Abnormality detection method and system for intelligent motor

Abstract

The invention relates to the technical field of equipment anomaly detection, and discloses an anomaly detection method and system for an intelligent motor, and the method comprises the steps: collecting multi-source heterogeneous data when the intelligent motor works, carrying out the feature-level fusion, and carrying out the potential local learning, and obtaining optimized data features; performing data fragment segmentation on the optimized data features to obtain segmentation features, and determining a time-frequency feature spectrum of the segmentation features under multiple resolutions; mining a feature association relationship of the multi-dimensional operation features of the intelligent motor to construct a reinforcement learning strategy, and performing target feature screening on the multi-dimensional operation features of the intelligent motor to obtain a screened feature subset; and carrying out confrontation generation processing on the screening feature subset to obtain a simulation sample, carrying out feature ensemble learning processing on the simulation sample and the screening feature subset to obtain a judgment feature, and carrying out anomaly analysis on the intelligent motor to obtain an anomaly detection report. According to the invention, the detection precision of non-obvious or weak-feature anomalies in the intelligent motor can be improved.

Description

Technical Field

[0001] The present invention relates to a method and system for abnormal detection of an intelligent motor, belonging to the technical field of equipment abnormal detection. Background Art

[0002] With the progress of modern science and technology and the continuous development of production applications, intelligent motors play an increasingly important role in industrial, commercial, aviation, and military fields. The failure of an intelligent motor will not only damage the motor itself but also affect the normal operation of the entire system, and even endanger personal safety and cause huge losses. By real-time monitoring the parameters of an intelligent motor during operation for abnormal detection and analysis of the intelligent motor, potential faults such as mechanical wear, winding short circuit, and bearing failure can be identified in advance, thereby reducing production stoppage losses, lowering maintenance costs, and improving energy utilization efficiency.

[0003] Currently, the abnormal detection of intelligent motors mainly relies on single-modal signal analysis. First, the vibration signal or acoustic signal of the intelligent motor is collected, then the operation parameters of the intelligent motor are identified using time-domain analysis or frequency-domain analysis, and then compared with a preset threshold to determine abnormalities. However, for the single-modal signal analysis method, it mostly relies on artificial experience to set rules, lacks the fusion analysis and self-learning ability of multi-source heterogeneous data, and often fails to discover some abnormal problems with unobvious or weak features. Summary of the Invention

[0004] The present invention provides a method and system for abnormal detection of an intelligent motor, and its main purpose is to improve the detection accuracy for abnormalities with unobvious or weak features in the intelligent motor.

[0005] To achieve the above object, a method for abnormal detection of an intelligent motor provided by the present invention includes: Collecting multi-source heterogeneous data when the intelligent motor is working, performing feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, and performing potential Bent learning on the feature fusion data to obtain optimized data features; Querying the operation fluctuation state of the intelligent motor, segmenting the optimized data features based on the operation fluctuation state to obtain segmented features, and performing time-frequency joint analysis on the segmented features to determine the time-frequency feature maps of the segmented features at multiple resolutions; Using the time-frequency feature maps to mine the feature correlation relationships of the multi-dimensional operation features of the intelligent motor, constructing a reinforcement learning strategy for abnormal detection of the intelligent motor based on the feature correlation relationships, and using the reinforcement learning strategy to perform target feature screening on the multi-dimensional operation features of the intelligent motor to obtain a screened feature subset; Perform adversarial generation processing on the screened feature subset to obtain a simulated feature subset sample. Perform feature integration learning processing on the simulated feature subset sample and the screened feature subset to obtain judgment features. Use the judgment features to perform anomaly analysis on the intelligent motor to obtain an anomaly detection report.

[0006] Optionally, perform feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, including: Perform data modality classification on the multi-source heterogeneous data to obtain classification data; Extract the features of different types of data in the classification data to obtain multi-modal features; Perform standardization processing on the multi-modal features to obtain standardized data features; Perform feature fusion on the standardized data features to obtain feature fusion data.

[0007] Optionally, perform latent Bent learning on the feature fusion data to obtain optimized data features, including: Perform tensor reconstruction processing on the feature fusion data to obtain reconstructed data; Perform spatio-temporal attention encoding on the reconstructed data to obtain spatio-temporal fusion features; Perform feature compression on the spatio-temporal fusion features to obtain latent variable distribution parameters; Perform feature screening on the latent variable distribution parameters to obtain optimized data features.

[0008] Optionally, perform feature screening on the latent variable distribution parameters to obtain optimized data features, including: Calculate the mutual information value of each parameter feature in the latent variable distribution parameters using the following formula ; where represents the mutual information value, represents a parameter feature in the latent variable distribution parameters, represents an anomaly label in the latent variable distribution parameters, represents the probability that feature X takes the value x and anomaly label Y takes the value y, represents the probability that feature X takes the value x, represents the probability that anomaly label Y takes the value y; Based on the mutual information value, perform feature screening on the latent variable distribution parameters to obtain optimized data features.

[0009] Optionally, based on the running fluctuation state, perform data segment segmentation on the optimized data features to obtain segmented features, including: Performing parameter quantization processing on the operation fluctuation state to obtain an operation fluctuation parameter; Performing dynamic threshold modeling on the operation fluctuation parameter to obtain an adaptive fluctuation threshold; Using the adaptive fluctuation threshold, segmenting the optimized data features into data segments to obtain initial data segments; The boundary gradient of the initial data segment is detected, and the initial data segment is dynamically optimized using the boundary gradient to obtain a segmentation feature.

[0010] Optionally, performing a time-frequency joint analysis on the segmentation feature to determine a time-frequency feature map of the segmentation feature at multiple resolutions includes: Performing short-time Fourier transform processing on the segmentation features to obtain a low-frequency time-frequency matrix; Performing continuous wavelet transform on the low-frequency time-frequency matrix to obtain a high-frequency time-frequency matrix; Performing multi-resolution splicing on the low-frequency time-frequency matrix and the high-frequency time-frequency matrix to obtain a composite time-frequency spectrum; Adaptive threshold noise reduction is performed on the composite time-frequency spectrum to obtain a time-frequency feature spectrum.

