Intelligent sensor data exception identification method and device and storage medium

Abstract

The invention relates to an intelligent sensor data anomaly identification method and device and a storage medium, and the method comprises the following steps: carrying out the abnormal region boundary extraction of a feature component probability distribution graph through a self-adaptive threshold segmentation algorithm when the feature component probability distribution graph represents that the original time series data is abnormal; obtaining a candidate abnormal event boundary sequence; performing topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal mode relation graph; and performing fault factor analysis on the intelligent sensor based on the multi-dimensional abnormal mode relational graph to obtain an abnormal factor of the intelligent sensor, and obtaining a corresponding maintenance strategy based on the abnormal factor, thereby solving the problem that a traditional abnormal detection method often depends on simple threshold setting or a statistical model, and is not easy to maintain. The method is difficult to adapt to variable and complex practical application scenes, and especially is poor in performance when facing non-linear and non-stable data.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent sensors, and in particular to an intelligent sensor data anomaly recognition method, device and storage medium. Background Art

[0002] In modern industrial production and daily life, smart sensors are increasingly used. As an important tool for information collection, they play a key role in ensuring the normal operation of the system and improving production efficiency. However, with the continuous expansion of the scale and complexity of sensor networks, how to accurately and efficiently identify anomalies in sensor data has become an urgent problem to be solved. Traditional anomaly detection methods often rely on simple threshold settings or statistical models, which are difficult to adapt to the changing and complex actual application scenarios, especially when facing nonlinear and non-stationary data.

[0003] In addition, the massive time series data generated by smart sensors contains rich information, but also brings huge processing challenges. These data are not only large in volume but also complex in structure. Traditional methods are difficult to effectively mine the deep features contained in them, thus limiting the accuracy and reliability of anomaly detection. This puts forward higher requirements for real-time monitoring and fault prediction, and existing technical means are difficult to meet this demand. Therefore, it is necessary to develop a new method that can automatically extract multi-level features from raw data and combine advanced data analysis technology for anomaly identification.

[0004] In response to the above problems, researchers have begun to explore more intelligent solutions, using advanced algorithms such as wavelet transform, kernel density estimation, and graph convolutional networks to improve the ability of intelligent sensor data anomaly identification. This method can not only deeply explore the potential patterns in the data, but also accurately analyze the fault factors by constructing a multi-dimensional abnormal pattern relationship diagram, thereby providing a basis for formulating effective maintenance strategies. However, integrating these advanced technologies into a unified framework and achieving the expected results in practical applications still faces challenges such as algorithm optimization and parameter adjustment. The existence of these problems has promoted the continued research and development in related fields. Summary of the invention

[0005] The main purpose of the present invention is to provide a method, device and storage medium for identifying anomaly in intelligent sensor data, which solves the technical problem that traditional anomaly detection methods often rely on simple threshold settings or statistical models, are difficult to adapt to changeable and complex practical application scenarios, especially when facing nonlinear and non-stationary data.

[0006] To achieve the above object, the present invention provides a method for identifying abnormal data of an intelligent sensor, comprising the following steps: Perform multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set; Perform distribution probability density mapping on the multi-level signal feature component set based on the non-parametric kernel density estimation algorithm to obtain a feature component probability distribution map; When the feature component probability distribution map indicates that there is an abnormality in the original time-series data, extract the boundary of the abnormal area from the feature component probability distribution map through the adaptive threshold segmentation algorithm to obtain a candidate abnormal event boundary sequence; Perform topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph; Perform fault factor analysis on the intelligent sensor based on the multi-dimensional abnormal pattern relationship graph to obtain the abnormal factors of the intelligent sensor, and obtain the corresponding maintenance strategy based on the abnormal factors.

