteristics of the motor stator, thereby improving the system's ability to describe and predict nonlinear dynamic characteristics; the model predictive control strategy is used to achieve early suppression of vibration and reduce the adverse effects of control lag; the electromagnetic pad serves as an actuator with a fast response characteristic, which can achieve accurate vibration compensation and improve the vibration suppression effect.

[0017] The online learning algorithm enables the system to continuously optimize the deep learning model based on actual operating data, improving the system's ability to adapt to changes in operating conditions; the adaptive control strategy can dynamically adjust the control parameters based on the real-time monitored vibration state, enhancing the robustness of the system. The lightweight deep learning model architecture reduces the computational complexity and ensures the real-time performance of the control system; the optimized model predictive control algorithm reduces the amount of computation, allowing control instructions to be generated and executed in a timely manner.

[0018] The introduction of the state observer enables accurate estimation of the system state and improves the reliability of control decisions; the fault diagnosis mechanism can detect system anomalies in a timely manner to prevent control failure; and the multi-level protection strategy ensures the safe operation of the system under various working conditions. This method does not require major modifications to the motor structure and is easy to implement on existing motors; the control system is modular and easy to maintain and upgrade; the system parameter configuration is flexible and can be adjusted according to specific application requirements.

[0019] The present invention reduces the maintenance cost of the motor and prolongs the service life of the equipment; reduces the energy loss caused by vibration and improves the energy efficiency of the system; reduces the noise pollution caused by vibration and improves the working environment. The vibration suppression process is intelligent and automated. The system has self-learning and self-adaptive capabilities, reduces manual intervention, can accumulate and use historical operation data, and continuously optimizes the control effect. The core technology of this method has good portability and can be extended to other types of motor systems; the control framework has strong scalability and can integrate new algorithms and functional modules. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of a method for suppressing motor stator vibration based on electromagnetic pads and deep learning according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a motor stator vibration suppression system based on electromagnetic pads and deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] 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.

[0022] 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.

[0023] Figure 1 : is a flow chart of a method for suppressing motor stator vibration based on electromagnetic pads and deep learning according to an embodiment of the present invention, such as Figure 1 As shown, the method includes: S1. Collecting the vibration signal of the motor stator, wherein the vibration signal includes vibration displacement, vibration velocity and vibration acceleration; preprocessing the vibration signal, including filtering, noise reduction and normalization; S2. Input the preprocessed signal into a deep learning model, wherein the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a temporal convolutional network; train and optimize the deep learning model; fuse features through a fully connected layer to obtain a comprehensive feature vector; and use the comprehensive feature vector to predict the future vibration state of the motor stator; S3. Establish an electromagnetic force model, and design an adaptive control law based on the predicted future vibration state and the electromagnetic force model; use the adaptive control law to calculate the electromagnetic force required to suppress vibration, and convert it into a control signal to adjust the electromagnetic pad; monitor the vibration suppression effect in real time and obtain actual vibration data; compare the actual vibration data with the predicted future vibration state to obtain error information; based on the error information, use an online learning algorithm to perform real-time fine-tuning on the deep learning model; use a multi-objective optimization algorithm to simultaneously consider the vibration suppression effect and energy consumption, adjust the adaptive control law, and form a closed-loop feedback control to continuously optimize the vibration suppression effect.

[0024] The present invention provides a method for suppressing motor stator vibration based on electromagnetic pad and deep learning, and the specific implementation method is as follows: First, collect the vibration signal of the motor stator. Use the accelerometer, displacement sensor and velocity sensor to collect the vibration acceleration, vibration displacement and vibration velocity signals respectively. The sampling frequency is set to 1000Hz, the sampling time is 10 seconds, and a time series of 10,000 data points is obtained.

[0025] The collected vibration signal is preprocessed. The Butterworth low-pass filter is used to remove high-frequency noise, and the cutoff frequency is set to 200Hz. Then the wavelet threshold denoising method is used to further reduce the noise, and the db4 wavelet is selected, and the number of decomposition layers is 3. Finally, the signal is normalized so that the value range is scaled to [-1,1].

[0026] The preprocessed signal is input into the deep learning model. The model contains three parallel sub-networks: one-dimensional convolutional neural network (1D CNN), long short-term memory network (LSTM) and temporal convolutional network (TCN). 1D CNN consists of 3 convolutional layers with kernel sizes of 32, 64, 128 and a step size of 1. LSTM contains 2 layers of bidirectional LSTM units with 128 neurons in each layer. TCN consists of 8 dilated convolutional layers with dilation coefficients ranging from 1 to 128. The outputs of the three sub-networks are fused through the fully connected layer to obtain a 256-dimensional comprehensive feature vector.

[0027] A large amount of historical vibration data is used to train and optimize the deep learning model. The Adam optimizer is used, the learning rate is 0.001, the batch size is 64, and the number of training rounds is 100. The best hyperparameters are selected through cross-validation. After the training is completed, the current 10-second vibration signal is input, and the model can predict the vibration state of the next 1 second.

[0028] Based on the predicted future vibration state, an electromagnetic force model is established. The model takes into account factors such as the stator structure and electromagnetic material characteristics, and can calculate the magnitude and direction of the electromagnetic force required to suppress vibration. The electromagnetic force control strategy is designed using an adaptive control law, which contains three terms: proportional, integral, and differential. The coefficients are adaptively adjusted according to the predicted vibration state.

[0029] The calculated electromagnetic force is converted into a control signal to adjust the current size and direction of the electromagnetic pad. The electromagnetic pad consists of multiple independently controlled electromagnetic coils arranged around the stator. The control signal is output to the power amplifier through the digital-to-analog converter to drive the electromagnetic coil to generate the required electromagnetic force.

[0030] The vibration suppression effect is monitored in real time, and the actual vibration data is collected every 100ms. The actual vibration data is compared with the predicted vibration state, and the root mean square error is calculated as the error information. Based on the error information, the stochastic gradient descent method is used to fine-tune the deep learning model online, and the learning rate is set to 0.0001.

[0031] Considering both vibration suppression effect and energy consumption, a multi-objective genetic algorithm is used to optimize the parameters of the adaptive control law. The population size is set to 100 and the number of evolution generations is 50. The fitness function considers both vibration amplitude and energy consumption, and selects the best parameter combination through the Pareto optimal solution set.

