An intelligent motor anomaly detection method and system

Abstract

The application relates to the technical field of device anomaly detection, and discloses an abnormality detection method and system of an intelligent motor. The method comprises the following steps: collecting multi-source heterogeneous data of the intelligent motor during operation and performing feature-level fusion and potential feature learning to obtain optimized data features; performing data segment segmentation on the optimized data features to obtain segmented features, and determining a time-frequency feature graph of the segmented features under multi-resolution; mining a feature correlation relationship of multi-dimensional operation features of the intelligent motor to construct a reinforcement learning strategy, performing 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 simulation samples, performing feature set integration learning processing on the simulation samples and the screened feature subset to obtain judgment features, performing abnormality analysis on the intelligent motor, and obtaining an abnormality detection report. The application can improve the detection accuracy of unobvious or weak-feature abnormalities of the intelligent motor.

Description

TECHNICAL FIELD

[0001] The application relates to an abnormality detection method and system of an intelligent motor, and belongs to the technical field of device abnormality detection. BACKGROUND

[0002] With the progress of modern science and technology and the continuous development of production and application, intelligent motors play an increasingly important role in the fields of industry, commerce, aviation and military affairs. The failure of an intelligent motor not only damages the motor itself but also affects the normal work of the entire system, and even endangers personal safety and causes huge losses. By monitoring the parameters of an intelligent motor in real time, the intelligent motor can be detected and analyzed for abnormalities, potential failures such as mechanical wear, winding short circuit and bearing failure can be identified in advance, and the production loss, maintenance cost and energy utilization efficiency can be reduced.

[0003] At present, the abnormality detection of an intelligent motor mainly relies on single-mode signal analysis. First, the vibration signal or acoustic signal of the intelligent motor is collected, then the running parameters of the intelligent motor are identified by using time domain analysis or frequency domain analysis, and finally the abnormality is judged by comparing with a preset threshold. However, for the single-mode signal analysis method, the rules are mainly set by artificial experience, and there is a lack of fusion analysis and self-learning ability for multi-source heterogeneous data, so some abnormal problems with unobvious or weak characteristics cannot be found. SUMMARY

[0004] The application provides an abnormality detection method and system of an intelligent motor, which mainly aims to improve the detection accuracy of unobvious or weak characteristics of an intelligent motor.

[0005] To achieve the above-mentioned purpose, the application provides an abnormality detection method of an intelligent motor, which comprises the following steps:

[0006] Multi-source heterogeneous data during the operation of the intelligent motor is collected, feature-level fusion is performed on the multi-source heterogeneous data to obtain feature fusion data, and potential intrinsic learning is performed on the feature fusion data to obtain optimized data features;

[0007] The running fluctuation state of the intelligent motor is queried, data segment segmentation is performed on the optimized data features based on the running fluctuation state to obtain segmented features, time-frequency joint analysis is performed on the segmented features to determine a time-frequency feature map of the segmented features under multi-resolution;

[0008] The time-frequency feature map is used to mine the feature correlation of multi-dimensional running features of the intelligent motor, a reinforcement learning strategy for the abnormality detection of the intelligent motor is constructed based on the feature correlation, the reinforcement learning strategy is used to perform target feature screening on the multi-dimensional running features of the intelligent motor to obtain a screened feature subset;

[0009] The screening feature subset is subjected to a generative adversarial process to obtain a simulation feature subset sample, the simulation feature subset sample and the screening feature subset are subjected to feature set integration learning processing to obtain a judgment feature, the judgment feature is used for abnormal analysis of the intelligent motor to obtain an abnormal detection report.

[0010] Optionally, the multi-source heterogeneous data is subjected to feature-level fusion to obtain feature fusion data, including:

[0011] The multi-source heterogeneous data is subjected to data modal classification to obtain classified data.

[0012] Features of different types of data in the classified data are extracted to obtain multi-modal features.

[0013] The multi-modal features are subjected to standardization processing to obtain standardized data features.

[0014] The standardized data features are subjected to feature fusion to obtain feature fusion data.

[0015] Optionally, the feature fusion data is subjected to latent intrinsic learning to obtain optimized data features, including:

[0016] The feature fusion data is subjected to tensor reconstruction processing to obtain reconstructed data.

[0017] The reconstructed data is subjected to spatio-temporal attention coding to obtain spatio-temporal fusion features.

[0018] The spatio-temporal fusion features are subjected to feature compression to obtain hidden variable distribution parameters.

[0019] The hidden variable distribution parameters are subjected to feature screening to obtain optimized data features.

[0020] Optionally, the hidden variable distribution parameters are subjected to feature screening to obtain optimized data features, including:

[0021] The mutual information value of each parameter feature in the hidden variable distribution parameters is calculated using the following formula,

[0022] ;

[0023] wherein, represents the mutual information value, represents one parameter feature in the hidden variable distribution parameters, represents one abnormal label in the hidden variable distribution parameters, represents the probability of feature X taking value x and abnormal label Y taking value y, represents the probability of feature X taking value x, a probability that an abnormality label Y takes a value y;

[0024] Based on the mutual information value, the hidden variable distribution parameter is screened for features to obtain an optimized data feature.

[0025] Optionally, based on the running fluctuation state, the optimized data feature is segmented for data segments to obtain a segmented feature, including:

[0026] The running fluctuation state is quantitatively processed for parameters to obtain a running fluctuation parameter.

[0027] The running fluctuation parameter is modeled for a dynamic threshold to obtain an adaptive fluctuation threshold.

[0028] The adaptive fluctuation threshold is used to segment the optimized data feature for data segments to obtain an initial data segment.

[0029] The boundary gradient of the initial data segment is detected to dynamically optimize the initial data segment using the boundary gradient to obtain a segmented feature.

[0030] Optionally, the segmented feature is analyzed jointly in time and frequency to determine a time-frequency feature map of the segmented feature at multiple resolutions, including:

[0031] The segmented feature is processed for a short-time Fourier transform to obtain a low-frequency time-frequency matrix.

