An AI-based method and system for detecting motor anomalies

Abstract

The present invention relates to the technical field of motor detection, and specifically provides a method and system for motor anomaly detection based on artificial intelligence. The method collects signal data of motor operation through sensors to establish a data set; obtains preliminary modal components of the preprocessed data set through variational mode extraction; extracts residual features of each preliminary modal component through a deep residual network to obtain enhanced modal components; performs sparse shrinkage on the enhanced modal components and substitutes them back into the objective function to output an iteratively updated set of modal components; constructs a motor anomaly detection model, and trains the motor anomaly detection model based on a training set; detects and classifies different types of attacks on the operation of motors in an industrial control system through the motor anomaly detection model. By combining variational mode extraction and a deep residual network, it can more accurately extract the subtle changes in the motor operation state, improve the accuracy of the motor anomaly detection model, and achieve the classification of different types of attacks.

Description

Technical Field

[0001] The present invention relates to the technical field of motor detection, and particularly to an abnormal motor detection method and system based on artificial intelligence. Background Art

[0002] With the rapid development of technologies such as industrial Internet and Internet of Things, the network security threats faced by motor devices in industrial control systems (ICS) are increasing day by day. As a core device in industrial production, the operating state of the motor is directly related to the security of the industrial system and production efficiency. Since a large amount of sensor data is generated during the operation of the motor, the access methods and sources of these data are diverse, which increases the complexity of data processing. At the same time, abnormal motor detection needs to find a balance between real-time performance and accuracy, and traditional detection methods relying on static rules and features are difficult to cope with complex and changeable abnormal patterns.

[0003] In order to improve the accuracy and efficiency of abnormal motor detection, some researchers explore new technical paths, mainly including data fusion technology, deep learning technology, and abnormal detection algorithms. Data fusion technology can effectively fuse time-domain and frequency-domain data, and improve the accuracy of abnormal detection through the complementarity of multi-source data. For example, through time-frequency joint analysis, synchronous analysis of the vibration signal and current signal of the motor can more comprehensively reflect the operating state of the motor. Deep learning technology has advantages in processing complex non-linear data. By constructing a deep neural network model, feature extraction and pattern recognition can be performed on the operating data of the motor. For example, convolutional neural network (CNN) and recurrent neural network (RNN) are used to analyze motor data to achieve efficient detection of unknown abnormal patterns. Unsupervised learning methods based on machine learning, such as Isolation Forest and Principal Component Analysis (PCA), can effectively discover abnormal points in motor data. At the same time, ensemble learning methods, such as Random Forest and Boosting Trees, can also improve the detection performance by combining multiple weak classifiers.

[0004] However, in practical applications, there are still many deficiencies in traditional security protection measures relying on vulnerability databases, and innovative technical means are needed to improve the security protection capabilities of motors and their control systems. For example, developing an intrusion detection system (IDS) based on deep learning can monitor and analyze the operating data of motors in real time, and promptly discover and prevent abnormal behaviors and unknown attacks. However, with the continuous evolution of network threats and the continuous intelligence of motor devices, abnormal motor detection technology needs to be continuously developed and innovated. Summary of the Invention

[0005] To solve the above problems, on the one hand, the present invention provides a method for detecting motor anomalies based on artificial intelligence, including:

[0006] S1. Collect signal data of motor operation through sensors to establish a data set, and preprocess the data set;

[0007] S2. Obtain the preliminary modal components of the preprocessed data set through variational mode decomposition;

[0008] S2.1. Initialize the number of modal components for variational mode decomposition K , and initialize the center frequency parameters of each modal component in a linearly averaged distribution manner ω k and the penalty coefficient α k ;

[0009] S2.2. Construct the objective function for variational mode decomposition with quadratic penalty terms and bandwidth constraints;

[0010] S2.3. Based on the objective function, use the alternating direction method of multipliers to iteratively optimize the data set to obtain the preliminary modal components ;

[0011] S3. Extract the residual features of each preliminary modal component through a deep residual network, and superimpose the residual features on the preliminary modal components to obtain enhanced modal components ;

[0012] S4. Perform sparse shrinkage on the enhanced modal components , substitute the sparsely shrunk modal components back into the objective function of S2, and iteratively update the modal components and the center frequencies of the modal components through adaptive gradient descent until the objective function converges to a preset threshold, and output the iteratively updated set of modal components as the final decomposition result of variational mode decomposition;

[0013] S5. Construct a motor anomaly detection model, and construct a focal loss function as the loss function of the motor anomaly detection model; divide the decomposition result of the data set obtained by variational mode decomposition into a training set and a test set, train the motor anomaly detection model based on the training set, and perform training evaluation through the test set to obtain a trained motor anomaly detection model;

[0014] S6. Detect and classify different types of attacks on the operation of motors in industrial control systems through the motor anomaly detection model.