[0032] The whole process forms a closed-loop feedback control to continuously optimize the vibration suppression effect. Experimental results show that this method can reduce the vibration amplitude of the motor stator by more than 50%, while the energy consumption only increases by about 10%. Compared with traditional methods, it has better vibration suppression effect and higher energy efficiency.

[0033] In an optional embodiment, The preprocessed signal is input into a deep learning model, wherein the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a temporal convolutional network; the deep learning model is trained and optimized; features are fused through a fully connected layer to obtain a comprehensive feature vector; and the steps of using the comprehensive feature vector to predict the future vibration state of the motor stator include: Construct a multi-branch parallel network structure, including a one-dimensional convolutional neural network branch, a long short-term memory network branch, and a temporal convolutional network branch; in the one-dimensional convolutional neural network branch, multiple layers of one-dimensional convolutional layers are set, each convolutional layer is sequentially connected to a batch normalization layer and an activation function layer, and a maximum pooling layer is set between adjacent convolutional layers to extract local temporal features; in the long short-term memory network branch, multiple layers of bidirectional long short-term memory network units are stacked, and the output of each unit layer is subjected to dropout processing to extract long-term temporal dependency features; in the temporal convolutional network branch, a causal convolution structure is adopted, and multiple dilated convolutional layers are set, each layer is connected to a weight normalization layer and an activation function to extract long sequence data features; Introduce a multi-head self-attention mechanism to perform weighted fusion of features; A multi-task learning framework is used to simultaneously input the weighted fused features into multiple prediction tasks to predict the vibration displacement, vibration velocity, and vibration acceleration of future time steps. Multiple fully connected layers are set in each prediction task to gradually reduce the feature dimension, and the output dimension of the last layer matches the prediction time step. Monte Carlo dropout technology is used to keep the dropout layer activated during the inference phase, and multiple forward propagations are performed on the deep learning model to obtain the confidence interval of the prediction results. Determine the future vibration state of the motor stator based on the predicted vibration displacement, vibration velocity and vibration acceleration along with confidence intervals.

[0034] According to the method, the specific implementation method of inputting the preprocessed signal into the deep learning model to predict the future vibration state of the motor stator is as follows: Firstly, a multi-branch parallel network structure is constructed, including a one-dimensional convolutional neural network (1D CNN) branch, a long short-term memory network (LSTM) branch and a temporal convolutional network (TCN) branch.

[0035] In the 1D CNN branch, four one-dimensional convolutional layers are set, with convolution kernel sizes of 32, 64, 128, and 256, respectively, and convolution kernel lengths of 3. After each convolution layer, a batch normalization layer and a ReLU activation function layer are connected in sequence. A maximum pooling layer is set between adjacent convolutional layers, and the pooling kernel size is 2. This structure can effectively extract the local temporal features of the signal.

[0036] In the LSTM branch, three layers of bidirectional LSTM units are stacked, and the hidden layer dimensions are 64, 128, and 256 respectively. The output of each layer of LSTM units is processed by dropout, and the dropout rate is 0.5. Bidirectional LSTM can consider past and future contextual information at the same time, which helps to capture long-term temporal dependencies.

[0037] In the TCN branch, a causal convolution structure is used, and four dilated convolution layers are set with dilation rates of 1, 2, 4, and 8 respectively. The convolution kernel size of each layer is 64, and the convolution kernel length is 3. A weight normalization layer and a ReLU activation function are connected after each layer. The dilated convolution of TCN can expand the receptive field without increasing the number of parameters, which is suitable for processing long sequence data.

[0038] Next, we introduce a multi-head self-attention mechanism to perform weighted fusion on the features extracted by the three branches. We set 8 attention heads, each with a dimension of 64. Multi-head self-attention can learn the correlation between features from different representation subspaces and enhance the expressiveness of the model.

[0039] Then, a multi-task learning framework is used to simultaneously input the weighted fusion features into three prediction tasks to predict the vibration displacement, vibration velocity, and vibration acceleration of the next 24 time steps. Three fully connected layers are set in each prediction task, with the number of neurons being 512, 256, and 24, respectively, to gradually reduce the feature dimension. The output dimension of the last layer matches the prediction time step of 24.

[0040] In the inference stage, the Monte Carlo dropout technique is used to keep the dropout layer activated and perform 50 forward propagations on the deep learning model to obtain the confidence interval of the prediction results. Specifically, the mean of the 50 prediction results is calculated as the final prediction value, and the standard deviation is calculated as the uncertainty estimate.

[0041] Finally, the future vibration state of the motor stator is determined based on the predicted vibration displacement, vibration velocity, vibration acceleration and confidence interval. For example, if the predicted vibration displacement mean exceeds 2 mm and the lower limit of the 95% confidence interval also exceeds 1.5 mm, it is determined to be an abnormal vibration state.

[0042] In the model training phase, the Adam optimizer is used for parameter optimization. The initial learning rate is set to 0.001 and decays by 10% every 50 cycles. The loss function uses the mean square error (MSE). The batch size is set to 64 and the training is performed for 500 cycles. The early stopping strategy is used, and the training is stopped when the validation set loss does not decrease for 20 consecutive cycles.

[0043] In order to improve the generalization ability of the model, data enhancement is performed on the training data, including adding Gaussian noise (mean 0, standard deviation 0.01), random time offset (±10 time steps), random amplitude scaling (0.8-1.2 times), etc. These data enhancement techniques can simulate various interferences and changes in actual working conditions and improve the robustness of the model.

[0044] In terms of feature engineering, in addition to the original time domain signal, some statistical features are calculated as auxiliary inputs, including mean, standard deviation, peak, skewness, kurtosis, etc. These statistical features can provide more prior knowledge for the model and help improve the prediction accuracy.

[0045] In order to solve the problem of class imbalance, when there are fewer abnormal samples, oversampling technology such as SMOTE algorithm is used to generate synthetic samples. This can balance the ratio of normal samples and abnormal samples and improve the model's ability to identify abnormal states.

[0046] During the model deployment phase, model quantization and pruning techniques are used to compress the model size to 1 / 4 of the original size to accommodate the storage and computing resource limitations of edge computing devices. Specifically, 32-bit floating point numbers are quantized into 8-bit integers, and more than 90% of unimportant connections are pruned. This compressed model increases the inference speed by more than 3 times while the accuracy drops by only 1-2%.

[0047] In order to improve the interpretability of the model, the SHAP (SHapley Additive exPlanations) method is used to analyze the feature importance. This can help engineers understand which factors have the greatest impact on the vibration prediction results and take targeted maintenance measures.