[0032] The low-frequency time-frequency matrix is processed for a continuous wavelet transform to obtain a high-frequency time-frequency matrix.

[0033] The low-frequency time-frequency matrix and the high-frequency time-frequency matrix are spliced for multiple resolutions to obtain a composite time-frequency map.

[0034] The composite time-frequency map is processed for adaptive threshold denoising to obtain a time-frequency feature map.

[0035] Optionally, the time-frequency feature map is used to mine feature correlation relationships of multi-dimensional running features of an intelligent motor, including:

[0036] The time-frequency feature map is used to construct a feature correlation graph of the intelligent motor.

[0037] The feature correlation graph is processed for graph neural convolution to obtain a local feature correlation relationship of the intelligent motor.

[0038] Based on the local feature correlation relationship, the intelligent motor is mined for a global feature using an attention mechanism to obtain a global feature correlation relationship.

[0039] According to the global feature correlation, a feature correlation of the intelligent motor multi-dimensional operation characteristic is identified.

[0040] Optionally, based on the feature correlation, a reinforcement learning strategy for intelligent motor anomaly detection is constructed, including:

[0041] Using the feature correlation, a state space of the intelligent motor is constructed to obtain a state vector;

[0042] The state vector is subjected to action discretization processing to obtain a discrete action set;

[0043] A multi-objective reward function of the discrete action set is constructed, wherein the multi-objective reward function can be represented by the following formula:

[0044] ;

[0045] Wherein F represents a multi-objective reward function, represents an anomaly detection accuracy reward of a detection decision action in the discrete action set, represents a decision response time reward of a detection decision action in the discrete action set, represents a misjudgment loss reward of a detection decision action in the discrete action set, represents a weight of represents a weight of represents a weight of represents a weight of represents a weight of ;

[0046] Based on the discrete action set and the multi-objective reward function, a reinforcement learning strategy for intelligent motor anomaly detection is constructed.

[0047] Optionally, the simulation feature subset sample and the screening feature subset are subjected to feature set integration learning processing to obtain a judgment feature, including:

[0048] The simulation feature subset sample and the screening feature subset are subjected to data fusion alignment to obtain a mixed data set;

[0049] Using the trained heterogeneous base classifier, the mixed data set is subjected to anomaly decision analysis to obtain a preliminary anomaly judgment result;

[0050] The preliminary anomaly judgment result is subjected to dynamic weighted integration processing to obtain an integrated decision result;

[0051] The integrated decision result is subjected to adversarial optimization processing to obtain an optimized decision result;

[0052] The optimized decision result is subjected to edge lightweight deployment to obtain a judgment feature.

[0053] To address the aforementioned problems, the present invention also provides an anomaly detection system for an intelligent motor, the system comprising:

[0054] The device feature processing module is used to collect multi-source heterogeneous data when the smart motor is working, 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.

[0055] The equipment feature analysis module is used to query the operating fluctuation status of the intelligent motor, perform data segmentation on the optimized data features based on the operating fluctuation status to obtain segmentation features, and perform time-frequency joint analysis on the segmentation features to determine the time-frequency feature spectrum of the segmentation features under multi-resolution.

[0056] The equipment feature enhancement module is used to mine the feature associations of the multi-dimensional operation features of the intelligent motor using the time-frequency feature map, construct a reinforcement learning strategy for intelligent motor anomaly detection based on the feature associations, and use the reinforcement learning strategy to filter the target features of the multi-dimensional operation features of the intelligent motor to obtain a subset of filtered features.

[0057] The equipment anomaly detection module is used to perform adversarial generation processing on the selected feature subset to obtain simulated feature subset samples, perform feature ensemble learning processing on the simulated feature subset samples and the selected feature subset to obtain judgment features, and use the judgment features to perform anomaly analysis on the smart motor to obtain an anomaly detection report.

[0058] Compared with the problems described in the background art, the embodiment of the application first deploys multi-modal sensors such as sound, vibration and current, collects original data (such as vibration time domain waveform, current spectrum, etc.) in the operation of the motor, after modal classification, feature extraction (such as current harmonic amplitude, vibration root mean square value), Z-score standardization, uses a deep neural network with attention mechanism for feature fusion, forms a unified format of feature fusion data, which can cover multi-dimensional data of machinery, acoustics and electricity, avoid missed detection caused by missing single modal information, and reconstruct the feature fusion data into a three-dimensional tensor, then use a variational autoencoder to compress it into a hidden variable distribution parameter, and finally select key features based on mutual information, which can reduce the computational complexity and improve the efficiency of subsequent algorithms; further, the application calculates the variance, change rate and other fluctuation parameters of the 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, ensuring that the vibration impact and current mutation correspond accurately in time, improving the reliability of correlation analysis, and for the segmented vibration / current signals, high-frequency details are extracted through short-time Fourier transform and continuous wavelet transform, and a composite time-frequency spectrum is formed to realize fault type visualization; further, the application maps the time-frequency spectrum feature points to graph nodes and constructs a feature correlation graph to generate feature correlation rules, then based on the feature correlation relationship, a reinforcement learning environment containing a 100-dimensional state vector (fusion time-frequency features, correlation weights, real-time parameters), 12 discrete actions (such as early warning, shutdown, adjusting sampling frequency) is constructed, a multi-objective reward function is designed, and the weight is flexibly set to improve the detection accuracy; further, the application uses a generative adversarial network (GAN) to generate simulated abnormal samples, and after fusion with real filtered features, the samples are detected in parallel by a heterogeneous base classifier, and then the model is compressed through dynamic weighted integration, adversarial optimization and knowledge distillation, and finally a 32-dimensional judgment feature vector is deployed on the edge device, which can make up for the lack of real fault data, combine vibration frequency, current harmonic and other multi-feature comprehensive judgment, and improve the accuracy of abnormal judgment. Therefore, the application can improve the detection accuracy of unobvious or weak feature abnormalities of intelligent motors. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A flowchart of an abnormality detection method for an intelligent motor provided by an embodiment of the application;

[0060] Figure 2 A module diagram of an abnormality detection system for an intelligent motor provided by an embodiment of the application.