[0011] Optionally, the time-frequency feature map is used to mine feature correlations of multi-dimensional operating features of the intelligent motor, including: Using the time-frequency feature map, construct a feature correlation map of the smart motor; Performing graph neural convolution processing on the feature association graph to obtain a local feature association relationship of the intelligent motor; Based on the local feature association relationship, the attention mechanism is used to perform global feature mining on the smart motor to obtain a global feature association relationship; According to the global feature association relationship, the feature association relationship of the multi-dimensional operation feature of the intelligent motor is identified.

[0012] Optionally, based on the feature association relationship, a reinforcement learning strategy for intelligent motor anomaly detection is constructed, including: Using the characteristic association relationship, constructing the state space of the intelligent motor to obtain a state vector; Performing action discretization processing on the state vector to obtain a discrete action set; Construct a multi-objective reward function for the discrete action set, wherein the multi-objective reward function can be expressed by the following formula: ; Where F represents the multi-objective reward function, represents the anomaly detection accuracy reward for detecting decision actions in discrete action sets, represents the decision response time reward for detecting decision actions in a discrete action set, represents the misjudgment loss reward for detecting decision actions in discrete action sets, express The weight of express The weight of express The weight of Based on the discrete action set and the multi-objective reward function, a reinforcement learning strategy for intelligent motor anomaly detection is constructed.

[0013] Optionally, performing feature ensemble learning processing on the simulated feature subset samples and the screened feature subset to obtain judgment features includes: Performing data fusion and alignment on the simulated feature subset samples and the screened feature subset to obtain a mixed data set; Using the trained heterogeneous base classifier, anomaly decision analysis is performed on the mixed data set to obtain a preliminary anomaly judgment result; Performing dynamic weighted integration processing on the preliminary abnormality judgment results to obtain an integrated decision result; Performing adversarial optimization processing on the integrated decision result to obtain an optimized decision result; The optimization decision result is deployed in a lightweight manner at the edge to obtain a judgment feature.

[0014] In order to solve the above problems, the present invention further provides an abnormality detection system for an intelligent motor, the system comprising: The device feature processing module is used to collect multi-source heterogeneous data when the intelligent motor is working, perform feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, and perform latent feature learning on the feature fusion data to obtain optimized data features; an equipment feature analysis module, configured to query the operating fluctuation state of the intelligent motor, segment the optimized data features based on the operating fluctuation state to obtain segmentation features, and perform a time-frequency joint analysis on the segmentation features to determine a time-frequency feature spectrum of the segmentation features at multiple resolutions; a device feature enhancement module, configured to utilize the time-frequency feature map to mine feature correlations of the multi-dimensional operating features of the intelligent motor, construct a reinforcement learning strategy for intelligent motor anomaly detection based on the feature correlations, and utilize the reinforcement learning strategy to perform target feature screening on the multi-dimensional operating features of the intelligent motor to obtain a screening feature subset; The device anomaly detection module is used to perform adversarial generation processing on the screened feature subset to obtain a simulated feature subset sample, perform feature integration learning processing on the simulated feature subset sample and the screened feature subset to obtain judgment features, and use the judgment features to perform anomaly analysis on the intelligent motor to obtain an anomaly detection report.

[0015] Compared with the problems described in the background art, in the embodiments of the present invention, first, by deploying multi-modal sensors such as sound, vibration, and current, the original data during the operation of the motor is collected (such as vibration time-domain waveforms, current frequency spectra, etc.). After modal classification, feature extraction (such as current harmonic amplitude, vibration root mean square value), and Z-score standardization, a deep neural network with an attention mechanism is used for feature fusion to form feature fusion data in a unified format, which can cover multi-dimensional data of machinery, acoustics, and electricity, avoiding missed detection caused by the lack of single-modal information. Then, the feature fusion data is reconstructed into a three-dimensional tensor, and then compressed into latent variable distribution parameters using a variational autoencoder. Finally, key features are screened based on mutual information, which can reduce the computational complexity and improve the efficiency of subsequent algorithms. Further, the present invention calculates the fluctuation parameters such as the variance and change rate of vibration / current data in real time, then establishes an adaptive threshold through kernel density estimation, and then uses a sliding window algorithm to segment the data segments to ensure the precise correspondence of vibration shocks and current mutations in time, improving the reliability of correlation analysis. For the segmented vibration / current signals, high-frequency details are extracted through short-time Fourier transform and continuous wavelet transform, and stitched to form a composite time-frequency map to realize the visualization of fault types. Further, the present invention maps the feature points of the time-frequency map to graph nodes, constructs a feature correlation graph, generates feature correlation rules, and then constructs a reinforcement learning environment containing a 100-dimensional state vector (fusing time-frequency features, correlation weights, real-time parameters) and 12 discrete actions (such as early warning, shutdown, adjusting the sampling frequency) based on the feature correlation relationship, designs a multi-objective reward function, and flexibly sets the weights to improve the detection accuracy. Even further, the present invention uses a generative adversarial network (GAN) to generate simulated anomaly samples, and after fusing them with real screened features, parallel detection is performed through a heterogeneous base classifier, and then through dynamic weighted integration, adversarial optimization, and knowledge distillation to compress the model. Finally, a 32-dimensional judgment feature vector is deployed on the edge device, which can make up for the lack of real fault data, and comprehensively judge by combining multiple features such as vibration frequency and current harmonics, improving the accuracy of anomaly judgment. Therefore, the present invention can improve the detection accuracy for non-obvious or weakly featured anomalies in intelligent motors. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic flow chart of an anomaly detection method for an intelligent motor provided by an embodiment of the present invention; Figure 2It is a module schematic diagram of an abnormal detection system for the intelligent motor provided by an embodiment of the present invention.