[0015] As a specific implementation, the construction of the motor anomaly detection model in S5 is to construct a motor anomaly detection model through an LSTM artificial neural network, and obtain an optimized improved motor anomaly detection model through particle swarm optimization.

[0016] Among them, the motor anomaly detection model includes a three-layer LSTM artificial neural network, and the number of nodes in each layer of LSTM artificial neural network decreases layer by layer, namely 128, 64, and 32 respectively; the channel attention mechanism is introduced after the third layer of LSTM artificial neural network, and its output layer is connected to the fully connected network.

[0017] The improved motor anomaly detection model optimized by particle swarm optimization is specifically:

[0018] a. Initialize the particle swarm size and set the maximum number of iterations; initialize and determine the initial velocity and position of the particle. A single particle represents an LSTM hyperparameter.

[0019] b. Train the motor anomaly detection model using the training set and calculate the loss value using the fitness function; based on the loss value, update the individual and global optimal values of the particles, and update the particle speed and position;

[0020] The updating formula of the particle velocity and position is:

[0021] ,

[0022] ,

[0023] in: y i represents the model parameters, v i (t) is the amplitude of control parameter adjustment, w represents the inertia weight, p i represents the individual optimal position of particle i, g represents the global optimal position, t Indicates the current iteration number, c 1 , c 2 are learning factors; r 1 , r 2 is a random number between [0,1];

[0024] c. If the number of iterations reaches the maximum number of iterations or the loss value is less than the adaptation threshold, the optimal particle is determined, and the hyperparameters of the optimal particle are selected as the parameters of the motor anomaly detection model to obtain the optimized improved motor anomaly detection model.

[0025] As a specific implementation, the specific operations of the preprocessing are:

[0026] Step 1: Based on the dataset, determine the minority-class samples and majority-class samples. Use the Synthetic Minority Over-sampling Technique (SMOTE) to synthesize new samples among the neighbors of the minority-class samples and add the new samples to the dataset.

[0027] Step 2: Identify and remove the overlapping samples in the dataset through the duplicate sample removal technique.

[0028] Step 3: Normalize all the signal data in the dataset after removing the overlapping samples, and map the signal data into the interval [0, 1].

[0029] The objective function of the variational mode extraction is:

[0030] ,

[0031] Among them, u represents the extracted modal component, ω represents the central frequency of the modal component, is the time differential operator, α is the regularization parameter, j represents the imaginary unit.

[0032] In S3, extracting the residual features of each preliminary modal component through the deep residual network and superimposing the residual features with the preliminary modal components specifically includes:

[0033] For each modal component u k (t), extract the residual features, and the calculation formula for extracting the residual features is:

[0034] ,

[0035] Among them, f(u k (t)) is the non-linear mapping function of the deep residual network, and R(u k (t)) represents the residual features extracted from the modal component;

[0036] Superimpose the extracted residual features with the preliminary modal components to generate enhanced modal components which is:

[0037] .

[0038] In order to dynamically adjust the attention weights of the model for different samples during training, the formula of the focal loss function in S5 is:

[0039] ,

[0040] Among them, is the predicted probability that the current sample is correctly classified, is the weight factor; γ is an adjustable parameter;

[0041] The decomposition results of the training set are fused by the focal loss function at multiple resolutions and then trained.

[0042] The method for detecting motor anomalies based on artificial intelligence further includes evaluating the reliability of the motor anomaly detection model through cross-validation when training the motor anomaly detection model. The specific method is as follows:

[0043] Construct a multi-class confusion matrix, and calculate the parameter indicators for evaluating the motor anomaly detection model according to the multi-class confusion matrix. The parameter indicators include Precision, Recall, and the harmonic mean.