[0007] Further, the performing multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set includes: Perform spectral analysis preprocessing on the original time-series data collected by the intelligent sensor to obtain a frequency-domain signal distribution map, and perform multi-resolution adaptive segmentation on the frequency-domain signal distribution map to obtain a frequency band division matrix; wherein, the frequency band division matrix includes the main frequency band interval boundary value and the sub-frequency band interval energy distribution characteristics; Perform discrete wavelet packet transform on the original time-series data based on the frequency band division matrix to obtain a multi-level wavelet coefficient set, and perform denoising processing on the multi-level wavelet coefficient set through a non-linear threshold shrinkage function to obtain an optimized wavelet coefficient group; wherein, the optimized wavelet coefficient group includes high-frequency change feature components and low-frequency trend feature components; Perform feature component extraction on the optimized wavelet coefficient group to obtain a multi-level signal feature component set.

[0008] Further, the performing distribution probability density mapping on the multi-level signal feature component set based on the non-parametric kernel density estimation algorithm to obtain a feature component probability distribution map includes: Perform feature space orthogonal transformation on the multi-level signal feature component set to obtain a multi-dimensional feature representation matrix, and perform kernel density sampling calculation on the multi-dimensional feature representation matrix to obtain a multi-level signal feature density sequence; Perform kernel function expansion on the multi-level signal feature density sequence to obtain a kernel function expansion coefficient group, and perform bandwidth selection and optimization on the kernel function expansion coefficient group to obtain an adaptive density estimation sequence; Perform probability density reconstruction on the adaptive density estimation sequence through integral transform technology to obtain an initial probability distribution field, and perform boundary correction calculation on the initial probability distribution field to obtain a corrected probability density map; Perform multi-scale decomposition operation on the corrected probability density map based on the non-parametric kernel density estimation algorithm to obtain a hierarchical density feature group, and perform feature space mapping on the hierarchical density feature group to obtain a density feature mapping set; Construct a probability field based on the density feature mapping set to obtain a multi-dimensional probability density field, and perform feature fusion operation on the multi-dimensional probability density field to obtain a feature component probability distribution map.

[0009] Further, perform abnormal region boundary extraction on the feature component probability distribution map through an adaptive threshold segmentation algorithm to obtain a candidate abnormal event boundary sequence, including: Perform gradient calculation and direction analysis on the feature component probability distribution map to obtain a probability distribution gradient field, and perform local extreme value detection on the probability distribution gradient field to obtain an initial threshold candidate set; Perform probability distribution region segmentation on the initial threshold candidate set through an adaptive threshold segmentation algorithm to obtain an initial segmentation region group, and perform boundary curvature calculation on the initial segmentation region group to obtain a region boundary feature sequence; Perform boundary expansion on the region boundary feature sequence through a preset region growing technique to obtain an extended boundary contour set, and perform morphological processing on the extended boundary contour set to obtain an optimized boundary sequence; Perform topological structure analysis on the optimized boundary sequence to obtain a boundary topological relationship graph, and perform region merging calculation on the boundary topological relationship graph to obtain a candidate abnormal region set; Perform temporal correlation analysis on the candidate abnormal region set to obtain a temporal correlation feature group, and perform boundary sequence screening and fusion on the temporal correlation feature group to obtain a candidate abnormal event boundary sequence; wherein, the candidate abnormal event boundary sequence includes temporal abnormal interval boundary markers, abnormal event topological contour features, abnormal region density gradient vectors, and abnormal boundary curvature feature descriptors.