[0017] The implementation, functional features and advantages of the present invention will be further described with reference to the embodiments and the accompanying drawings. Specific embodiments

[0018] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] The embodiments of the present application provide an abnormal detection method for an intelligent motor. The execution subject of the abnormal detection method for the intelligent motor includes but is not limited to at least one of electronic devices such as a server, a terminal, etc. that can be configured to execute the method provided by the embodiments of the present application. In other words, the abnormal detection system for the intelligent motor can be executed by software or hardware installed on a terminal device or a server device. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc.

[0020] Embodiment 1: Refer to Figure 1 As shown, it is a flow schematic diagram of an abnormal detection method for an intelligent motor provided by an embodiment of the present invention. In this embodiment, the abnormal detection method for the intelligent motor includes: S1. Collect multi-source heterogeneous data during the operation of the intelligent motor, perform feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, and perform latent bent learning on the feature fusion data to obtain optimized data features.

[0021] In the embodiment of the present invention, by collecting the multi-source heterogeneous data during the operation of the intelligent motor, information such as vibration, sound, and current during the operation of the motor can be obtained to ensure a comprehensive and original information basis and provide data guarantee for motor abnormal detection.

[0022] Among them, the multi-source heterogeneous data refers to data with different sources, different types, and different structures during the operation of the intelligent motor, such as vibration data, temperature data, and current data.

[0023] Optionally, the multi-source heterogeneous data can be obtained by deploying multi-modal sensors in the working area of the intelligent motor, such as sound sensors, temperature sensors, etc.

[0024] Furthermore, in the embodiment of the present invention, by performing feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, the original data of different types can be converted into feature fusion data in a unified format, thereby retaining multi-dimensional complementary information and providing comprehensive and refined analysis data support for subsequent abnormal detection.

[0025] As an embodiment of the present invention, performing feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, including: Performing data modality classification on the multi-source heterogeneous data to obtain classification data; Extracting the features of different types of data in the classification data to obtain multi-modal features; Performing normalization processing on the multi-modal features to obtain normalized data features; Performing feature fusion on the normalized data features to obtain feature fusion data.

[0026] Among them, the multi-modal features refer to data features of different modalities, such as current data features, voltage data features, etc.

[0027] Optionally, the classification data can be obtained by classifying the collected multi-source heterogeneous data such as vibration, sound, and current into corresponding modal categories such as mechanical vibration data, acoustic data, and electrical data. The normalized data features can be obtained by performing normalization processing on the multi-modal features using the Z-score normalization method. The feature fusion data can be obtained by constructing a deep neural network containing a multi-layer perceptron, taking the normalized data features as input, and through internal feature transformation and interaction in the network, and combining the attention mechanism to automatically allocate modal weights for output.

[0028] Furthermore, in the embodiment of the present invention, by performing latent bent learning on the feature fusion data, the optimized data features can map the data from high dimension to low dimension, effectively extract the essential features of the data while reducing the data dimension, generate samples that conform to the original data distribution, and further optimize the analysis data.

[0029] As an embodiment of the present invention, performing latent bent learning on the feature fusion data to obtain optimized data features, including: Performing tensor reconstruction processing on the feature fusion data to obtain reconstructed data; Performing spatio-temporal attention encoding on the reconstructed data to obtain spatio-temporal fusion features; Performing feature compression on the spatio-temporal fusion features to obtain latent variable distribution parameters; Performing feature screening on the latent variable distribution parameters to obtain optimized data features.

[0030] Among them, the spatio-temporal fusion features refer to the features obtained by fusing the information of the time dimension and the space dimension, and weighting and integrating the key information of the data in the time series change and the spatial correlation characteristics. The latent variable distribution parameters refer to the parameters used to describe the probability distribution state of the latent variables in the probability model, usually including the mean, variance, etc., and are used to characterize the latent distribution characteristics of the data.

[0031] Optionally, the reconstructed data can be obtained by reconstructing the feature fusion data into a three-dimensional tensor based on the time series characteristics and modal dimensions of the vibration, sound and current data in the feature fusion data. The spatiotemporal fusion features can be generated by constructing a spatiotemporal attention mechanism network, calculating the attention weights on the time dimension and modal dimension respectively, focusing on the key information of the motor operation data in the time series changes and the correlation between modalities, and weighted fusion. The latent variable distribution parameters can be obtained by utilizing the encoder structure in the variational autoencoder, inputting the spatiotemporal fusion features into a multi-layer perceptron, and outputting the mean vector and variance vector of the latent variable through layer-by-layer dimensionality reduction compression.

[0032] Preferably, the feature screening of the latent variable distribution parameters to obtain optimized data features includes: The mutual information value of each parameter feature in the latent variable distribution parameter is calculated using the following formula: ; in, represents the mutual information value, Represents a parameter feature in the latent variable distribution parameter, represents an abnormal label in the latent variable distribution parameter, represents the probability that feature X takes the value x and abnormal label Y takes the value y, represents the probability that feature X takes the value x, Indicates the probability that the abnormal label Y takes the value y; Based on the mutual information value, feature screening is performed on the latent variable distribution parameters to obtain optimized data features.

[0033] The mutual information value refers to a quantitative indicator used to measure the degree of dependence and information sharing between two random variables.

[0034] In specific implementation, we can first set a mutual information threshold (such as 0.2), eliminate parameter features that are less than the threshold, and retain parameter features with high mutual information values and strong correlation with abnormal labels. Then, we form an optimized data feature subset. For example, the current fifth harmonic amplitude feature with a mutual information value of 0.3 is retained, and a temperature fluctuation feature with a mutual information value of 0.15 is discarded. Finally, we form an optimized data feature that focuses on key information and can improve the efficiency of anomaly detection.