[0061] The purpose of the application, functional features and advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0062] It is to be understood that the specific embodiments described herein are merely illustrative of the present application and do not limit the scope of the application.

[0063] The embodiment of the present application provides an abnormality detection method of an intelligent motor. The execution subject of the abnormality detection method of the intelligent motor includes but is not limited to at least one of electronic devices such as a server, a terminal and the like which can be configured to execute the method provided by the embodiment of the present application. In other words, the abnormality detection system of the intelligent motor can be executed by software or hardware installed in 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 and the like.

[0064] Embodiment 1

[0065] Referring to Figure 1 Fig. 1 shows a flowchart of an abnormality detection method of an intelligent motor provided by an embodiment of the present application. In the embodiment, the abnormality detection method of the intelligent motor includes:

[0066] S1, collecting multi-source heterogeneous data of the intelligent motor during operation, performing feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, performing latent feature learning on the feature fusion data to obtain optimized data features.

[0067] The embodiment of the present application can obtain vibration, sound and current and other information during motor operation by collecting multi-source heterogeneous data of the intelligent motor during operation, so as to ensure to obtain comprehensive and original information basis and provide data guarantee for motor abnormality detection.

[0068] The multi-source heterogeneous data refers to data of different sources, different types and different structures during operation of the intelligent motor, such as vibration data, temperature data and current data and the like.

[0069] 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 and the like.

[0070] Further, the embodiment of the present application can convert different types of original data into feature fusion data in a unified format by performing feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data, so as to retain multi-dimensional complementary information and provide comprehensive and refined analysis data support for subsequent abnormality detection.

[0071] As an embodiment of the present application, performing feature-level fusion on the multi-source heterogeneous data to obtain feature fusion data includes:

[0072] performing data modal classification on the multi-source heterogeneous data to obtain classified data;

[0073] Extracting features of different types of data in the classification data to obtain multi-modal features;

[0074] Standardizing the multi-modal features to obtain standardized data features;

[0075] Performing feature fusion on the standardized data features to obtain feature fusion data.

[0076] Among them, the multi-modal features refer to data features of different modalities, such as current data features, voltage data features, etc.

[0077] Optionally, the classification data can classify 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 standardized data features can be obtained by standardizing the multi-modal features using the Z-score standardization method, and the feature fusion data can be obtained by constructing a deep neural network containing a multi-layer perception, taking the standardized data features as input, and automatically assigning modal weights through network internal feature transformation and interaction combined with an attention mechanism.

[0078] Further, the embodiment of the present application can map data from high dimension to low dimension by performing latent intrinsic learning on the feature fusion data to obtain optimized data features, effectively extract data intrinsic features while reducing data dimension, generate samples conforming to the original data distribution, and further optimize the analysis data.

[0079] As an embodiment of the present application, the latent intrinsic learning on the feature fusion data to obtain the optimized data features comprises:

[0080] Performing tensor reconstruction processing on the feature fusion data to obtain reconstructed data;

[0081] Performing spatio-temporal attention encoding on the reconstructed data to obtain spatio-temporal fusion features;

[0082] Performing feature compression on the spatio-temporal fusion features to obtain latent variable distribution parameters;

[0083] Performing feature screening on the latent variable distribution parameters to obtain the optimized data features.

[0084] Among them, the spatio-temporal fusion features refer to features obtained by weighting and integrating key information of data in time series change and spatial correlation characteristics, and the latent variable distribution parameters refer to parameters used to describe the probability distribution state of the latent variable in the probability model, usually including mean, variance, etc., used to describe the latent distribution characteristics of the data.

[0085] Optionally, the reconstruction data can be obtained by reconstructing the feature fusion data into a three-dimensional tensor according to the time sequence characteristics and modal dimension of the vibration, sound and current data in the feature fusion data, the spatio-temporal fusion feature can be obtained by constructing a spatio-temporal attention mechanism network, calculating the attention weights on the time dimension and the modal dimension respectively, focusing on the key information of the motor operation data in the time sequence change and the correlation between modes, and generating a weighted fusion, the hidden variable distribution parameter can be obtained by inputting the spatio-temporal fusion feature into a multilayer perceptron through an encoder structure in a variational autoencoder, and outputting the mean vector and variance vector of the hidden variable through layer-by-layer dimension reduction compression.

[0086] Preferably, the feature screening of the hidden variable distribution parameter to obtain the optimized data feature comprises:

[0087] The mutual information value of each parameter feature in the hidden variable distribution parameter is calculated by the following formula,

[0088] ;

[0089] Wherein, represents the mutual information value, represents a parameter feature in the hidden variable distribution parameter, represents an abnormal label in the hidden variable distribution parameter, 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.

[0090] Based on the mutual information value, the feature screening of the hidden variable distribution parameter is performed to obtain the optimized data feature.

[0091] The mutual information value refers to a quantitative index for measuring the degree of dependence and information sharing between two random variables.

[0092] In specific implementation, a mutual information threshold (such as 0.2) can be set first, the parameter features less than the threshold are removed, the parameter features with high mutual information value and strong association with the abnormal label are retained, and then an optimized data feature subset is formed, such as the current 5th harmonic amplitude feature with a mutual information value of 0.3 is retained, and the certain temperature fluctuation feature with a mutual information value of 0.15 is discarded, finally forming the optimized data feature focusing on the key information and capable of improving the abnormal detection efficiency.

[0093] S2, query the running fluctuation state of the intelligent motor, segment the optimized data feature based on the running fluctuation state to obtain segmented features, and perform time-frequency joint analysis on the segmented features to determine the time-frequency feature atlas of the segmented features under multiple resolutions.