[0044] On the other hand, a motor anomaly detection system based on artificial intelligence is provided, including:

[0045] A data acquisition module for collecting signal data of the motor operation through sensors to establish a data set and preprocessing the data set;

[0046] A modal extraction module for obtaining the preliminary modal components of the preprocessed data set through variational mode decomposition; initializing the number of modal components of the variational mode decomposition K and initializing the center frequency parameters of each modal component in a linearly evenly distributed manner ω k and the penalty coefficient α k ; constructing an objective function for variational mode decomposition with quadratic penalty terms and bandwidth constraints; based on the objective function, using the alternating direction method of multipliers to iteratively optimize the data set to obtain the preliminary modal components ;

[0047] The modal extraction module is also used to extract the residual features of each preliminary modal component through a deep residual network, and superimpose the residual features on the preliminary modal components to obtain enhanced modal components ;

[0048] The modal extraction module is also used to perform sparse shrinkage on the enhanced modal components and substitute the sparsely shrunk modal components back into the objective function of S2, and iteratively update the modal components and the center frequencies of the modal components through adaptive gradient descent until the objective function converges to a preset threshold, and output the iteratively updated set of modal components as the final decomposition result of the variational mode decomposition;

[0049] A model construction module for constructing a motor anomaly detection model and constructing a focal loss function as the loss function of the motor anomaly detection model; dividing the decomposition result extracted from the dataset by variational mode decomposition into a training set and a test set, training the motor anomaly detection model based on the training set, and performing training evaluation through the test set to obtain a trained motor anomaly detection model;

[0050] A model detection module for detecting and classifying different types of attacks on the operation of motors in an industrial control system through the motor anomaly detection model.

[0051] Beneficial effects:

[0052] (1) By processing the signal data of motor operation in the present invention, the purpose is to balance the dataset and prevent data imbalance.

[0053] (2) Through variational mode decomposition in the present invention, complex and non-stationary signals can be decomposed into modal components with physical meanings, significantly improving the robustness of the model to noise signals and ensuring stable detection performance in complex industrial environments; combining variational mode decomposition and deep residual network can more accurately extract small changes in the motor operation state, thereby improving the accuracy of fault detection.

[0054] (3) By dynamically extracting the residual features of modal components and further optimizing the signal features through sparse shrinkage in the present invention, the characteristics of abnormal signals are made more prominent, improving the detection ability of the model for minority-class fault signals.

[0055] (4) By using particle swarm optimization to optimize the hyperparameters of LSTM and dynamically adjusting the parameters to obtain optimal parameters in the present invention, the accuracy of the motor anomaly detection model can be improved, realizing the classification of different types of attacks in industry and improving the detection ability for minority-class faults. Description of the Drawings

[0056] Figure 1 It is a flowchart of a method for motor anomaly detection based on artificial intelligence. Detailed Embodiments

[0057] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the drawings.

[0058] See Figure 1 , this embodiment provides a method for motor anomaly detection based on artificial intelligence, including the following steps:

[0059] S1. Collect signal data of motor operation through sensors to establish a dataset and preprocess the dataset;

[0060] The three-phase winding current signal and the axial vibration acceleration signal data during the operation of the motor are respectively collected through a Hall current sensor and a MEMS triaxial vibration sensor, and a data set is established with the collected three-phase winding current signal and axial vibration signal data.

[0061] By preprocessing the data set, relevant data for training the motor anomaly detection model is obtained. The specific operations of the preprocessing are as follows:

[0062] Step 1, determine the minority class samples and majority class samples based on the data set, and use the Synthetic Minority Over-sampling Technique (SMOTE) to synthesize new samples among the neighbors of the minority class samples and add the new samples to the data set;

[0063] For any minority class sample x i , calculate its k neighboring samples; randomly select several samples from these k neighboring samples, and perform linear interpolation with the original minority class sample to obtain a new synthetic sample. The formula for the synthetic new sample is:

[0064] ,

[0065] where, x new represents the new sample, x i is the minority class sample, represents x i 's neighboring sample, δ ∈[0,1] is a random number;

[0066] Step 2, identify and remove the overlapping samples in the data set through the Tomek Links technique;

[0067] Obtain the Euclidean distance of different class samples, that is, set d ( x m ,y n ) represents the Euclidean distance, x m is the minority class sample, y n is the majority class sample; if there is no third sample z ] , such that d ( x m ,z ) < d ( x m ,yn ) or d( y n ,z ) < d ( x m , y n ), then determine ( x m ,y n ) pair as Tomek Links pair and remove the Tomek Links pair.