[0035] S2. Query the operating fluctuation state of the intelligent motor, segment the optimized data features based on the operating fluctuation state to obtain segmentation features, and perform a time-frequency joint analysis on the segmentation features to determine a time-frequency feature map of the segmentation features at multiple resolutions.

[0036] In the embodiments of the present invention, by querying the operation fluctuation state of the intelligent motor, the dynamic change of the motor operation parameters can be obtained in real time, the operation characteristics of the motor under different working conditions can be accurately grasped, and the fluctuation trend of the motor operation state can be visually presented.

[0037] Among them, the operation fluctuation state refers to the dynamic change of operation parameters such as vibration, sound, and voltage of the intelligent motor during actual operation, as well as characteristics such as parameter change rate, fluctuation amplitude, and change trend, reflecting the working stability and performance of the motor under different working conditions, such as large changes in operation parameters within a short time and exceeding the normal fluctuation range.

[0038] Optionally, the operation fluctuation state can be determined by real-time monitoring of multi-source heterogeneous data such as vibration, sound, and current during the operation of the motor, calculating statistical indicators such as variance and change rate of the data, and comparing with the fluctuation range of historical normal operation data.

[0039] Furthermore, in the embodiments of the present invention, by segmenting the optimized data features based on the operation fluctuation state, the segmentation features can dynamically adjust the key data segments according to the actual operation change of the motor.

[0040] As an embodiment of the present invention, segmenting the optimized data features based on the operation fluctuation state to obtain segmentation features includes: Performing parameter quantization processing on the operation fluctuation state to obtain operation fluctuation parameters; Performing dynamic threshold modeling on the operation fluctuation parameters to obtain an adaptive fluctuation threshold; Using the adaptive fluctuation threshold to segment the optimized data features to obtain initial data segments; Detecting the boundary gradient of the initial data segments, and using the boundary gradient to dynamically optimize the initial data segments to obtain segmentation features.

[0041] Among them, the adaptive fluctuation threshold refers to a critical value dynamically adjusted according to the real-time change of data, used to judge whether the data fluctuation is abnormal, and the operation fluctuation parameter refers to an index obtained by quantifying characteristics such as the amplitude, frequency, and change rate of data fluctuation, used to describe the data fluctuation state.

[0042] Optionally, the operation fluctuation parameters can be obtained by extracting the time-domain statistical features (root mean square value, peak factor) and frequency-domain features (spectrum energy distribution) from the vibration, sound, and current signals of the intelligent motor, constructing a multi-dimensional feature vector by combining the parameter change rate and fluctuation amplitude, and then obtaining it after dimensionality reduction through principal component analysis. The adaptive fluctuation threshold can be based on the fluctuation parameters of the historical normal operation data of the intelligent motor, using kernel density estimation (KDE) to fit the probability distribution. Based on the fluctuation parameters of the historical normal operation data of the intelligent motor, using kernel density estimation (KDE) to fit the probability distribution, calculate the 95% confidence interval of the operation fluctuation parameters as the initial threshold, and then obtain it through dynamic updating of the threshold by the exponentially weighted moving average (EWMA) algorithm during real-time monitoring. The initial data segment can be obtained by designing a sliding window algorithm based on the fluctuation intensity. When the fluctuation parameter exceeds the threshold, the window size is reduced (e.g., from 100 ms to 50 ms), and when it is stable, the window is enlarged (e.g., extended to 200 ms). The window step size is dynamically adjusted according to the fluctuation frequency, and then the data segment is segmented using the set sliding window. The segmentation features can first calculate the gradient change rate of the fluctuation parameters between adjacent windows, mark the positions exceeding the set threshold as candidate boundaries, and then use the dynamic programming algorithm to optimize the boundary positions to minimize the variance within the segment and maximize the difference between segments. Finally, data synchronization and alignment processing are performed through multi-modal data timestamp matching and dynamic time warping (DTW) algorithm.

[0043] Furthermore, in the embodiment of the present invention, time-frequency joint analysis is performed on the segmentation features to determine that the time-frequency feature map of the segmentation features at multiple resolutions can clearly capture the characteristics of abnormal signals at different times and frequencies, and intuitively display the overall time-frequency characteristics of the motor operation state.

[0044] Among them, the time-frequency feature map refers to the visualization result obtained by performing time-frequency analysis on the dynamic time series segment of the intelligent motor by using the short-time Fourier transform combined with the wavelet transform.

[0045] As an embodiment of the present invention, performing time-frequency joint analysis on the segmentation features to determine the time-frequency feature map of the segmentation features at multiple resolutions includes: Performing short-time Fourier transform processing on the segmentation features to obtain a low-frequency time-frequency matrix; Performing continuous wavelet transform processing on the low-frequency time-frequency matrix to obtain a high-frequency time-frequency matrix; Performing multi-resolution splicing on the low-frequency time-frequency matrix and the high-frequency time-frequency matrix to obtain a composite time-frequency map; Performing adaptive threshold denoising on the composite time-frequency map to obtain a time-frequency feature map.

[0046] Among them, the low-frequency time-frequency matrix refers to the matrix data that presents the distribution of low-frequency signals in the time domain and frequency domain after the segmentation features are subjected to short-time Fourier transform using a fixed window function, which can reflect the changes in the low-frequency components of the signal over time. The high-frequency time-frequency matrix refers to the matrix formed after applying continuous wavelet transform to the low-frequency time-frequency matrix to capture the fine time-frequency characteristics of the high-frequency components of the signal, highlighting the frequency details of the high-frequency signal at different time points. The composite time-frequency spectrum refers to the alignment and splicing of the low-frequency time-frequency matrix and the high-frequency time-frequency matrix along the time axis, integrating the low-bandwidth time resolution and high-frequency high-frequency resolution characteristics, and comprehensively displaying the visual spectrum of the time-frequency characteristics of the signal at multiple resolutions.