[0094] The embodiment of the present application can obtain the dynamic change of the motor operation parameter in real time by querying the operation fluctuation state of the intelligent motor, accurately master the operation characteristics of the motor under different working conditions, and intuitively present the fluctuation trend of the motor operation state.

[0095] The operation fluctuation state refers to the dynamic change of the vibration, sound, voltage and other operation parameters of the intelligent motor in the actual operation process, and the parameter change rate, fluctuation amplitude, change trend and other characteristics, reflecting the working stability and performance of the motor under different working conditions, such as a large change of the operation parameter in a short time, exceeding the normal fluctuation range.

[0096] Optionally, the operation fluctuation state can be determined by monitoring the vibration, sound, current and other multi-source heterogeneous data in the motor operation process, calculating the variance and change rate of the data, and comparing with the fluctuation range of the historical normal operation data.

[0097] Further, the embodiment of the present application can obtain the segmented feature by segmenting the data segment of the optimized data feature based on the operation fluctuation state, which can dynamically adjust the data segment of focus according to the actual operation change of the motor.

[0098] As an embodiment of the present application, the segmented feature is obtained by segmenting the data segment of the optimized data feature based on the operation fluctuation state, which includes:

[0099] The operation fluctuation state is quantitatively processed to obtain an operation fluctuation parameter;

[0100] The operation fluctuation parameter is dynamically threshold modeled to obtain an adaptive fluctuation threshold;

[0101] The adaptive fluctuation threshold is used to segment the data segment of the optimized data feature to obtain an initial data segment;

[0102] The boundary gradient of the initial data segment is detected to dynamically optimize the initial data segment by using the boundary gradient to obtain a segmented feature.

[0103] The adaptive fluctuation threshold refers to a critical value dynamically adjusted according to the real-time change of data, which is used to judge whether the data fluctuation is abnormal, and the operation fluctuation parameter refers to an index obtained by quantifying the amplitude, frequency, change rate and other characteristics of data fluctuation, which is used to describe the data fluctuation state.

[0104] Optionally, the running fluctuation parameter can be extracted from the time domain statistical features (root mean square value, peak factor) and frequency domain features (spectrum energy distribution) of the vibration, sound and current signals of the intelligent motor, a multi-dimensional feature vector is constructed by combining the parameter change rate and fluctuation amplitude, and then the running fluctuation parameter is obtained through principal component analysis dimension reduction, the adaptive fluctuation threshold value can be based on the fluctuation parameters of the historical normal running data of the intelligent motor, the probability distribution is fitted by using kernel density estimation (KDE), the probability distribution is fitted by using kernel density estimation (KDE) based on the fluctuation parameters of the historical normal running data of the intelligent motor, the 95% confidence interval of the running fluctuation parameter is calculated to serve as an initial threshold value, and then the threshold value is dynamically updated through the exponential weighted moving average (EWMA) algorithm in real-time monitoring, the initial data segment can be obtained by designing a sliding window algorithm based on fluctuation intensity, when the fluctuation parameter exceeds the threshold value, the window size is reduced (for example, from 100ms to 50ms), and when it is stable, the window is enlarged (for example, to 200ms), and the window step is dynamically adjusted according to the fluctuation frequency, and then the data segment is segmented by using the set sliding window to obtain the segmented feature, the segmented feature can be obtained by first calculating the gradient change rate of the fluctuation parameter between adjacent windows, marking the positions exceeding the set threshold value as candidate boundaries, then optimizing the boundary positions by using a dynamic programming algorithm to minimize the variance within the segment and maximize the difference between segments, and finally performing data synchronization alignment processing through multi-modal data timestamp matching and dynamic time warping (DTW) algorithm.

[0105] Further, the embodiment of the present application can clearly capture the characteristics of abnormal signals at different times and frequencies by performing time-frequency joint analysis on the segmented feature to determine the time-frequency feature atlas of the segmented feature at multiple resolutions, and can intuitively display the time-frequency feature of the motor running state.

[0106] The time-frequency feature atlas refers to a visual result obtained by performing time-frequency analysis on the dynamic time series segment of the intelligent motor by using short-time Fourier transform combined with wavelet transform.

[0107] As an embodiment of the present application, the time-frequency joint analysis on the segmented feature to determine the time-frequency feature atlas of the segmented feature at multiple resolutions comprises:

[0108] Performing short-time Fourier transform processing on the segmented feature to obtain a low-frequency time-frequency matrix;

[0109] Performing continuous wavelet transform processing on the low-frequency time-frequency matrix to obtain a high-frequency time-frequency matrix;

[0110] Performing multi-resolution splicing on the low-frequency time-frequency matrix and the high-frequency time-frequency matrix to obtain a composite time-frequency atlas;

[0111] Performing adaptive threshold denoising on the composite time-frequency atlas to obtain a time-frequency feature atlas.

[0112] The low-frequency time-frequency matrix refers to matrix data of low-frequency signals in time and frequency domain distribution after short-time Fourier transform of segmented features using a fixed window function, which can reflect the change of signal low-frequency components over time, the high-frequency time-frequency matrix refers to a matrix formed after capturing fine time-frequency characteristics of signal high-frequency components using continuous wavelet transform on the low-frequency time-frequency matrix, which highlights the frequency details of high-frequency signals at different time points, and the composite time-frequency spectrum refers to a visualized spectrum of time-frequency characteristics of signals under multi-resolution, which is obtained by aligning and splicing the low-frequency time-frequency matrix and the high-frequency time-frequency matrix along the time axis, fusing the low-frequency wide time resolution and high-frequency high frequency resolution characteristics, and comprehensively displaying the time-frequency characteristics of signals under multi-resolution.