[0068] Step 3, perform normalization processing on all signal data in the dataset after removing overlapping samples, and map the signal data into the interval [0, 1];

[0069] The formula for the normalization processing is:

[0070] ,

[0071] where, x represents the original signal data, x min represents the minimum value of the signal data, x max represents the maximum value of the signal data, X represents the normalized data.

[0072] S2. Obtain the preliminary modal components of the preprocessed dataset through variational mode extraction;

[0073] S2.1. Initialize the number of modal components of variational mode extraction (Variational Mode Extraction, VME) K , and initialize the center frequency parameter of each modal component in a linearly averaged distribution manner ω k and the penalty coefficient α k ;

[0074] Initialize variational mode extraction according to the spectral characteristics of the signal data, and determine the significant frequency components in the signal data through spectral analysis; if the signal characteristics are unknown, use a uniform distribution for initialization and set the initial frequencies uniformly within the frequency range.

[0075] Set the number of modal components according to the complexity of the signal data and the analysis requirements K , the number of modal components K is strongly correlated with the bearing fault characteristic frequency band (2 kHz - 8 kHz → K = 5).

[0076] Center frequency parameter ω k Initialization, with a 1500 Hz interval covering the typical motor fault frequency band; the penalty coefficient is used to control the sparsity of the modal components. A larger α k value will enhance the sparsity, while a smaller value allows the modal components to contain more details. Set the penalty coefficient α k to the empirical value of signal processing α k = 3000.

[0077] S2.2. Construct the objective function of variational mode decomposition with quadratic penalty term and bandwidth constraint;

[0078] The objective function of the variational mode decomposition is:

[0079] ,

[0080] where u represents the extracted modal component, ω represents the center frequency of the modal component, is the time differential operator, α is the regularization parameter, j represents the imaginary unit.

[0081] S2.3. Based on the objective function, use the alternating direction method of multipliers to iteratively optimize the dataset to obtain the preliminary modal components ;

[0082] Introduce the constraint condition into the objective function to get L ( u , ω , y ), define the multiplier vector y 0 and the regularization parameter α > 0 , and initialize the modal component u 0 and the center frequency of the modal component ω 0;

[0083] For k = 0, 1, 2,..., repeat the following steps:

[0084] Update u k+1, u k+1 = argmin u L ( u , ω k , y k );

[0085] Update ω k+1, ω k+1 = argmin ω L( u k+1 , ω , y k );

[0086] Update y k+1, y k+1 = y k +α ( Au k+1 +Bω k+1 -c ),

[0087] where A, B, and c represent constant matrices respectively.

[0088] When the function value of the objective function converges to a preset threshold, the preliminary modal components are output .

[0089] S3. Extract the residual features of each preliminary modal component through a deep residual network, and superimpose the residual features with the preliminary modal components to obtain enhanced modal components ;

[0090] The preliminary modal component u k (t) is the result obtained by preliminary decomposition through VME, and each modal component corresponds to the characteristics of different frequency bands in the signal.

[0091] Extract the residual features for each modal component u k (t). The calculation formula for extracting the residual features is:

[0092] ,

[0093] where f(u k (t)) is the non-linear mapping function of the deep residual network, and R(u k (t)) represents the residual features extracted from the modal component;

[0094] Superimpose the extracted residual features with the preliminary modal components to generate enhanced modal components as:

[0095] .

[0096] The Residual Mapping Shrinkage Mechanism (RMSM) extracts high-frequency and non-stationary characteristics, enabling the enhanced modal components to better reflect the subtle changes in the motor operating state and providing more accurate input features for subsequent fault detection.