[0047] In specific implementation, a fixed-size Hanning window can be used to perform a short-time Fourier transform on the vibration, sound, and current data in the segmented features. The Fourier transform converts the time-domain signal within each time window into a frequency-domain signal, obtaining the distribution of the low-frequency signal in the time and frequency domains and generating a low-frequency time-frequency matrix. The Morlet wavelet is selected as the mother wavelet, and a continuous wavelet transform is performed on the data in the low-frequency time-frequency matrix. The high temporal resolution of the wavelet transform in the high-frequency band is utilized to finely analyze the time-frequency characteristics of the high-frequency components of the signal and generate a high-frequency time-frequency matrix. The low-frequency and high-frequency time-frequency matrices are aligned along the time axis, and the wide temporal resolution and low-frequency resolution features of the low-frequency band are vertically spliced with the narrow temporal resolution and high-frequency resolution features of the high-frequency band to form a composite time-frequency spectrum containing both high- and low-frequency information. Based on the data distribution of the composite time-frequency spectrum, a local adaptive threshold algorithm (such as the SureShrink algorithm) is used to dynamically calculate the threshold according to the signal energy in different regions. The noise components in the spectrum are soft-thresholded to shrink, retaining the effective signal features and generating the final time-frequency feature spectrum.

[0048] S3. Utilize the time-frequency feature map to mine the feature correlation relationships of the multidimensional operating features of the intelligent motor. Based on the feature correlation relationships, construct a reinforcement learning strategy for intelligent motor anomaly detection. Utilize the reinforcement learning strategy to perform target feature screening on the multidimensional operating features of the intelligent motor to obtain a screening feature subset.

[0049] The embodiment of the present invention utilizes the time-frequency feature map to mine the feature correlation relationship of the multi-dimensional operating characteristics of the intelligent motor, thereby analyzing the intrinsic connection and potential pattern of multi-dimensional data such as motor vibration, current, temperature, etc. at the time-frequency level, intuitively presenting the interaction mechanism between each feature, and clearly showing the correlation characteristics of different factors in the motor operating state.

[0050] As an embodiment of the present invention, the time-frequency feature map is used to mine the feature correlation relationship of the multi-dimensional operation characteristics of the intelligent motor, including: Using the time-frequency feature map, construct a feature correlation map of the smart motor; Perform graph neural convolution processing on the feature association graph to obtain the local feature association relationship of the intelligent motor; Based on the local feature association relationship, use the attention mechanism to mine the global features of the intelligent motor and obtain the global feature association relationship; Identify the feature association relationship of the multi-dimensional operating features of the intelligent motor according to the global feature association relationship.

[0051] Among them, the feature association graph refers to a graph structure formed by taking data features as nodes and constructing edges based on the correlations, similarities, etc. between features, which is used to intuitively display the association relationships between features. The local feature association relationship refers to the association pattern and interaction relationship presented by data features within its neighboring area or the range of directly connected nodes. The global feature association relationship refers to the long-distance dependence and mutual association relationship spanning the entire data range, covering features in different regions and levels, reflecting the overall connection between features.

[0052] In the specific implementation process, each time-frequency feature point in the time-frequency feature map can be mapped to a graph node. Based on the spatio-temporal proximity and similarity of vibration, sound, and current signals in the time-frequency domain, the edge connections between nodes are constructed by calculating the cosine similarity or mutual information to form an initial feature association graph. Use a graph convolutional network (GCN) to perform convolutional operations on the feature association graph, aggregate the neighborhood information of nodes through the message passing mechanism, and extract the first-order neighborhood feature association pattern between nodes to obtain the local feature association relationship. Introduce a graph attention network (GAT) to calculate the attention weights between different nodes, adaptively focus on key nodes and important association edges, mine the long-distance dependence relationships across regions and modalities, and generate the global feature association relationship. Based on the node representation and edge weights in the global feature association relationship, screen out strong association relationships by setting thresholds, and then combine domain knowledge to convert the association pattern in the graph structure into interpretable multi-dimensional operating feature association rules to achieve the recognition of feature association relationships. For example, in the feature association graph, the edge weight between the node with a vibration frequency of 120Hz and the node with the amplitude of the 3rd harmonic of the current is 0.85, exceeding the set threshold of 0.7. Combining the knowledge of the motor field, it can be known that rotor imbalance will cause 120Hz vibration and at the same time increase the 3rd harmonic of the current. Therefore, the association rule of "abnormal vibration 120Hz frequency component and abnormal amplitude of the 3rd harmonic of the current, then the rotor is unbalanced" is obtained.

[0053] Furthermore, in the embodiment of the present invention, by constructing a reinforcement learning strategy for intelligent motor anomaly detection based on the feature association relationship, it is possible to dynamically optimize the feature selection decision with the goal of the accuracy and timeliness of motor anomaly detection through the interaction between the intelligent agent and the environment and the reward feedback mechanism.

[0054] Among them, the reinforcement learning policy refers to the mechanism by which an agent selects the optimal action based on the current state during the interaction with the environment.

[0055] As an embodiment of the present invention, based on the feature association relationship, a reinforcement learning policy for intelligent motor anomaly detection is constructed, including: Using the feature association relationship, construct the state space of the intelligent motor to obtain a state vector; Perform action discretization processing on the state vector to obtain a discrete action set; Construct a multi-objective reward function for the discrete action set, where the multi-objective reward function can be expressed by the following formula: ; Among them, F represents the multi-objective reward function, represents the anomaly detection accuracy reward of the detection decision action in the discrete action set, represents the decision response time reward of the detection decision action in the discrete action set, represents the misjudgment loss reward of the detection decision action in the discrete action set, represents the weight of represents the weight of represents the weight of Based on the discrete action set and the multi-objective reward function, construct a reinforcement learning policy for intelligent motor anomaly detection.