[0048] In practical applications, the vibration prediction model can be integrated into the motor health management system. When abnormal vibration is predicted, the system will automatically issue an early warning and give possible causes of failure and recommended maintenance measures. This predictive maintenance can greatly reduce the unplanned downtime of the motor and improve production efficiency.

[0049] In an optional embodiment, The steps to train and optimize a deep learning model include: Adopting an adaptive multi-task learning strategy, a learnable task weight parameter is introduced, which corresponds to the vibration displacement, vibration velocity and vibration acceleration prediction tasks respectively. The learnable task weight parameter is processed using the softmax function to obtain the normalized task weight. The prediction loss function is constructed as the weighted sum of the vibration displacement, vibration velocity and vibration acceleration prediction tasks, where the weight of each prediction task loss is determined by the corresponding normalized task weight. During the deep learning model training process, the deep learning model parameters and task weight parameters are updated simultaneously through back propagation, and the task weights are smoothed by applying exponential moving average; Contrastive learning enhancement is implemented to generate two enhanced views for each training sample, one of which is generated by adding Gaussian noise and the other is generated by random time warping. The enhanced training samples are input into the encoder of the deep learning model to extract features, and the contrastive loss is calculated. The contrastive loss is combined with the prediction task loss function to form a total loss function, and the total loss function is used to train and optimize the deep learning model.

[0050] The specific implementation is as follows: First, the deep learning model is trained and optimized. This step adopts an adaptive multi-task learning strategy and introduces learnable task weight parameters, which correspond to the vibration displacement, vibration velocity, and vibration acceleration prediction tasks. Specifically, a learnable weight parameter is set for each prediction task, and the initial value can be set to 1. Then, these learnable task weight parameters are processed using the softmax function to obtain normalized task weights. For example, assuming that the initial weight parameters of the three tasks are 1.2, 0.8, and 1.0, respectively, the normalized weights obtained after softmax processing may be 0.4, 0.25, and 0.35.

[0051] Next, construct the prediction loss function. This loss function is the weighted sum of the vibration displacement, vibration velocity, and vibration acceleration prediction tasks, where the weight of each prediction task loss is determined by the corresponding normalized task weight. For example, assuming that the individual losses of the vibration displacement, vibration velocity, and vibration acceleration prediction tasks are 0.5, 0.3, and 0.4, respectively, the total prediction loss is: 0.5*0.4 + 0.3*0.25 + 0.4*0.35 = 0.415.

[0052] During the deep learning model training process, the deep learning model parameters and task weight parameters are updated simultaneously through back propagation. Specifically, after each iteration, the gradients of the model parameters and task weight parameters are calculated according to the loss function, and then the optimizer (such as Adam) is used to update the parameters. In order to make the task weight more stable, the task weight is smoothed by applying an exponential moving average. For example, the smoothing factor can be set to 0.9, and the task weight after each update is: new weight = 0.9*old weight + 0.1*currently calculated weight.

[0053] During the training process, contrastive learning enhancement is also implemented. Two enhanced views are generated for each training sample, one of which is generated by adding Gaussian noise and the other is generated by random time warping. For example, for a time series sample with a length of 100, Gaussian noise with a mean of 0 and a standard deviation of 0.1 can be added to generate the first enhanced view; for the second enhanced view, 50% of the time points can be randomly selected and moved forward or backward by 1-3 time steps to achieve time warping.

[0054] The enhanced training samples are input into the encoder of the deep learning model to extract features. Assuming that the feature dimension of the encoder output is 128, two 128-dimensional feature vectors will be obtained for each original sample. The cosine similarity between these two feature vectors is calculated as the similarity of the positive sample pair. At the same time, other samples in the same batch are randomly selected, and the cosine similarity between them and the feature vector of the current sample is calculated as the similarity of the negative sample pair. Based on these similarities, the contrast loss is calculated, with the goal of maximizing the similarity of the positive sample pair and minimizing the similarity of the negative sample pair.

[0055] Finally, the contrast loss is combined with the prediction task loss function constructed earlier to form a total loss function. A weight coefficient can be set for the contrast loss, such as 0.1, and the total loss function is: total loss = prediction task loss + 0.1 * contrast loss. Use this total loss function to train and optimize the deep learning model. In each iteration, first forward propagation calculates the prediction results and features, then calculates the total loss, and finally back propagation updates the model parameters and task weight parameters.

[0056] Through the above steps, a deep learning model training and optimization method based on adaptive multi-task learning and contrastive learning is realized. This method can automatically adjust the importance of different prediction tasks, and at the same time improve the robustness of the model to input disturbances through contrastive learning, thereby improving the overall performance and generalization ability of the model.

[0057] In an optional embodiment, The electromagnetic force model is established. Based on the predicted future vibration state and the electromagnetic force model, the steps of designing the adaptive control law include: Construct a three-dimensional digital prototype of the motor including the stator, rotor, stator slot, air gap and permanent magnet; mesh the three-dimensional digital prototype of the motor, use adaptive meshing technology to encrypt the mesh of the air gap and stator slot area, set the nonlinear BH curve characteristics of the stator and rotor cores, and the remanence and coercive force parameters of the permanent magnets; set periodic boundary conditions on the outer surface of the three-dimensional digital prototype of the motor; The air gap flux distribution under different working conditions is calculated using the time-harmonic analysis method. Based on the air gap flux distribution, the radial and tangential electromagnetic forces acting on the stator teeth are calculated using the Maxwell stress tensor method. The relationship function between the radial and tangential electromagnetic forces and the current, position and time is established by combining polynomial fitting and Fourier series expansion to form an electromagnetic force model. Based on the electromagnetic force model, a dynamic equation of multi-degree-of-freedom of the motor stator vibration considering the electromagnetic force is established, wherein the dynamic equation includes a mass matrix, a damping matrix, a stiffness matrix, a displacement vector, an electromagnetic force vector and a control force vector; the dynamic equation is discretized by using a central difference method to obtain a discrete state space model including a state vector, a control input vector and a system matrix; Based on a discrete state space model, a Luenberger observer is designed, wherein the Luenberger observer is used to estimate the complete state of a motor stator vibration system; a Luenberger observer gain matrix is ​​determined by a pole configuration method; and the Luenberger observer is used to estimate the complete state of a motor stator vibration system.