[0097] S4. Perform sparse shrinkage on the enhanced modal components Substitute the sparsely shrunk modal components back into the objective function of S2, and iteratively update the modal components and the central frequencies of the modal components through adaptive gradient descent until the objective function converges to a preset threshold, and output the iteratively updated set of modal components as the final decomposition result of variational mode extraction;

[0098] Perform sparse shrinkage on the enhanced modal components The sparse shrinkage formula is:

[0099]

[0100] where represents the k-th enhanced modal component, represents the sign of the returned modal component; λ k represents the sparse shrinkage threshold.

[0101] Output the shrunk modal components:

[0102] 。

[0103] The sparse shrinkage function is a regularization term added to the objective function, which encourages sparse solutions by penalizing non-zero coefficients. It extracts the most important features from complex data while removing redundant information that has less impact on the results. In signal processing, modal components usually represent different components or features of a signal. The sparse shrinkage function weakens or removes those unimportant modal components by penalizing smaller coefficients, thus enhancing those key modal components. In high-dimensional data, there is often a large amount of redundant information. The sparse shrinkage function effectively removes this redundant information by introducing sparsity constraints, making most elements in the solution vector zero or close to zero. While removing redundant information, the sparse shrinkage function also retains those key features that have the greatest impact on the results. These features can usually reflect the essence and core information of the data and are of great significance for subsequent analysis and decision-making.

[0104] Substitute the sparsely shrunk modal components back into the objective function:

[0105]

[0106] Among them, represents the mode component after sparse shrinkage of the input, ω represents the center frequency of the mode component, ∂t() is the time differential operator, α is the regularization parameter, j represents the imaginary unit.

[0107] By adopting the adaptive gradient descent with Nesterov momentum (NAG-Adam) optimization method to iteratively update the mode component and the center frequency parameter, given the initial learning rate of NAG-Adam ( is the sampling frequency, unit kHz); the momentum factor μ = 0.95, iteratively update the mode component after sparse shrinkage until the objective function converges to the preset threshold, and output the iteratively optimized mode component. The finally optimized set of mode components is used as the final decomposition result of VME.

[0108] S5. Build a motor anomaly detection model, and build a focal loss function as the loss function of the motor anomaly detection model; divide the decomposition result extracted by variational mode from the data set into a training set and a test set, train the motor anomaly detection model based on the training set, and conduct training evaluation through the test set to obtain a trained motor anomaly detection model;

[0109] The building of the motor anomaly detection model specifically includes: building a motor anomaly detection model through an LSTM artificial neural network, and obtaining an optimized improved motor anomaly detection model through particle swarm optimization;

[0110] The motor anomaly detection model includes three layers of LSTM, and the number of nodes in each layer of LSTM decreases layer by layer, which are 128, 64, and 32 respectively; a channel attention mechanism is introduced after the third layer of LSTM in the model, and its output layer is connected to a fully connected network. Decreasing the number of nodes layer by layer helps to reduce the model complexity and enhance the generalization ability of the model; the third LSTM learns the importance of each channel (feature dimension), dynamically adjusts the contribution of different channels to the final decision, and the fully connected layer adopts a feature recalibration technology to perform Hadamard product operation on the attention weight and the original feature to realize the optimization of the non-linear mapping of the feature space, helping the LSTM artificial neural network to focus on the key features in the motor data set and improving the detection performance of the model.

[0111] Adopt particle swarm optimization (PSO) to globally optimize the initial weights and learning rate parameters of the built motor anomaly detection model, and dynamically adjust the parameters of the motor anomaly detection model through an evolutionary algorithm. The specific operation is as follows:

[0112] a. Initialize the particle swarm size and set the maximum number of iterations; initialize and determine the initial velocity and position of the particles, where a single particle represents an LSTM hyperparameter.

[0113] b. Train the motor anomaly detection model using the training set and calculate the loss value through the fitness function; update the individual optimal and global optimal of the particles based on the loss value, and update the particle velocity and position.

[0114] The smaller the value of the loss value, the better the training effect of the model. Update the individual optimal and global optimal of the particles, and update the particle velocity and position. The update formulas for the particle velocity and position are as follows:

[0115] ,

[0116] ,

[0117] Where: y i Represents the model parameter. v i (t) is the amplitude of the control parameter adjustment. w Represents the inertia weight (dynamically adjusted to ensure a balance between global search and local search). p i Represents the individual optimal position of particle i. g Represents the global optimal position. t Represents the current iteration number. c 1 , c 2 Are the learning factors respectively, controlling the degree to which the particles move towards the individual optimal and global optimal. r 1 , r 2 Is a random number between [0, 1], increasing the search randomness.