[0056] Among them, the state vector refers to a multi-dimensional numerical vector formed by quantifying and encoding various state information such as system operation and environmental conditions, and is used to comprehensively describe the state of the system at a certain moment. The discrete action set refers to a set of actions obtained by classifying the executable operations or behaviors and presented in a discrete form, and each action is clear, independent and selectable for execution. The multi-objective reward function refers to a mathematical function designed to evaluate the quality of behaviors and feedback reward values according to the weights of different objectives (such as efficiency, cost, benefit, etc.) based on the results generated by different behaviors.

[0057] Optionally, the state vector can perform feature engineering on the feature node states (such as the amplitude of vibration frequency components), edge weight relationships (such as the correlation strength between current harmonics and vibration characteristics), and real-time operating parameters of the motor (such as temperature and speed) in the time-frequency feature map. After dimensionality reduction through principal component analysis (PCA), it is mapped to a high-dimensional vector space to obtain a state vector containing 100 dimensions (which can be specifically set according to actual applications). The discrete action set can classify and encode anomaly detection decision actions (continuous monitoring, warning prompt, first-level shutdown, or second-level shutdown) and data processing actions (adjusting the sampling frequency, changing the time-frequency analysis algorithm, or switching the feature extraction method). Using a hierarchical Q-learning architecture, the upper layer processes detection decision actions (4), and the lower layer processes data processing actions (3) to form an action set containing 12 discrete actions (the specific processing actions and the number of action steps also need to be designed according to actual applications). The reinforcement learning strategy can adopt the deep deterministic policy gradient (DDPG) algorithm, combined with an experience replay mechanism and a target network. The agent is trained in a simulation environment containing 1000 historical fault cases, and then prioritized experience replay is used to improve the training efficiency. The target network is updated every 100 time steps, and finally a strategy is formed that can adaptively select the optimal detection and decision actions according to the real-time state of the motor.

[0058] It should be further noted that the multi-objective reward function is obtained by designing a weighted sum form of the reward function, which includes three sub-objectives: anomaly detection accuracy (weight 0.5), decision response time (weight 0.3), and misjudgment loss (weight 0.2). For example, a successful detection of a bearing fault is rewarded +15 points, a false alarm is punished -10 points, and an additional -5 points are deducted if the response time exceeds 3 seconds, thus realizing the quantitative evaluation and optimization of multi-objective rewards.

[0059] Furthermore, in the embodiment of the present invention, by using the reinforcement learning strategy, target feature screening is performed on the multi-dimensional operating characteristics of the intelligent motor, and the obtained screened feature subset can screen out the most discriminative feature subset, efficiently and accurately mining the key features closely related to the abnormal state of the motor.

[0060] Optionally, the screened feature subset can use the constructed reinforcement learning strategy, guided by the multi-objective reward function, to let the agent perform screening operations on the multi-dimensional operating characteristics by executing the discrete action set in the simulation environment. According to the state vector change and reward feedback, the screening strategy is iteratively optimized, and finally the feature subset that makes the anomaly detection performance optimal is determined.

[0061] S4. Perform adversarial generation processing on the screened feature subset to obtain a simulated feature subset sample. Perform feature integration learning processing on the simulated feature subset sample and the screened feature subset to obtain judgment features. Use the judgment features to perform anomaly analysis on the intelligent motor to obtain an anomaly detection report.

[0062] In the embodiment of the present invention, through the adversarial generation processing on the screened feature subset to obtain a simulated feature subset sample, diverse simulated abnormal feature samples can be generated, effectively expanding the amount of abnormal data samples, enriching data diversity, and alleviating the problem of lack of actual abnormal data.

[0063] Among them, the simulated feature subset sample refers to a sample that is simulated and generated through data generation technologies (such as generative adversarial network GAN, variational autoencoder VAE) and focuses on the key feature subset.

[0064] Optionally, the simulated feature subset sample can be obtained by using a generative adversarial network (GAN). The screened feature subset is input into the generator to generate simulated data. The discriminator distinguishes between real and simulated data. The parameters are optimized through adversarial training of the two, and finally the generator outputs a simulated sample that conforms to the real data distribution.

[0065] In the embodiment of the present invention, through the feature integration learning processing on the simulated feature subset sample and the screened feature subset to obtain judgment features, the original and augmented feature data can be integrated to improve the recognition accuracy and stability of complex abnormal situations of the intelligent motor.

[0066] As an embodiment of the present invention, performing feature integration learning processing on the simulated feature subset sample and the screened feature subset to obtain judgment features includes: Fuse and align the simulated feature subset sample and the screened feature subset to obtain a mixed data set; Use the trained heterogeneous base classifier to perform anomaly decision analysis on the mixed data set to obtain a preliminary anomaly judgment result; Perform dynamic weighted integration processing on the preliminary anomaly judgment result to obtain an integrated decision result; Perform adversarial optimization processing on the integrated decision result to obtain an optimized decision result; Perform edge lightweight deployment on the optimized decision result to obtain judgment features.

[0067] Among them, the preliminary anomaly judgment result refers to the original judgment results of each classifier on whether the sample is an anomaly and the specific anomaly type (such as bearing fault, rotor imbalance, etc.) after independent anomaly detection of the mixed data set. The integrated decision result refers to the comprehensive decision result after fusing the preliminary anomaly judgment results of multiple heterogeneous base classifiers through dynamic weighted integration technology.