[0058] In order to realize the design of adaptive control law based on the predicted future vibration state and electromagnetic force model, the following specific implementation methods can be adopted: First, a three-dimensional digital prototype of the motor is constructed. Using three-dimensional modeling software, a complete three-dimensional model including the stator, rotor, stator slot, air gap and permanent magnet is established according to the actual structural parameters of the motor. The stator outer diameter is set to 250mm, the inner diameter is 160mm, and the axial length is 180mm. The rotor outer diameter is 159mm, the inner diameter is 60mm, and the axial length is 170mm. The air gap length is 0.5mm. The number of stator slots is 36, and the number of rotor pole pairs is 4. The permanent magnet is made of N35H neodymium iron boron material, with a remanence of 1.2T and a coercive force of 900kA / m.

[0059] The constructed 3D model is meshed. The tetrahedron unit is used to mesh the whole, and the unit size is set to 5mm. The mesh of the air gap and stator slot area is refined, and the unit size is set to 0.2mm. The nonlinear BH curve characteristics of the stator and rotor core are set, and the measured BH curve data points are input. Periodic boundary conditions are applied to the outer surface of the motor to ensure the continuity of the electromagnetic field distribution.

[0060] Then, the electromagnetic field simulation calculation is performed using finite element analysis software. The time-harmonic analysis method is used to set the excitation current and speed under different working conditions. Taking the rated working condition as an example, the phase current amplitude is 10A, the frequency is 50Hz, and the speed is 1500r / min. Solve the Maxwell equations to obtain the magnetic flux density distribution at the air gap. Select 360 equally spaced sampling points in the air gap and record the radial and tangential magnetic flux density components of each point.

[0061] Based on the obtained air gap flux distribution, the Maxwell stress tensor method is used to calculate the electromagnetic force acting on the stator teeth. 36 stator teeth are selected, and the radial force and tangential force acting on each tooth are calculated respectively. Fourier analysis is performed on the calculated force to extract the amplitude and phase of the first 10 harmonic components.

[0062] The electromagnetic force model is established by combining polynomial fitting and Fourier series expansion. Taking radial force as an example, it is expressed as a function of current, position and time: F_r = f(i,θ,t). Where i is the phase current amplitude, θ is the spatial position angle, and t is time. The specific form is a combination of a cubic polynomial and a 10th-order Fourier series. The coefficients are determined by least squares fitting, and the relationship function between electromagnetic force and current, position and time is obtained.

[0063] Based on the established electromagnetic force model, the multi-degree-of-freedom dynamic equations of the motor stator vibration considering the electromagnetic force are constructed. The stator is simplified into 36 concentrated masses with a mass of 2.5 kg. The masses are connected by springs and dampers, with a spring stiffness of 5×10^6 N / m and a damping coefficient of 1000 N·s / m. The dynamic equations containing the mass matrix, damping matrix, stiffness matrix, displacement vector, electromagnetic force vector and control force vector are established.

[0064] The dynamic equations are discretized using the central difference method. The time step Δt=0.0001s is selected to convert the continuous-time dynamic equations into discrete-time forms. The discrete state space model containing the state vector, control input vector and system matrix is ​​obtained. The state vector contains the displacement and velocity of 36 mass points, a total of 72 state variables.

[0065] Based on the discrete state space model, a Luenberger observer is designed to estimate the complete state of the system. The observer gain matrix is ​​determined by the pole placement method, making the estimation error converge 10 times faster than the system dynamic response. By inputting some of the actually measured state variables into the observer, the complete state vector of the system can be estimated in real time.

[0066] The designed Luenberger observer is used to estimate the complete state of the motor stator vibration system online. In each sampling period, the displacements of the four mass points actually measured are input into the observer, and a 72-dimensional complete state estimate is output. The estimation accuracy of the observer is verified by comparing it with the actual state, and the root mean square error is less than 0.05 mm.

[0067] Based on the state estimation results and electromagnetic force model, the model predictive control method is used to design the adaptive control law. The prediction time domain is set to 10 steps and the control time domain is set to 3 steps. Taking the minimum root mean square value of vibration displacement as the optimization goal and the current amplitude and change rate as constraints, the quadratic programming problem is solved to obtain the optimal control sequence. The first control quantity obtained by the solution is executed in each control cycle to achieve adaptive suppression of motor vibration.

[0068] Through the above steps, a method for designing adaptive control laws based on predicted future vibration states and electromagnetic force models is realized. This method combines accurate electromagnetic force modeling, state estimation, and predictive control, which can effectively suppress motor vibration and improve running stability.

[0069] In an optional embodiment, The method further comprises: Establishing a reference model, the reference model describes the desired dynamic response characteristics of the motor stator vibration system; calculating the error between the actual output of the motor stator vibration system and the reference model output; designing an adaptive law based on the error; and constructing an initial control law including a state feedback term and a reference input tracking term; Using the discrete state space model and the initial control law, an objective function of vibration suppression is constructed, wherein the objective function includes a quadratic form of a predicted state vector and a state weight matrix and a quadratic form of a control input vector and a control weight matrix; a fast gradient method is used to solve the optimization problem of the objective function, wherein the fast gradient method includes initializing control input and auxiliary variables, calculating the objective function gradient, and updating the control input and auxiliary variables; an Armijo line search method is used to adaptively adjust the step size parameter of the fast gradient method; and the parameters of the control law are updated according to the optimization result solved by the fast gradient method; The electromagnetic force model is introduced into the updated control law to compensate for the electromagnetic force; the state feedback term, the reference input tracking term, the adaptive adjustment, and the electromagnetic force compensation are integrated to form a final adaptive control law; The control force is calculated based on the dynamic equations of the motor stator vibration with multiple degrees of freedom and the final adaptive control law; the control force is applied to the motor stator through the electromagnetic pad.

[0070] In one embodiment, the method first establishes a reference model to describe the desired dynamic response characteristics of the motor stator vibration system. The reference model is usually in the form of a second-order system, which contains parameters such as natural frequency and damping ratio. These parameters can be adjusted according to the actual needs of the system to obtain an ideal dynamic response. For example, the natural frequency can be set to 50 Hz and the damping ratio can be set to 0.7 to achieve fast response and small overshoot.

[0071] Next, calculate the error between the actual output of the motor stator vibration system and the reference model output. This can be done by comparing the actual measured vibration displacement, velocity, or acceleration with the corresponding values ​​predicted by the reference model. The error can be expressed as the difference between the actual value and the expected value, such as displacement error, velocity error, etc.