[0118] Meanwhile, the inertia weight is dynamically adjusted:

[0119] .

[0120] The inertia weight is initially set to a relatively large value (such as ), increasing the global search ability; gradually decreasing in the later stage (such as ), thereby improving the local convergence ability. T is the total number of iterations, used to control the rate of weight adjustment, and t is the current iteration number.

[0121] c. If the number of iterations reaches the maximum number of iterations or the loss value is less than the adaptation threshold, determine the optimal particle, select the hyperparameters of the optimal particle as the parameters of the motor anomaly detection model, and obtain the optimized improved motor anomaly detection model;

[0122] After the motor anomaly detection model is optimized by the particle swarm optimization, by constructing a focal loss function (Focal Loss, abbreviated as FL) as the loss function of the improved motor anomaly detection model, which is used to dynamically adjust the attention weights of the model to different samples during training. The formula of the focal loss function is:

[0123] ,

[0124] where, is the predicted probability that the current sample is correctly classified, is the weight factor; γ is an adjustable parameter;

[0125] By adjusting the parameter γ, the attention degree of the focal loss function to easy-to-classify and difficult-to-classify samples can be controlled.

[0126] S6. Detect and classify different types of attacks on the operation of the motor in the industrial control system through the motor anomaly detection model.

[0127] During the training process of the motor anomaly detection model in S5 of this embodiment, it also includes evaluating the reliability of the motor anomaly detection model through cross-validation. The specific operation is as follows:

[0128] Construct a multi-class confusion matrix, and calculate the parameter indicators for evaluating the motor anomaly detection model respectively according to the multi-class confusion matrix. The parameter indicators include Precision accuracy, Recall recall rate, and harmonic mean.

[0129] Among them, Precision accuracy represents the proportion of samples that are actually in class among all samples that the model thinks are in class ;

[0130] .

[0131] Recall recall rate indicates how many samples that are actually in class are predicted to be in class ;

[0132] .

[0133] F1-score is the harmonic mean of the above Precision accuracy and Recall recall rate;

[0134] 。

[0135] In addition, the present invention also provides an artificial intelligence-based motor anomaly detection system, including:

[0136] A data acquisition module, configured to collect signal data of the motor operation through sensors to establish a data set, and preprocess the data set;

[0137] A modal extraction module, configured to obtain preliminary modal components of the preprocessed data set through variational mode extraction; initialize the number of modal components of variational mode extraction K , and initialize the center frequency parameters of each modal component in a linearly evenly distributed manner ω k and the penalty coefficient α k ; construct an objective function of variational mode extraction containing a quadratic penalty term and bandwidth constraint; based on the objective function, use the alternating direction multiplier method to iteratively optimize the data set to obtain preliminary modal components ;

[0138] The modal extraction module is further configured to extract residual features of each preliminary modal component through a deep residual network, and superimpose the residual features on the preliminary modal components to obtain enhanced modal components ;

[0139] The modal extraction module is further configured to perform sparse shrinkage on the enhanced modal components , re-substitute the sparsely shrunk modal components into the objective function of S2, and iteratively update the modal components and the center frequencies of the modal components through adaptive gradient descent until the objective function converges to a preset threshold, and output the iteratively updated modal component set as the final decomposition result of variational mode extraction;

[0140] A model construction module, configured to construct a motor anomaly detection model, and construct a focal loss function as the loss function of the motor anomaly detection model; divide the decomposition result of the data set obtained by variational mode extraction into a training set and a test set, train the motor anomaly detection model based on the training set, and perform training evaluation through the test set to obtain a trained motor anomaly detection model;

[0141] A model detection module, configured to detect and classify different types of attacks on the operation of the motor in the industrial control system through the motor anomaly detection model.

[0142] To verify the feasibility of the present invention, a simulation experiment was conducted. The laboratory simulation motor dataset was used, with the sensor model: three PCB 352C33 unidirectional acceleration sensors, the sensor sampling frequency: 12.8 Khz, the load: 1.5 kW, the rotational speed: 2600 rpm, and the bearing model: 6203 type. Among them, 3 channels collected vibration signals and 3 channels collected voltage signals. The fault types included five types: inner race fault of the bearing, outer race fault of the bearing, inter-turn short circuit, eccentricity, and broken bar. In the simulation example, the population size was set to 30 for all, and the number of iterations was set to 300 times. To verify the effectiveness of the proposed algorithm, the present invention compared the proposed method with other feature selection and classification methods.