[0068] In the specific implementation process, the dynamic time warping (DTW) algorithm can be used to align the time series data of the simulated feature subset samples and the screened feature subset. Then, the tensor decomposition technology is used to reconstruct multi-modal features such as vibration, current, and temperature into a third-order tensor (sample × time step × feature) with a unified dimension, forming a mixed data set with consistent dimensions and synchronized time series. For example, align the vibration characteristics of the simulated bearing fault with the current harmonic characteristics screened from the real data according to the time stamp; deploy three base classifiers of support vector machine (SVM), random forest (RF), and long short-term memory network (LSTM) that have been trained on historical fault data to perform parallel anomaly detection on the mixed data set respectively; SVM detects bearing faults based on vibration characteristics, RF analyzes current harmonics to identify rotor imbalance, and LSTM captures the time series changes of temperature to judge winding anomalies, and finally outputs the preliminary anomaly detection results including categories such as "normal", "bearing fault", and "rotor imbalance"; based on the performance (accuracy, recall rate, F1 value) of each base classifier on the validation set, design an adaptive weight adjustment mechanism to calculate the credibility weights of each classifier under the current data characteristics in real time; for example, when the vibration characteristics are dominant, increase the weight of LSTM to 0.6, and generate the integrated decision result through the weighted voting method; construct a generative adversarial network (GAN) architecture, the generator generates anomaly samples close to the real distribution (such as the characteristics of a slightly unbalanced motor rotor), the discriminator distinguishes the integrated decision result from the real label, and iteratively optimizes the integrated decision result through adversarial training to improve the detection accuracy of the model for rare fault modes; use knowledge distillation technology to compress the optimized decision result into a lightweight neural network, and finally deploy it to the edge computing device to generate a low-dimensional judgment feature vector with a dimension of 32.

[0069] In the embodiment of the present invention, by using the judgment features to perform anomaly analysis on the intelligent motor, the obtained anomaly detection report can accurately identify the abnormal state of the motor based on the comprehensively optimized key features, and clearly present information such as the anomaly type, location, and degree in an intuitive report form, providing a direct and reliable basis for motor maintenance and decision-making.

[0070] During implementation, judgment features are input into a lightweight convolutional neural network model deployed on an edge computing device. The model's pre-trained parameters are used to classify the features and compare them with the intelligent motor anomaly database to determine the anomaly type, probability, and severity. The model then outputs an anomaly detection report containing the fault location (e.g., bearing, rotor), probability of occurrence (e.g., 85%), fault level (Level 3 severe fault), and recommended actions. For example, if the vibration frequency anomaly eigenvalue in the input judgment feature vector reaches 0.8 and the current harmonic eigenvalue reaches 0.7, the model identifies a bearing fault with an 85% probability of occurrence, a Level 3 severe fault, and recommends immediate shutdown for maintenance.

[0071] Example 2: like Figure 2 FIG. 1 is a functional module diagram of an abnormality detection system for an intelligent motor according to the present invention.

[0072] The intelligent motor anomaly detection system 200 described in the present invention can be installed in an electronic device. Depending on the functionality implemented, the intelligent motor anomaly detection system may include a device feature processing module 201, a device feature analysis module 202, a device feature enhancement module 203, and a device anomaly detection module 204. A module, also referred to as a unit, is a series of computer program segments that can be executed by an electronic device processor and perform a fixed function. These are stored in the electronic device's memory.

[0073] In the embodiment of the present invention, the functions of each module / unit are as follows: The device feature processing module 201 is used to collect multi-source heterogeneous data when the intelligent motor is working, perform feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, and perform latent feature learning on the feature fusion data to obtain optimized data features; The device feature analysis module 202 is configured to query the operating fluctuation state of the intelligent motor, segment the optimized data features based on the operating fluctuation state to obtain segmentation features, and perform a time-frequency joint analysis on the segmentation features to determine a time-frequency feature spectrum of the segmentation features at multiple resolutions; The device feature enhancement module 203 is configured to utilize the time-frequency feature map to mine feature correlations of the multi-dimensional operating features of the intelligent motor, construct a reinforcement learning strategy for intelligent motor anomaly detection based on the feature correlations, and utilize the reinforcement learning strategy to perform target feature screening on the multi-dimensional operating features of the intelligent motor to obtain a screening feature subset; The device anomaly detection module 204 is configured to perform adversarial generation processing on the screened feature subset to obtain a simulated feature subset sample, perform feature integration learning processing on the simulated feature subset sample and the screened feature subset to obtain judgment features, and use the judgment features to perform anomaly analysis on the intelligent motor to obtain an anomaly detection report.

[0074] Specifically, each module in the anomaly detection system 200 of the intelligent motor in the embodiment of the present invention adopts the same technical means as those in the Figure 1 anomaly detection method of an intelligent motor described above, and can produce the same technical effects, which will not be elaborated here.

[0075] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An abnormal detection method for an intelligent motor, characterized in that, The method includes: Collecting multi-source heterogeneous data during the operation of the intelligent motor, performing feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, and performing potential Bent learning on the feature fusion data to obtain optimized data features; Querying the operation fluctuation state of the intelligent motor, segmenting the optimized data features based on the operation fluctuation state to obtain segmented features, and performing time-frequency joint analysis on the segmented features to determine the time-frequency feature maps of the segmented features at multiple resolutions; Using the time-frequency feature maps to mine the feature correlation relationships of the multi-dimensional operation features of the intelligent motor, constructing a reinforcement learning strategy for abnormal detection of the intelligent motor based on the feature correlation relationships, and using the reinforcement learning strategy to perform target feature screening on the multi-dimensional operation features of the intelligent motor to obtain a screened feature subset; Performing adversarial generation processing on the screened feature subset to obtain a simulated feature subset sample, performing feature integration learning processing on the simulated feature subset sample and the screened feature subset to obtain judgment features, and using the judgment features to perform abnormal analysis on the intelligent motor to obtain an abnormal detection report.

2. The abnormal detection method of an intelligent motor according to claim 1, characterized in that Performing feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, including: Classifying the data modalities of the multi-source heterogeneous data to obtain classified data; Extracting the features of different types of data in the classified data to obtain multi-modal features; Performing normalization processing on the multi-modal features to obtain normalized data features; Performing feature fusion on the normalized data features to obtain feature fusion data.