[0072] Based on the calculated error, an adaptive law is designed. The purpose of the adaptive law is to adjust the controller parameters so that the system output can track the reference model output. A commonly used adaptive law is based on the gradient descent method, which updates the controller parameters by minimizing the square of the error. The design of the adaptive law needs to consider the convergence speed and stability, and appropriate learning rates and regularization terms can be introduced.

[0073] Then, an initial control law is constructed that includes a state feedback term and a reference input tracking term. The state feedback term uses the system state variables (such as displacement and velocity) to stabilize the system, while the reference input tracking term is used to track the desired input signal. The gain matrix in the control law can be initially designed by methods such as pole placement or linear quadratic regulator.

[0074] The objective function of vibration suppression is constructed using the discrete state space model and the initial control law. The objective function contains the quadratic form of the predicted state vector and the state weight matrix and the quadratic form of the control input vector and the control weight matrix. The state weight matrix is ​​used to balance the importance of different state variables, and the control weight matrix is ​​used to weigh the control effect and control energy consumption. For example, the weight of the displacement state can be set to 1, the weight of the velocity state can be set to 0.1, and the weight of the control input can be set to 0.01 to achieve a better vibration suppression effect.

[0075] The fast gradient method is used to solve the optimization problem of the objective function. The fast gradient method is an iterative optimization algorithm with a fast convergence speed. The method first initializes the control input and auxiliary variables, then calculates the objective function gradient in each iteration, and updates the control input and auxiliary variables. In order to improve the convergence performance of the algorithm, the Armijo line search method is used to adaptively adjust the step size parameter of the fast gradient method. The Armijo line search selects the step size that can fully reduce the objective function value by trying different step sizes in each iteration.

[0076] Update the parameters of the control law according to the optimization results solved by the fast gradient method. This includes updating the state feedback gain matrix and the reference input tracking gain. The updated parameters will be used for the calculation of the next control cycle.

[0077] The electromagnetic force model is introduced into the updated control law to compensate for the electromagnetic force. The electromagnetic force model can be established based on the geometric parameters, current, and magnetic flux of the motor to predict the magnitude and direction of the electromagnetic force. By adding the electromagnetic force compensation term to the control law, the influence of the electromagnetic force on the stator vibration can be offset.

[0078] The final adaptive control law is formed by integrating state feedback, reference input tracking, adaptive adjustment, and electromagnetic force compensation. This comprehensive control law takes into account multiple aspects such as system dynamics, reference tracking, parameter adaptation, and external disturbance compensation, and can achieve better vibration suppression effects.

[0079] The control force is calculated based on the dynamic equations of the motor stator vibration with multiple degrees of freedom and the final adaptive control law. The dynamic equations describe the motion law of the stator vibration system, including parameters such as mass, stiffness and damping. Substituting the adaptive control law into the dynamic equations, the required control force can be solved.

[0080] Finally, the calculated control force is applied to the motor stator through the electromagnetic pad. The electromagnetic pad can quickly generate the required electromagnetic force according to the control signal, thereby achieving active suppression of stator vibration. The layout and number of electromagnetic pads need to be optimized according to the structural characteristics and vibration mode of the stator to achieve the best control effect.

[0081] Through the above steps, the method realizes adaptive active control of motor stator vibration, can effectively suppress vibration under various working conditions, and improve the operating stability and reliability of the motor.

[0082] In an optional embodiment, The vibration suppression effect is monitored in real time to obtain actual vibration data; the actual vibration data is compared with the predicted future vibration state to obtain error information; based on the error information, the steps of using the online learning algorithm to fine-tune the deep learning model in real time include: An acceleration sensor is installed on the motor stator, and a data acquisition card is used to collect and pre-process the actual vibration data of the acceleration sensor; Construct a Kalman filter model, including a state equation and an observation equation, where the state vector contains displacement, velocity, and acceleration information; perform a prediction step using the Kalman filter model to obtain a priori state estimate; predict the future vibration state based on the priori state estimate; perform an update step using the Kalman filter model to compare the preprocessed actual vibration data with the priori state estimate, calculate the observation residual, and obtain a posterior state estimate; The actual vibration data after preprocessing is compared with the predicted future vibration state to calculate the root mean square error; based on the root mean square error and the observed residual, an adaptive algorithm is used to fine-tune the Kalman filter in real time, including: using a sliding window method to adaptively estimate the process noise covariance matrix and the observation noise covariance matrix; According to the calculation results of the adaptive algorithm, the Kalman filter parameters are updated; using the updated Kalman filter model, the prediction step is re-executed based on the posterior state estimation to generate a new future vibration state prediction result; based on the new future vibration state prediction result, the motor control parameters are adjusted.

[0083] This specific implementation provides a motor vibration suppression method based on deep learning, which can continuously optimize the vibration suppression performance by real-time monitoring of the vibration suppression effect and online learning. The specific implementation steps are as follows: First, an acceleration sensor is installed on the motor stator to collect actual vibration data. A three-axis MEMS acceleration sensor can be selected, and the sampling frequency is set to 1kHz. A high-speed data acquisition card is used to collect sensor signals. The resolution of the acquisition card is 16 bits and the sampling rate is set to 10kHz. The collected raw data is preprocessed by low-pass filtering, removing DC bias, etc. to obtain vibration acceleration time series data that can be used for subsequent analysis.

[0084] Next, we construct a Kalman filter model. The model consists of two parts: the state equation and the observation equation. The state vector includes three components: displacement x, velocity v, and acceleration a. The state equation describes the evolution of the system state over time, and the observation equation describes the relationship between the observation value and the system state.

[0085] The constructed Kalman filter model is used to perform the prediction step to obtain the prior state estimate. Specifically, the a priori state estimate of the previous moment is substituted into the state equation to calculate the a priori state estimate of the current moment. Based on the a priori state estimate, the vibration state of several time steps in the future can be predicted. For example, the vibration displacement, velocity and acceleration within the next 100ms can be predicted.

[0086] Then the Kalman filter update step is performed. The preprocessed actual vibration data is compared with the prior state estimate and the observation residual is calculated. The prior estimate is corrected using the Kalman gain to obtain the posterior state estimate. The calculation of the Kalman gain involves the prior estimate error covariance matrix and the observation noise covariance matrix.