[0113] In the implementation process, a fixed-size Hann window can be used to perform short-time Fourier transform on vibration, sound and current data in the segmented features, the time-domain signals in each time window are converted into frequency-domain signals through Fourier transform, the distribution of low-frequency signals in time and frequency domain is obtained, and the low-frequency time-frequency matrix is generated; Morlet wavelet is selected as the mother wavelet, continuous wavelet transform is performed on the data in the low-frequency time-frequency matrix, the high time resolution characteristics of wavelet transform in the high-frequency band are used to finely analyze the time-frequency characteristics of signal high-frequency components, and the high-frequency time-frequency matrix is generated; the low-frequency time-frequency matrix and the high-frequency time-frequency matrix are aligned along the time axis, the low-frequency wide time resolution and low-frequency resolution characteristics are vertically spliced with the high-frequency narrow time resolution and high-frequency resolution characteristics, and the composite time-frequency spectrum containing high and low frequency information is formed; based on the data distribution of the composite time-frequency spectrum, a local adaptive threshold algorithm (such as SureShrink algorithm) is used to dynamically calculate the threshold according to the signal energy in different regions, the noise components in the spectrum are processed by soft threshold shrinkage, the effective signal characteristics are retained, and the final time-frequency characteristic spectrum is generated.

[0114] S3, using the time-frequency characteristic spectrum, mining the feature correlation of the multi-dimensional running features of the intelligent motor, based on the feature correlation, constructing a reinforcement learning strategy for intelligent motor anomaly detection, using the reinforcement learning strategy, target feature screening is performed on the multi-dimensional running features of the intelligent motor, and a screened feature subset is obtained.

[0115] By using the time-frequency characteristic spectrum, mining the feature correlation of the multi-dimensional running features of the intelligent motor, the internal relationship and potential mode of the multi-dimensional data such as motor vibration, current and temperature in the time-frequency layer can be analyzed, the interaction mechanism between features is intuitively presented, and the correlation characteristics of different factors in the motor running state are clearly displayed.

[0116] As an embodiment of the application, using the time-frequency characteristic spectrum, mining the feature correlation of the multi-dimensional running features of the intelligent motor, includes:

[0117] construct a feature correlation graph of the intelligent motor by using the time-frequency feature map;

[0118] perform graph neural convolution processing on the feature correlation graph to obtain a local feature correlation relationship of the intelligent motor;

[0119] based on the local feature correlation relationship, perform global feature mining on the intelligent motor by using an attention mechanism to obtain a global feature correlation relationship;

[0120] identify the feature correlation relationship of the multi-dimensional running features of the intelligent motor according to the global feature correlation relationship.

[0121] The feature correlation graph refers to a graph structure formed by taking data features as nodes and constructing edges according to the correlation and similarity between the features, and is used to visually display the correlation relationship between the features. The local feature correlation relationship refers to the correlation mode and interaction relationship exhibited by the data features within their adjacent regions or directly connected nodes. The global feature correlation relationship refers to the long-distance dependence and mutual correlation relationship between different regional and hierarchical features across the entire data range, reflecting the overall connection between the features.

[0122] 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 spatial proximity and similarity of vibration, sound, and current signals in the time-frequency domain, the edge connection between nodes is constructed by calculating the cosine similarity or mutual information to form an initial feature correlation graph. A graph convolution network (GCN) is used to perform convolution operation on the feature correlation graph, and the neighborhood information of the nodes is aggregated through the message passing mechanism to extract the first-order neighborhood feature correlation mode between the nodes and obtain the local feature correlation relationship. A graph attention network (GAT) is introduced to calculate the attention weights between different nodes, adaptively focusing on key nodes and important correlation edges, mining long-distance dependence relationships across regions and modalities, and generating global feature correlation relationships. Based on the node representation and edge weight in the global feature correlation relationship, strong correlation relationships are filtered by setting a threshold, and then the correlation mode in the graph structure is converted into an interpretable multi-dimensional running feature correlation rule combined with domain knowledge, to realize the identification of the feature correlation relationship. For example, in the feature correlation graph, the edge weight between the node with a vibration frequency of 120 Hz and the node with a current 3rd harmonic amplitude is 0.85, which exceeds the set threshold of 0.7. Combined with the knowledge of the motor field, it can be known that rotor imbalance can cause 120 Hz vibration and simultaneously cause the current 3rd harmonic to increase. Therefore, the correlation rule "vibration 120 Hz frequency component is abnormal and current 3rd harmonic amplitude is abnormal, then the rotor is unbalanced" is obtained.

[0123] Furthermore, the embodiments of the present invention, by constructing a reinforcement learning strategy for intelligent motor anomaly detection based on the aforementioned feature association relationship, can aim at the accuracy and timeliness of motor anomaly detection, and dynamically optimize feature selection decisions through agent-environment interaction and reward feedback mechanisms.

[0124] The reinforcement learning strategy refers to the mechanism by which an agent selects the optimal action based on the current state during its interaction with the environment.

[0125] As an embodiment of the present invention, a reinforcement learning strategy for intelligent motor anomaly detection is constructed based on the aforementioned feature association relationship, including:

[0126] Using the aforementioned feature associations, the state space of the intelligent motor is constructed to obtain the state vector;

[0127] The state vector is discretized to obtain a discrete action set;

[0128] Construct a multi-objective reward function for the discrete action set, wherein the multi-objective reward function can be expressed by the following formula:

[0129] ;

[0130] Where F represents the multi-objective reward function, This represents the reward for the anomaly detection accuracy of the discrete action set detection decision action. This represents the reward for the decision response time of a discrete action set for detecting decision-making actions. This represents the loss or reward for misjudgment in the centralized detection and decision-making process for discrete actions. express The weight, express The weight, express The weights;

[0131] Based on the discrete action set and the multi-objective reward function, a reinforcement learning strategy for intelligent motor anomaly detection is constructed.