3. The abnormal detection method of an intelligent motor according to claim 1, wherein, Performing potential Bent learning on the feature fusion data to obtain optimized data features, including: Performing tensor reconstruction processing on the feature fusion data to obtain reconstructed data; Performing spatio-temporal attention encoding on the reconstructed data to obtain spatio-temporal fusion features; Performing feature compression on the spatio-temporal fusion features to obtain latent variable distribution parameters; Performing feature screening on the latent variable distribution parameters to obtain optimized data features.

4. The abnormal detection method of an intelligent motor according to claim 3, wherein Performing feature screening on the latent variable distribution parameters to obtain optimized data features, including: Calculating the mutual information values of each parameter feature in the latent variable distribution parameters using the following formula, ； Among them, represents the mutual information value, represents a parameter feature in the latent variable distribution parameters, represents an abnormal label in the latent variable distribution parameters, represents the probability that the feature X takes the value x and the abnormal label Y takes the value y, represents the probability that the feature X takes the value x, represents the probability that the abnormal label Y takes the value y; Based on the mutual information values, performing feature screening on the latent variable distribution parameters to obtain optimized data features.

5. The abnormal detection method of an intelligent motor according to claim 1, characterized in that, Segmenting the optimized data features based on the operation fluctuation state to obtain segmented features, including: Performing parameter quantization processing on the operation fluctuation state to obtain operation fluctuation parameters; Performing dynamic threshold modeling on the operation fluctuation parameters to obtain an adaptive fluctuation threshold; Using the adaptive fluctuation threshold to segment the optimized data features to obtain initial data segments; Detecting the boundary gradient of the initial data segments to dynamically optimize the initial data segments using the boundary gradient to obtain segmented features.

6. The abnormal detection method of an intelligent motor according to claim 1, characterized in that, Performing time-frequency joint analysis on the segmented features to determine the time-frequency feature maps of the segmented features at multiple resolutions, including: Performing short-time Fourier transform processing on the segmented features to obtain a low-frequency time-frequency matrix; Performing continuous wavelet transform on the low-frequency time-frequency matrix to obtain a high-frequency time-frequency matrix; Performing multi-resolution splicing on the low-frequency time-frequency matrix and the high-frequency time-frequency matrix to obtain a composite time-frequency spectrum; Adaptive threshold noise reduction is performed on the composite time-frequency spectrum to obtain a time-frequency feature spectrum.

7. The abnormal detection method of an intelligent motor according to claim 1, characterized in that, By using the time-frequency feature map, the feature correlation relationship of the multi-dimensional operation characteristics of the intelligent motor is mined, including: Using the time-frequency feature map, construct a feature correlation map of the smart motor; Performing graph neural convolution processing on the feature association graph to obtain a local feature association relationship of the intelligent motor; Based on the local feature association relationship, the attention mechanism is used to perform global feature mining on the smart motor to obtain a global feature association relationship; According to the global feature association relationship, the feature association relationship of the multi-dimensional operation feature of the intelligent motor is identified.

8. The abnormal detection method of an intelligent motor according to claim 1, characterized in that Based on the feature association relationship, a reinforcement learning strategy for intelligent motor anomaly detection is constructed, including: Using the characteristic association relationship, constructing the state space of the intelligent motor to obtain a state vector; Performing action discretization processing on the state vector to obtain a discrete action set; Construct a multi-objective reward function for the discrete action set, wherein the multi-objective reward function can be expressed by the following formula: ； Among them, F represents the multi-objective reward function, represents the anomaly detection accuracy reward for the detection decision action in the discrete action set, represents the decision response time reward for the detection decision action in the discrete action set, represents the misjudgment loss reward for the detection decision action in the discrete action set, represents the weight of represents the weight of represents the weight of; Based on the discrete action set and the multi-objective reward function, a reinforcement learning strategy for intelligent motor anomaly detection is constructed.

9. The abnormal detection method of an intelligent motor according to claim 1, characterized in that Performing feature ensemble learning processing on the simulated feature subset samples and the screened feature subset to obtain judgment features includes: Performing data fusion and alignment on the simulated feature subset samples and the screened feature subset to obtain a mixed data set; Using the trained heterogeneous base classifier, anomaly decision analysis is performed on the mixed data set to obtain a preliminary anomaly judgment result; Performing dynamic weighted integration processing on the preliminary abnormality judgment results to obtain an integrated decision result; Performing adversarial optimization processing on the integrated decision result to obtain an optimized decision result; The optimization decision result is deployed in a lightweight manner at the edge to obtain a judgment feature.

10. An abnormal detection system for an intelligent motor, characterized in that, The system comprises: The device feature processing module is used to collect multi-source heterogeneous data when the intelligent motor is working, perform feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, and perform latent feature learning on the feature fusion data to obtain optimized data features; an equipment feature analysis module, configured to query the operating fluctuation state of the intelligent motor, segment the optimized data features based on the operating fluctuation state to obtain segmentation features, and perform a time-frequency joint analysis on the segmentation features to determine a time-frequency feature spectrum of the segmentation features at multiple resolutions; a device feature enhancement module, configured to utilize the time-frequency feature map to mine feature correlations of the multi-dimensional operating features of the intelligent motor, construct a reinforcement learning strategy for intelligent motor anomaly detection based on the feature correlations, and utilize the reinforcement learning strategy to perform target feature screening on the multi-dimensional operating features of the intelligent motor to obtain a screening feature subset; The device anomaly detection module is used to perform adversarial generation processing on the screened feature subset to obtain a simulated feature subset sample, perform feature integration learning processing on the simulated feature subset sample and the screened feature subset to obtain judgment features, and use the judgment features to perform anomaly analysis on the intelligent motor to obtain an anomaly detection report.

Images