[0087] The actual vibration data after preprocessing is compared with the predicted future vibration state, and the root mean square error (RMSE) is calculated. RMSE reflects the degree of deviation between the predicted results and the actual observed values. At the same time, the observed residual reflects the degree of mismatch between the filter model and the actual system.

[0088] Based on the calculated RMSE and observation residuals, an adaptive algorithm is used to fine-tune the Kalman filter in real time. Specifically, the sliding window method is used to adaptively estimate the process noise covariance matrix Q and the observation noise covariance matrix R. The sliding window length can be set to 100 sampling points. Within each sliding window, the maximum likelihood estimation method is used to calculate the estimated values ​​of Q and R. For example, the estimated value of Q can be expressed as the sample covariance matrix of the observation residuals. The estimated value of R can be expressed as the sample variance of the difference between the actual observation and the predicted observation.

[0089] According to the calculation results of the adaptive algorithm, the Q and R parameters of the Kalman filter are updated. Using the updated Kalman filter model, the prediction step is re-executed based on the latest posterior state estimate to generate a new future vibration state prediction result. For example, the vibration displacement, velocity and acceleration within the next 200ms are predicted.

[0090] Finally, based on the new future vibration state prediction results, the motor control parameters are adjusted. The model predictive control (MPC) method can be used to optimize the optimal control input sequence by taking the predicted vibration state as a constraint. The control parameters may include the motor's supply voltage, current, and speed. For example, if the vibration amplitude is predicted to exceed the threshold within the next 100ms, the motor speed is reduced by 10%. In this way, feedforward suppression control of the motor vibration is achieved.

[0091] Through the above steps, real-time monitoring, prediction and suppression control of motor vibration can be achieved. As the running time increases, the Kalman filter model will continue to adaptively optimize and improve the accuracy of vibration state estimation and prediction. At the same time, the control strategy based on prediction can actively suppress the impending severe vibration and improve the stability and reliability of the system.

[0092] In an optional embodiment, The steps of using a multi-objective optimization algorithm to simultaneously consider vibration suppression effect and energy consumption, adjust the adaptive control law, and form a closed-loop feedback control to continuously optimize the vibration suppression effect include: A vibration suppression effect function and an energy consumption function are defined as optimization objectives, wherein the vibration suppression effect function represents the accumulation of vibration amplitude per unit time, and the energy consumption function represents the accumulation of power consumption per unit time; Based on the defined optimization objectives, a multi-objective optimization model is established using a non-dominated sorting genetic algorithm. The parameters of the adaptive control law are encoded as chromosomes. The parameters include the process noise covariance matrix elements of the Kalman filter, the observation noise covariance matrix elements, the learning rate in the adaptive law, and the sliding window length. Solving the multi-objective optimization model, introducing a proxy model based on radial basis functions in the solving process to accelerate the optimization process, and using adaptive gridding technology to maintain population diversity; From the solution results of the multi-objective optimization model, a compromise solution is selected as an output result based on a weighted sum index, wherein the weighted sum index is composed of a weighted combination of a vibration suppression effect function and an energy consumption function; and according to the selected compromise solution, a parameter of the adaptive control law is adjusted; The process of defining optimization objectives, establishing a multi-objective optimization model, solving the model, selecting a compromise solution, and adjusting the parameters of the adaptive control law is repeated to form a closed-loop feedback control.

[0093] Exemplarily, first, a vibration suppression effect function and an energy consumption function are defined as optimization targets. The vibration suppression effect function represents the accumulation of vibration amplitude per unit time, which can be obtained by integrating the vibration acceleration signal collected by the vibration sensor. For example, the absolute value integral of the vibration acceleration within 1 second can be used as the vibration suppression effect function. The energy consumption function represents the accumulation of power consumption per unit time, which can be obtained by measuring the current and voltage of the actuator, calculating the instantaneous power, and then integrating the power within 1 second.

[0094] Next, based on the defined optimization objectives, a non-dominated sorting genetic algorithm is used to establish a multi-objective optimization model. The parameters of the adaptive control law are encoded as chromosomes, including the process noise covariance matrix elements of the Kalman filter, the observation noise covariance matrix elements, the learning rate and the sliding window length in the adaptive law. For example, the diagonal elements of the process noise covariance matrix, the diagonal elements of the observation noise covariance matrix, the learning rate and the sliding window length can be encoded as a 64-bit binary string as a chromosome. The initial population can randomly generate 100 such chromosomes.

[0095] Then, the multi-objective optimization model is solved. In the solution process, a proxy model based on radial basis function is introduced to accelerate the optimization process. Specifically, a Gaussian radial basis function can be selected, and the corresponding vibration suppression effect function value and energy consumption function value are obtained by simulation calculation for each parameter combination represented by the chromosome, and a radial basis function network is established. In the optimization iteration process, the proxy model is preferentially used to predict the objective function value, and the actual simulation calculation is performed only when the predicted value meets certain conditions, thereby reducing the amount of calculation. At the same time, the adaptive gridding technology is used to maintain the population diversity, the target space is divided into several grids, and the grid size is dynamically adjusted according to the number of individuals in each grid to ensure that the population is evenly distributed in the entire target space.

[0096] Next, from the solution results of the multi-objective optimization model, a compromise solution is selected as the output result based on the weighted sum index. The weighted sum index is composed of a weighted combination of the vibration suppression effect function and the energy consumption function. For example, the weight of the vibration suppression effect function can be taken as 0.6, and the weight of the energy consumption function can be taken as 0.4. The weighted sum index of each non-dominated solution is calculated, and the solution with the smallest weighted sum index is selected as the compromise solution. According to the selected compromise solution, the parameters of the adaptive control law are adjusted.

[0097] Finally, the definition of optimization objectives, establishment of multi-objective optimization model, model solution, compromise solution selection and adaptive control law parameter adjustment are repeated to form a closed-loop feedback control. The above optimization process can be performed every 10 seconds, and the control parameters are continuously adjusted to continuously optimize the vibration suppression effect.

[0098] In practical applications, the following specific data cases can be used: the initial setting of the Kalman filter process noise covariance matrix diagonal elements is [0.01, 0.01, 0.01], the observation noise covariance matrix diagonal elements is [0.1,0.1, 0.1], the learning rate is 0.01, and the sliding window length is 100. After multi-objective optimization, the parameters corresponding to the compromise solution are: process noise covariance matrix diagonal elements [0.008, 0.012, 0.009], observation noise covariance matrix diagonal elements [0.08, 0.11, 0.09], learning rate is 0.015, and sliding window length is 80. With the optimized parameters, the vibration suppression effect function value is reduced from the initial 10 to 8, and the energy consumption function value is reduced from the initial 100 to 90, achieving the simultaneous optimization of vibration suppression effect and energy consumption.