[0132] 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 that are classified into executable operations or behaviors and presented in discrete form. Each action is clear, independent and selectable to be executed. The multi-objective reward function refers to a mathematical function designed to evaluate the quality of behavior and provide feedback reward values ​​based on the results of different behaviors and according to the weight of each objective in order to optimize multiple objectives (such as efficiency, cost, benefit, etc.) at the same time.

[0133] Optionally, the state vector can perform feature engineering processing on the feature node state (such as the amplitude of the vibration frequency component) in the time-frequency feature map, the edge weight relationship (such as the association strength of the current harmonic and the vibration feature), and the real-time operation parameters (such as the temperature and the rotating speed) of the motor, and then map to a high-dimensional vector space after dimension reduction by principal component analysis (PCA), to obtain a state vector containing 100 (which can be set according to actual application) dimensions. The discrete action set can classify and encode abnormal detection decision actions (continuous monitoring, early warning prompt, primary shutdown or secondary shutdown) and data processing actions (adjusting the sampling frequency, replacing the time-frequency analysis algorithm, or switching the feature extraction method), adopt a hierarchical Q-learning architecture, process detection decision actions (4) in the upper layer, process data processing actions (3) in the lower layer, 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 application), and adopt a deep deterministic policy gradient (DDPG) algorithm in combination with an experience replay mechanism and a target network to train the agent in a simulation environment containing 1000 historical fault cases, then use priority experience replay to improve the training efficiency, and update the target network once every 100 time steps, to finally form a strategy capable of selecting the optimal detection and decision action according to the real-time state of the motor.

[0134] It should be further explained that the multi-objective reward function is obtained by designing a weighted sum form of the reward function, including three sub-targets of abnormal detection accuracy (weight 0.5), decision response time (weight 0.3), and misjudgment loss (weight 0.2), for example, a reward of +15 points for successfully detecting a bearing fault, a penalty of -10 points for a false alarm, and an additional deduction of -5 points for a response time exceeding 3 seconds, so as to realize the quantitative evaluation and optimization of multi-objective rewards.

[0135] Furthermore, the embodiment of the present application can select the most discriminative feature subset by using the reinforcement learning strategy to perform target feature screening on the multi-dimensional operating features of the intelligent motor, and efficiently and accurately mine the key features closely related to the abnormal state of the motor.

[0136] Optionally, the filtered feature subset can use the constructed reinforcement learning strategy to guide the agent to perform filtering operations on the multi-dimensional operating features in the simulation environment according to the multi-objective reward function, iteratively optimize the filtering strategy according to the state vector change and the reward feedback, and finally determine the feature subset that optimizes the abnormal detection performance.

[0137] S4, performing an adversarial generation process on the filtered feature subset to obtain a simulated feature subset sample, performing feature integration learning processing on the simulated feature subset sample and the filtered feature subset to obtain a judgment feature, using the judgment feature to perform abnormal analysis on the intelligent motor to obtain an abnormal detection report.

[0138] The embodiment of the application can generate diversified simulated abnormal feature samples by performing the adversarial generation process on the filtered feature subset to obtain a simulated feature subset sample, effectively expand the amount of abnormal data samples, enrich the data diversity, and alleviate the problem of lack of actual abnormal data.

[0139] The simulated feature subset sample refers to a sample simulated and generated by focusing on a key feature subset through a data generation technology (such as a generative adversarial network GAN, a variational autoencoder VAE).

[0140] Optionally, the simulated feature subset sample can use a generative adversarial network (GAN) to input the filtered feature subset into a generator to generate simulated data, distinguish real data from simulated data through a discriminator, optimize parameters through adversarial training of the two, and finally output simulated samples conforming to the real data distribution by the generator.

[0141] The embodiment of the application can comprehensively use original and expanded feature data by performing the feature integration learning processing on the simulated feature subset sample and the filtered feature subset to obtain a judgment feature, and improve the recognition accuracy and stability of complex abnormal conditions of the intelligent motor.

[0142] As an embodiment of the application, the feature integration learning processing on the simulated feature subset sample and the filtered feature subset to obtain a judgment feature includes:

[0143] Performing data fusion alignment on the simulated feature subset sample and the filtered feature subset to obtain a mixed data set;

[0144] Using a trained heterogeneous base classifier to perform abnormal decision analysis on the mixed data set to obtain a preliminary abnormal judgment result;

[0145] Performing dynamic weighted integration processing on the preliminary abnormal judgment result to obtain an integrated decision result;

[0146] Performing adversarial optimization processing on the integrated decision result to obtain an optimized decision result;

[0147] Performing edge lightweight deployment on the optimized decision result to obtain a judgment feature.

[0148] The preliminary abnormality judgment result refers to the original judgment result of each classifier for whether a sample is abnormal or not and a specific abnormal type (such as bearing failure, rotor imbalance, etc.) output after independent abnormality detection on the mixed data set, and the integrated decision result refers to a comprehensive decision result fused by a dynamic weighted integration technology on the preliminary abnormality judgment results of multiple heterogeneous base classifiers.

[0149] In the implementation process, a dynamic time warping (DTW) algorithm can be used to perform time sequence alignment on the time sequence data of the simulation feature subset sample and the screening feature subset, and then a tensor decomposition technology is used to reconstruct the multi-modal features such as vibration, current and temperature into a unified three-order tensor (sample x time step x feature) of a uniform dimension, forming a mixed data set with consistent dimensions and synchronous time sequences, for example, aligning the simulation-generated bearing failure vibration features with the real screened current harmonic features according to the time stamp; deploying three base classifiers, support vector machine (SVM), random forest (RF) and long short-term memory network (LSTM), which have been trained on historical fault data, to perform parallel abnormality detection on the mixed data set; the SVM detects bearing failure based on vibration features, the RF analyzes current harmonics to identify rotor imbalance, and the LSTM captures temperature time sequence changes to judge winding abnormalities, and finally outputs preliminary abnormality detection results containing categories such as “normal”, “bearing failure” and “rotor imbalance”; based on the performance (accuracy, recall rate, F1 value) of each base classifier on the verification set, an adaptive weight adjustment mechanism is designed to calculate the credibility weight of each classifier under the current data features in real time; for example, when the vibration features dominate, the weight of the LSTM is increased to 0.6, and the integrated decision result is generated by weighted voting; a generative adversarial network (GAN) architecture is constructed, the generator generates abnormal samples (such as features of a slightly unbalanced motor rotor) close to the real distribution, and the discriminator distinguishes between the integrated decision result and 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; the knowledge distillation technology is used to compress the optimized decision result into a lightweight neural network, and finally deployed to an edge computing device to generate a low-dimensional judgment feature vector with a dimension of 32.