[0143] To reflect the influence of the particle swarm optimization on LSTM, in this paper, the traditional LSTM and the LSTM improved by PSO were used as comparison models. Using the same dataset, the detection results are shown in Table 1. It can be seen from the table that the motor anomaly detection model proposed by the present invention performs better.

[0144] Table 1 Experimental results of LSTM and PSO-LSTM

[0145]

[0146] To verify the effectiveness of the VME after sparse shrinkage used in the present invention for data signal processing, it was compared with the ordinary standard VME. Table 2 shows the performance on the motor anomaly dataset compared with the ordinary standard VME, including the precision rate, false alarm rate, and missed alarm rate. It can be seen from Table 2 that the VME-RMSM proposed by the present invention performs better than the VME and has higher performance.

[0147] Table 2 Experimental results of the model using different dimensionality reduction methods based on VME-RMSM-LSTM

[0148]

[0149] Certainly, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions, or substitutions made by those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based motor anomaly detection method, characterized in that, Including: S1. Collect signal data of motor operation through sensors to establish a data set, and preprocess the data set. S2. Extract the initial modal components of the preprocessed data set through variational mode decomposition. S2.

1. Initialize the number of modal components extracted by the variational mode K , and initialize the center frequency parameters of each modal component in a linearly averaged distribution ω k and the penalty coefficient α k ; S2.

2. Construct an objective function for variational mode decomposition with quadratic penalty terms and bandwidth constraints. S2.

3. Iteratively optimize the dataset using the alternating direction method of multipliers based on the objective function to obtain the preliminary modal components ; S3. Extract the residual features of each preliminary modal component through a deep residual network, and superimpose the residual features on the preliminary modal components to obtain enhanced modal components. ; For each modal component u k (t) extract residual features, and the calculation formula for extracting the residual features is: , Among them, is the non-linear mapping function of the deep residual network, and R(u k (t)) represents the residual feature extracted by the modal component; S4. Perform sparse shrinkage on the enhanced modal component Substitute the modal component after sparse shrinkage back into the objective function of S2, and iteratively update the modal component and the center frequency of the modal component through adaptive gradient descent until the objective function converges to a preset threshold, and output the set of iteratively updated modal components as the final decomposition result of variational mode extraction; The enhanced modal component is subjected to sparse shrinkage, and the sparse shrinkage formula is as follows: ， Among them, represents the k-th enhanced modal component, represents the sign of the returned modal component; λ k represents the sparse shrinkage threshold; Output the shrunk modal components: ； S5. Construct a motor anomaly detection model, and construct a focal loss function as the loss function of the motor anomaly detection model; divide the decomposition result of the data set by variational mode decomposition into a training set and a test set, train the motor anomaly detection model based on the training set, and conduct training evaluation through the test set to obtain a trained motor anomaly detection model. S6. Detect and classify different types of attacks on the operation of motors in industrial control systems through the motor anomaly detection model.

2. The method for detecting motor anomalies based on artificial intelligence according to claim 1, wherein, The construction of the motor anomaly detection model in S5 is to construct a motor anomaly detection model through an LSTM artificial neural network, and obtain an optimized improved motor anomaly detection model through particle swarm optimization.

3. The method for detecting motor anomalies based on artificial intelligence according to claim 2, wherein The motor anomaly detection model includes three layers of LSTM artificial neural networks, and the number of nodes in each layer of the LSTM artificial neural network decreases layer by layer, which are 128, 64, and 32 respectively; a channel attention mechanism is introduced after the third layer of the LSTM artificial neural network, and its output layer is connected to a fully connected network.