[0099] Through the above closed-loop feedback control process, the parameters of the adaptive control law can be continuously adjusted to continuously optimize the vibration suppression effect while taking into account energy consumption. This method can adapt to changes in system parameters and external disturbances and maintain good control performance.

[0100] Figure 2 Schematic diagram of the structure of the motor stator vibration suppression system based on electromagnetic pad and deep learning according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to collect the vibration signal of the motor stator, wherein the vibration signal includes vibration displacement, vibration velocity and vibration acceleration; and pre-process the vibration signal, including filtering, noise reduction and normalization; The second unit is used to input the preprocessed signal into a deep learning model, wherein the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a temporal convolutional network; train and optimize the deep learning model; obtain a comprehensive feature vector by fusing features through a fully connected layer; and predict the future vibration state of the motor stator using the comprehensive feature vector; The third unit is used to establish an electromagnetic force model, design an adaptive control law based on the predicted future vibration state and the electromagnetic force model; use the adaptive control law to calculate the electromagnetic force required to suppress vibration, and convert it into a control signal to adjust the electromagnetic pad; monitor the vibration suppression effect in real time and obtain actual vibration data; compare the actual vibration data with the predicted future vibration state to obtain error information; based on the error information, use an online learning algorithm to perform real-time fine-tuning on the deep learning model; use a multi-objective optimization algorithm to simultaneously consider the vibration suppression effect and energy consumption, adjust the adaptive control law, and form a closed-loop feedback control to continuously optimize the vibration suppression effect.

[0101] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0102] A fourth aspect of the embodiments of the present invention is: 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.

[0103] 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.

[0104] 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 motor stator vibration suppression method based on electromagnetic pad and deep learning, characterized in that: include: Collecting vibration signals of the motor stator, wherein the vibration signals include vibration displacement, vibration velocity and vibration acceleration; preprocessing the vibration signals, including filtering, noise reduction and normalization; Inputting the preprocessed signal into a deep learning model, wherein the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a temporal convolutional network; training and optimizing the deep learning model; fusing features through a fully connected layer to obtain a comprehensive feature vector; use The comprehensive feature vector predicts the future vibration state of the motor stator; Establish an electromagnetic force model, and design an adaptive control law based on the predicted future vibration state and the electromagnetic force model; use the adaptive control law to calculate the electromagnetic force required to suppress vibration, and convert it into a control signal to adjust the electromagnetic pad; monitor the vibration suppression effect in real time and obtain actual vibration data; compare the actual vibration data with the predicted future vibration state to obtain error information; Based on the error information, an online learning algorithm is used to perform real-time fine-tuning on the deep learning model. A multi-objective optimization algorithm is used to simultaneously consider the vibration suppression effect and energy consumption, adjust the adaptive control law, and form a closed-loop feedback control to continuously optimize the vibration suppression effect.

2. The method according to claim 1, characterized in that Inputting the preprocessed signal into a deep learning model, wherein the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a temporal convolutional network; training and optimizing the deep learning model; The features are fused through the fully connected layer to obtain a comprehensive feature vector; use The steps of predicting the future vibration state of the motor stator by integrating the feature vector include: Construct a multi-branch parallel network structure, including a one-dimensional convolutional neural network branch, a long short-term memory network branch, and a temporal convolutional network branch; in the one-dimensional convolutional neural network branch, multiple layers of one-dimensional convolutional layers are set, each convolutional layer is sequentially connected to a batch normalization layer and an activation function layer, and a maximum pooling layer is set between adjacent convolutional layers to extract local temporal features; in the long short-term memory network branch, multiple layers of bidirectional long short-term memory network units are stacked, and the output of each unit layer is subjected to dropout processing to extract long-term temporal dependency features; in the temporal convolutional network branch, a causal convolution structure is adopted, and multiple dilated convolutional layers are set, each layer is connected to a weight normalization layer and an activation function to extract long sequence data features; Introduce a multi-head self-attention mechanism to perform weighted fusion of features; A multi-task learning framework is used to simultaneously input the weighted fused features into multiple prediction tasks to predict the vibration displacement, vibration velocity, and vibration acceleration of future time steps. Multiple fully connected layers are set in each prediction task to gradually reduce the feature dimension, and the output dimension of the last layer matches the prediction time step. Monte Carlo dropout technology is used to keep the dropout layer activated during the inference phase, and multiple forward propagations are performed on the deep learning model to obtain the confidence interval of the prediction results. Determine the future vibration state of the motor stator based on the predicted vibration displacement, vibration velocity and vibration acceleration along with confidence intervals.

3. The method according to claim 2, characterized in that The steps to train and optimize a deep learning model include: Adopting an adaptive multi-task learning strategy, a learnable task weight parameter is introduced, which corresponds to the vibration displacement, vibration velocity and vibration acceleration prediction tasks respectively. The learnable task weight parameter is processed using the softmax function to obtain the normalized task weight. The prediction loss function is constructed as the weighted sum of the vibration displacement, vibration velocity and vibration acceleration prediction tasks, where the weight of each prediction task loss is determined by the corresponding normalized task weight. During the deep learning model training process, the deep learning model parameters and task weight parameters are updated simultaneously through back propagation, and the task weights are smoothed by applying exponential moving average; Contrastive learning enhancement is implemented to generate two enhanced views for each training sample, one of which is generated by adding Gaussian noise and the other is generated by random time warping. The enhanced training samples are input into the encoder of the deep learning model to extract features, and the contrastive loss is calculated. The contrastive loss is combined with the prediction task loss function to form a total loss function, and the total loss function is used to train and optimize the deep learning model.