[0150] The embodiment of the application can accurately identify the abnormal state of the motor based on the key features optimized comprehensively, and clearly present information such as abnormal type, position and degree in the form of an intuitive report, thereby providing a direct and reliable basis for motor maintenance and decision-making.

[0151] In the implementation process, the judgment feature input can be deployed in a lightweight convolutional neural network model of an edge computing device, the features are classified through model pre-training parameters, compared with an intelligent motor anomaly database, the abnormal type, occurrence probability and severity are determined, and an abnormal detection report containing the fault position (such as bearing, rotor), occurrence probability (such as 85%), fault level (3-level serious fault) and processing suggestion is output. For example, when the vibration frequency abnormal feature value in the input judgment feature vector reaches 0.8 and the current harmonic feature value is 0.7, the model determines that it is a bearing fault, the occurrence probability is 85%, it belongs to a 3-level serious fault, and it is recommended to stop immediately for maintenance.

[0152] Embodiment 2:

[0153] As shown in Figure 2 FIG. 1 is a functional module diagram of an abnormal detection system of an intelligent motor according to the present application.

[0154] The abnormal detection system 200 of the intelligent motor according to the present application can be installed in an electronic device. According to the functions implemented, the abnormal detection system of the intelligent motor can 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. The modules according to the present application can also be referred to as units, which refer to a series of computer program segments that can be executed by an electronic device processor and can complete a fixed function, which are stored in the memory of the electronic device.

[0155] In the embodiments of the present application, the functions of each module / unit are as follows:

[0156] The device feature processing module 201 is configured to collect multi-source heterogeneous data of the intelligent motor during operation, 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.

[0157] The device feature analysis module 202 is configured to query the running fluctuation state of the intelligent motor, perform data segment segmentation on the optimized data features based on the running fluctuation state to obtain segmented features, perform time-frequency joint analysis on the segmented features to determine a time-frequency feature map of the segmented features under multi-resolution, and output the time-frequency feature map.

[0158] The device feature enhancement module 203 is configured to use the time-frequency feature map to mine feature correlation relationships of multi-dimensional running features of the intelligent motor, construct a reinforcement learning strategy for intelligent motor anomaly detection based on the feature correlation relationships, use the reinforcement learning strategy to perform target feature screening on the multi-dimensional running features of the intelligent motor to obtain a filtered feature subset, and output the filtered feature subset.

[0159] The device anomaly detection module 204 is configured to perform an adversarial generation process on the screening feature subset to obtain simulated feature subset samples, perform feature set integration learning processing on the simulated feature subset samples and the screening feature subset to obtain a judgment feature, perform anomaly analysis on the intelligent motor by using the judgment feature, and obtain an anomaly detection report.

[0160] In detail, the modules in the intelligent motor anomaly detection system 200 in the embodiments of the present application use the same technical means as the intelligent motor anomaly detection method in the above-mentioned Figure 1

[0161] It is apparent for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application.​

Claims

1. A method for detecting anomalies in an intelligent motor, characterized in that, The method includes: Collect multi-source heterogeneous data during the operation of the intelligent motor. This multi-source heterogeneous data refers to vibration data, temperature data, and current data generated during the motor's operation. Perform feature-level fusion on this multi-source heterogeneous data to obtain feature-fused data. Then, perform latent Ben-Taiwanese learning on the feature-fused data to obtain optimized data features. Specifically, this includes: The feature fusion data is subjected to tensor reconstruction processing to obtain reconstructed data; Spatiotemporal attention encoding is performed on the reconstructed data to obtain spatiotemporal fusion features; The spatiotemporal fusion features are compressed to obtain the latent variable distribution parameters; The latent variable distribution parameters are subjected to feature filtering to obtain optimized data features; The process involves querying the operational fluctuation state of the intelligent motor, segmenting the optimized data features based on the operational fluctuation state to obtain segmentation features. Specifically, this includes: performing parameter quantization on the operational fluctuation state to obtain operational fluctuation parameters; performing dynamic threshold modeling on the operational 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 segmentation features; and performing time-frequency joint analysis on the segmentation features to determine the time-frequency feature spectrum of the segmentation features under multi-resolution conditions. Using the aforementioned time-frequency feature map, feature correlations of multi-dimensional operating features of the intelligent motor are mined. Based on these feature correlations, a reinforcement learning strategy for intelligent motor anomaly detection is constructed. Specifically, this includes: constructing the state space of the intelligent motor using the feature correlations to obtain a state vector; discretizing the state vector into actions to obtain a discrete action set; constructing a multi-objective reward function for the discrete action set; constructing a reinforcement learning strategy for intelligent motor anomaly detection based on the discrete action set and the multi-objective reward function; and using the reinforcement learning strategy to filter target features of the multi-dimensional operating features of the intelligent motor to obtain a filtered feature subset. The reinforcement learning strategy employs a deep deterministic policy gradient algorithm, combined with an experience replay mechanism and target network construction. Furthermore, the multi-objective reward function is obtained by designing a weighted sum form reward function, which includes three sub-objectives: anomaly detection accuracy, decision response time, and misjudgment loss. Adversarial generation processing is performed on the selected feature subset to obtain simulated feature subset samples. Feature ensemble learning processing is then performed on the simulated feature subset samples and the selected feature subset to obtain judgment features. Using the judgment features, anomaly analysis is performed on the smart motor to determine the anomaly type, probability of occurrence, and severity of the smart motor. An anomaly detection report containing the fault location, probability of occurrence, fault level, and handling suggestions is output.