4. The method for detecting motor anomalies based on artificial intelligence according to claim 2, characterized in that, The specific method for obtaining the optimized improved motor anomaly detection model through particle swarm optimization is as follows: a. Initialize the particle swarm size and set the maximum number of iterations; initialize and determine the initial velocity and position of the particles, and a single particle represents an LSTM hyperparameter. b. Train the motor anomaly detection model through the training set, and calculate the loss value through the fitness function; update the individual optimal and global optimal of the particles based on the loss value, and update the particle velocity and position. The update formulas for the particle velocity and position are: , , Wherein: y i represents the model parameters, v i (t) is the amplitude of the control parameter adjustment, w represents the inertia weight, p i represents the individual optimal position of particle i, g represents the global optimal position, t represents the current iteration number, c 1 , c 2 are respectively the learning factors; r 1 , r 2 is a random number between [0, 1]; If the number of iterations reaches the maximum number of iterations or the loss value is less than the fitness threshold, determine the optimal particle, select the hyperparameters of the optimal particle as the parameters of the motor anomaly detection model, and obtain the optimized improved motor anomaly detection model.

5. The method for detecting motor anomalies based on artificial intelligence according to claim 1, wherein The specific operations of the preprocessing are as follows: Step 1. Based on the data set, determine the minority class samples and majority class samples, use the synthetic minority over-sampling technique to synthesize new samples in the neighbors of the minority class samples, and add the new samples to the data set. Step 2. Identify and remove the overlapping samples in the data set through the duplicate sample removal technique. Step 3. Normalize all the signal data in the data set after removing the overlapping samples, and map the signal data to the interval [0, 1].

6. The method for detecting motor anomalies based on artificial intelligence according to claim 1, wherein The objective function of the variational mode decomposition is: , Among them, u represents the extracted modal component, ω represents the center frequency of the modal component, is the time differential operator, α is the regularization parameter, j represents the imaginary unit.

7. The method for detecting motor anomalies based on artificial intelligence according to claim 1, wherein The specific method for extracting the residual features of each initial modal component through a deep residual network in S3 and superimposing the residual features with the initial modal components includes: Superimpose the extracted residual features on the preliminary modal components to generate enhanced modal components It is: 。 8. The method for detecting motor anomalies based on artificial intelligence according to claim 1, characterized in that, The formula of the focal loss function in S5 is: , Among them, is the predicted probability that the current sample is correctly classified, is the weight factor; γ is an adjustable parameter.

9. The method for detecting motor anomalies based on artificial intelligence according to claim 1, characterized in that, When training the motor anomaly detection model in S5, it also includes evaluating the reliability of the motor anomaly detection model through cross-validation. The specific method is: Construct a multi-class confusion matrix, and calculate the parameter indicators for evaluating the motor anomaly detection model based on the multi-class confusion matrix. The parameter indicators include Precision, Recall, and the harmonic mean.

10. An artificial intelligence-based motor anomaly detection system for implementing the artificial intelligence-based motor anomaly detection method described in claim 1, characterized in that, It includes: A data acquisition module for collecting signal data of motor operation through sensors to establish a data set and preprocessing the data set; A modal extraction module for obtaining the preliminary modal components of the preprocessed data set through variational mode extraction; Initialize the number of modal components extracted by variational mode decomposition K , and initialize the center frequency parameters of each modal component in a linearly averaged distribution ω k and the penalty coefficient α k ; construct the objective function of variational mode decomposition with quadratic penalty term and bandwidth constraint; based on the objective function, use the alternating direction method of multipliers to iteratively optimize the dataset to obtain the preliminary modal components ; The modal extraction module is further configured to extract the residual features of each preliminary modal component through a deep residual network, and superimpose the residual features on the preliminary modal components to obtain enhanced modal components ; The modal extraction module is further configured to perform sparse shrinkage on the enhanced modal components Substitute the modal components after sparse shrinkage back into the objective function of S2, and iteratively update the modal components and the center frequencies of the modal components through adaptive gradient descent until the objective function converges to a preset threshold, and output the iteratively updated set of modal components As the final decomposition result of variational mode extraction; A model construction module for constructing a motor anomaly detection model and constructing a focal loss function as the loss function of the motor anomaly detection model; Divide the decomposition result of the data set obtained by variational mode extraction into a training set and a test set, train the motor anomaly detection model based on the training set, and conduct training evaluation through the test set to obtain a trained motor anomaly detection model; A model detection module for detecting and classifying different types of attacks on the operation of motors in an industrial control system through the motor anomaly detection model.

Images