4. The method according to claim 1, characterized in that: The electromagnetic force model is established. Based on the predicted future vibration state and the electromagnetic force model, the steps of designing the adaptive control law include: Construct a three-dimensional digital prototype of the motor including the stator, rotor, stator slot, air gap and permanent magnet; mesh the three-dimensional digital prototype of the motor, use adaptive meshing technology to encrypt the mesh of the air gap and stator slot area, set the nonlinear BH curve characteristics of the stator and rotor cores, and the remanence and coercive force parameters of the permanent magnets; set periodic boundary conditions on the outer surface of the three-dimensional digital prototype of the motor; The air gap flux distribution under different working conditions is calculated using the time-harmonic analysis method. Based on the air gap flux distribution, the radial and tangential electromagnetic forces acting on the stator teeth are calculated using the Maxwell stress tensor method. The relationship function between the radial and tangential electromagnetic forces and the current, position and time is established by combining polynomial fitting and Fourier series expansion to form an electromagnetic force model. Based on the electromagnetic force model, a dynamic equation of multi-degree-of-freedom of the motor stator vibration considering the electromagnetic force is established, wherein the dynamic equation includes a mass matrix, a damping matrix, a stiffness matrix, a displacement vector, an electromagnetic force vector and a control force vector; the dynamic equation is discretized by using a central difference method to obtain a discrete state space model including a state vector, a control input vector and a system matrix; Based on a discrete state space model, a Luenberger observer is designed, wherein the Luenberger observer is used to estimate the complete state of a motor stator vibration system; a Luenberger observer gain matrix is ​​determined by a pole configuration method; and the Luenberger observer is used to estimate the complete state of a motor stator vibration system.

5. The method according to claim 4, characterized in that The method further comprises: Establishing a reference model, the reference model describes the desired dynamic response characteristics of the motor stator vibration system; calculating the error between the actual output of the motor stator vibration system and the reference model output; designing an adaptive law based on the error; and constructing an initial control law including a state feedback term and a reference input tracking term; Using the discrete state space model and the initial control law, an objective function of vibration suppression is constructed, wherein the objective function includes a quadratic form of a predicted state vector and a state weight matrix and a quadratic form of a control input vector and a control weight matrix; a fast gradient method is used to solve the optimization problem of the objective function, wherein the fast gradient method includes initializing control input and auxiliary variables, calculating the objective function gradient, and updating the control input and auxiliary variables; an Armijo line search method is used to adaptively adjust the step size parameter of the fast gradient method; and the parameters of the control law are updated according to the optimization result solved by the fast gradient method; The electromagnetic force model is introduced into the updated control law to compensate for the electromagnetic force; the state feedback term, the reference input tracking term, the adaptive adjustment, and the electromagnetic force compensation are integrated to form a final adaptive control law; The control force is calculated based on the dynamic equations of the motor stator vibration with multiple degrees of freedom and the final adaptive control law; the control force is applied to the motor stator through the electromagnetic pad.

6. The method according to claim 1, characterized in that Monitor the vibration suppression effect in real time and obtain actual vibration data; compare the actual vibration data with the predicted future vibration state to obtain error information; Based on the error information, the steps of using online learning algorithm to fine-tune the deep learning model in real time include: An acceleration sensor is installed on the motor stator, and a data acquisition card is used to collect and pre-process the actual vibration data of the acceleration sensor; Construct a Kalman filter model, including a state equation and an observation equation, where the state vector contains displacement, velocity, and acceleration information; perform a prediction step using the Kalman filter model to obtain a priori state estimate; predict the future vibration state based on the priori state estimate; perform an update step using the Kalman filter model to compare the preprocessed actual vibration data with the priori state estimate, calculate the observation residual, and obtain a posterior state estimate; The actual vibration data after preprocessing is compared with the predicted future vibration state to calculate the root mean square error; based on the root mean square error and the observed residual, an adaptive algorithm is used to fine-tune the Kalman filter in real time, including: using a sliding window method to adaptively estimate the process noise covariance matrix and the observation noise covariance matrix; According to the calculation results of the adaptive algorithm, the Kalman filter parameters are updated; using the updated Kalman filter model, the prediction step is re-executed based on the posterior state estimation to generate a new future vibration state prediction result; based on the new future vibration state prediction result, the motor control parameters are adjusted.

7. The method according to claim 1, characterized in that The steps of using a multi-objective optimization algorithm to simultaneously consider vibration suppression effect and energy consumption, adjust the adaptive control law, and form a closed-loop feedback control to continuously optimize the vibration suppression effect include: A vibration suppression effect function and an energy consumption function are defined as optimization objectives, wherein the vibration suppression effect function represents the accumulation of vibration amplitude per unit time, and the energy consumption function represents the accumulation of power consumption per unit time; Based on the defined optimization objectives, a multi-objective optimization model is established using a non-dominated sorting genetic algorithm. The parameters of the adaptive control law are encoded as chromosomes. The parameters include the process noise covariance matrix elements of the Kalman filter, the observation noise covariance matrix elements, the learning rate in the adaptive law, and the sliding window length. Solving the multi-objective optimization model, introducing a proxy model based on radial basis functions in the solving process to accelerate the optimization process, and using adaptive gridding technology to maintain population diversity; From the solution results of the multi-objective optimization model, a compromise solution is selected as an output result based on a weighted sum index, wherein the weighted sum index is composed of a weighted combination of a vibration suppression effect function and an energy consumption function; and according to the selected compromise solution, a parameter of the adaptive control law is adjusted; The process of defining optimization objectives, establishing a multi-objective optimization model, solving the model, selecting a compromise solution, and adjusting the parameters of the adaptive control law is repeated to form a closed-loop feedback control.

8. A motor stator vibration suppression system based on electromagnetic pads and deep learning, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect the vibration signal of the motor stator, wherein the vibration signal includes vibration displacement, vibration velocity and vibration acceleration; and pre-process the vibration signal, including filtering, noise reduction and normalization; A second unit is used to input the preprocessed signal into a deep learning model, wherein the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a temporal convolutional network; Train and optimize the deep learning model; fuse features through the fully connected layer to obtain a comprehensive feature vector; use The comprehensive feature vector predicts the future vibration state of the motor stator; The third unit is used to establish an electromagnetic force model, design an adaptive control law based on the predicted future vibration state and the electromagnetic force model; use the adaptive control law to calculate the electromagnetic force required to suppress vibration, and convert it into a control signal to adjust the electromagnetic pad; monitor the vibration suppression effect in real time and obtain actual vibration data; compare the actual vibration data with the predicted future vibration state to obtain error information; Based on the error information, an online learning algorithm is used to perform real-time fine-tuning on the deep learning model. A multi-objective optimization algorithm is used to simultaneously consider the vibration suppression effect and energy consumption, adjust the adaptive control law, and form a closed-loop feedback control to continuously optimize the vibration suppression effect.

9. 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 described in any one of claims 1 to 7.

10. 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 7 is implemented.

Images