2. The anomaly detection method for an intelligent motor as described in claim 1, characterized in that, The multi-source heterogeneous data is fused at the feature level to obtain feature fused data, including: The multi-source heterogeneous data is subjected to data modality classification to obtain classified data; Features of different types of data are extracted from the classified data to obtain multimodal features; The multimodal features are standardized to obtain standardized data features; The standardized data features are fused to obtain fused data.

3. The anomaly detection method for an intelligent motor as described in claim 1, characterized in that, Feature filtering is performed on the latent variable distribution parameters to obtain optimized data features, including: The mutual information values ​​of each parameter feature in the latent variable distribution parameters are calculated using the following formula. ； in, Represents mutual information value, This represents a characteristic parameter among the latent variable distribution parameters. This represents an anomaly label in the latent variable distribution parameters. This represents the probability that feature X has the value x and anomaly label Y has the value y. This represents the probability that feature X takes the value x. This represents the probability that the anomaly label Y takes the value y. Based on the mutual information value, feature filtering is performed on the latent variable distribution parameters to obtain optimized data features.

4. The anomaly detection method for an intelligent motor as described in claim 1, characterized in that, Perform time-frequency joint analysis on the segmentation features to determine the time-frequency feature maps of the segmentation features at multiple resolutions, including: The segmentation features are subjected to short-time Fourier transform processing to obtain a low-frequency time-frequency matrix; The low-frequency time-frequency matrix is ​​subjected to continuous wavelet transform to obtain the high-frequency time-frequency matrix; The low-frequency time-frequency matrix and the high-frequency time-frequency matrix are spliced ​​together at multiple resolutions to obtain a composite time-frequency spectrum. Adaptive threshold denoising is performed on the composite time-frequency spectrum to obtain a time-frequency feature spectrum.

5. The anomaly detection method for an intelligent motor as described in claim 1, characterized in that, Using the aforementioned time-frequency feature map, feature correlations of multi-dimensional operating features of intelligent motors are mined, including: Using the time-frequency feature map, a feature association map of the intelligent motor is constructed; The feature association graph is processed by graph neural convolution to obtain the local feature association relationships of the smart motor; Based on the local feature associations, an attention mechanism is used to perform global feature mining on the smart motor to obtain global feature associations. Based on the global feature association relationship, the feature association relationship of the multi-dimensional operation characteristics of the intelligent motor is identified.

6. The anomaly detection method for an intelligent motor as described in claim 1, characterized in that, The simulated feature subset samples and the selected feature subset are subjected to feature ensemble learning processing to obtain judgment features, including: The simulated feature subset samples and the selected feature subset are fused and aligned to obtain a hybrid dataset; Using a trained heterogeneous base classifier, anomaly decision analysis is performed on the mixed dataset to obtain preliminary anomaly judgment results; The preliminary anomaly assessment results are dynamically weighted and integrated to obtain the integrated decision result; The integrated decision results are subjected to adversarial optimization processing to obtain optimized decision results; The optimized decision results are then deployed in a lightweight, edge-based manner to obtain the judgment features.

7. An anomaly detection system for an intelligent motor, characterized in that, The system includes: The device feature processing module is used to collect multi-source heterogeneous data during the operation of the intelligent motor. This multi-source heterogeneous data refers to vibration data, temperature data, and current data generated during the operation of the intelligent motor. Feature-level fusion is performed on this multi-source heterogeneous data to obtain feature fusion data. Latent Ben-Taiwan learning is then performed on the feature fusion data to obtain optimized data features, specifically including: The feature fusion data is subjected to tensor reconstruction processing to obtain reconstructed data; Spatiotemporal attention encoding is performed on the reconstructed data to obtain spatiotemporal fusion features; The spatiotemporal fusion features are compressed to obtain the latent variable distribution parameters; The latent variable distribution parameters are subjected to feature filtering to obtain optimized data features; The equipment feature analysis module is used to query the operating fluctuation status of the intelligent motor, and to segment the optimized data features based on the operating fluctuation status to obtain segmentation features. Specifically, it includes: performing parameter quantization processing on the operating fluctuation status to obtain operating fluctuation parameters; performing dynamic threshold modeling on the operating 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 segmentation features; and performing time-frequency joint analysis on the segmentation features to determine the time-frequency feature spectrum of the segmentation features under multi-resolution. The equipment feature enhancement module is used to mine the feature correlations of multi-dimensional operating features of the intelligent motor using the time-frequency feature map, and to construct a reinforcement learning strategy for intelligent motor anomaly detection based on the feature correlations. Specifically, it includes: constructing the state space of the intelligent motor using the feature correlations to obtain a state vector; discretizing the state vector into actions to obtain a discrete action set; constructing a multi-objective reward function for the discrete action set; constructing a reinforcement learning strategy for intelligent motor anomaly detection based on the discrete action set and the multi-objective reward function; and using the reinforcement learning strategy to filter target features of the multi-dimensional operating features of the intelligent motor to obtain a filtered feature subset. The reinforcement learning strategy adopts a deep deterministic policy gradient algorithm, combined with an experience replay mechanism and target network construction. Furthermore, the multi-objective reward function is obtained by designing a weighted sum form reward function, which includes three sub-objectives: anomaly detection accuracy, decision response time, and misjudgment loss. The equipment anomaly detection module is used to perform adversarial generation processing on the selected feature subset to obtain simulated feature subset samples, perform feature ensemble learning processing on the simulated feature subset samples and the selected feature subset to obtain judgment features, and use the judgment features to perform anomaly analysis on the smart motor to obtain an anomaly detection report.

Images