[0010] Further, perform topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph, including: Perform multi-dimensional feature encoding conversion on the candidate abnormal event boundary sequence to obtain a boundary event feature vector group, and perform graph structure construction on the boundary event feature vector group to obtain an initial abnormal event relationship graph; Perform Laplacian matrix decomposition on the initial abnormal event relationship graph through a preset spectral domain transformation technique to obtain a spectral feature descriptor set, and perform multi-level convolutional operations on the spectral feature descriptor set to obtain a hierarchical graph convolutional feature group; Through a graph convolutional network, perform cross-domain correlation mapping on a preset abnormal event based on the hierarchical graph convolutional feature group to obtain an abnormal pattern correlation matrix, and perform graph attention calculation on the abnormal pattern correlation matrix to obtain a weighted abnormal relationship network; Perform dynamic graph evolution analysis on the weighted abnormal relationship network to obtain a set of abnormal propagation paths, and perform multi-scale fusion processing on the set of abnormal propagation paths to obtain a multi-level abnormal relationship topology graph; Perform spatio-temporal correlation embedding on the multi-level abnormal relationship topology graph to obtain an abnormal event correlation embedding vector space, and perform abnormal pattern extraction on the abnormal event correlation embedding vector space to obtain a multi-dimensional abnormal pattern relationship graph; wherein, the multi-dimensional abnormal pattern relationship graph includes abnormal event type correlation metrics and abnormal pattern evolution characteristics.

[0011] Further, performing multi-dimensional feature encoding conversion on the candidate abnormal event boundary sequence to obtain a boundary event feature vector group, including: Perform time-frequency domain decomposition and reconstruction on the candidate abnormal event boundary sequence to obtain a multi-domain boundary feature matrix, and perform non-linear projection conversion on the multi-domain boundary feature matrix to obtain a dimensionality-reduced feature space; Perform local sensitive hashing encoding on the dimensionality-reduced feature space to obtain a feature hashing encoding set, and perform multi-scale redundancy analysis on the feature hashing encoding set to obtain a compressed feature encoding table; Perform high-order relationship extraction on the compressed feature encoding table through tensor decomposition technology to obtain a multi-modal feature interaction tensor, and perform nuclear norm minimization processing on the multi-modal feature interaction tensor to obtain a sparse feature descriptor set; Perform manifold learning and embedding on the sparse feature descriptor set to obtain a low-dimensional embedding space, and perform similarity measurement and clustering on the low-dimensional embedding space to obtain a boundary event feature vector group; wherein, the boundary event feature vector group includes abnormal event category encoding and spatio-temporal correlation intensity characterization.

[0012] Further, performing fault factor analysis on the intelligent sensor based on the multi-dimensional abnormal pattern relationship graph to obtain abnormal factors of the intelligent sensor, including: Perform abnormal event topological feature decomposition on the multi-dimensional abnormal pattern relationship graph to obtain an abnormal event topological feature vector set, and perform multi-scale singular value decomposition on the abnormal event topological feature vector set to obtain a sensor abnormal feature matrix; Perform multi-dimensional feature space mapping on the preset sensor state based on the sensor anomaly feature matrix to obtain a sensor state feature space representation, and perform Gaussian mixture clustering analysis on the sensor state feature space representation to obtain a set of sensor state clustering clusters; Perform dynamic Bayesian network modeling on the set of sensor state clustering clusters to obtain a sensor state transition probability matrix, and perform hidden Markov model parameter estimation on the sensor state transition probability matrix to obtain a sensor state hidden variable sequence; Perform multivariate statistical process control analysis on the sensor state hidden variable sequence to obtain a sensor fault control chart, and perform fault feature extraction and quantification on the sensor fault control chart to obtain a sensor anomaly factor.

[0013] The present invention also provides an intelligent sensor data anomaly recognition device, including: A decomposition module for performing multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a set of multi-level signal feature components; A mapping module for performing distribution probability density mapping on the set of multi-level signal feature components based on a non-parametric kernel density estimation algorithm to obtain a feature component probability distribution map; An extraction module for, when the feature component probability distribution map indicates that the original time-series data is abnormal, performing abnormal region boundary extraction on the feature component probability distribution map through an adaptive threshold segmentation algorithm to obtain a candidate abnormal event boundary sequence; A first analysis module for performing topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph; A second analysis module for performing fault factor analysis on the intelligent sensor based on the multi-dimensional abnormal pattern relationship graph to obtain an anomaly factor of the intelligent sensor, and obtaining a corresponding maintenance strategy based on the anomaly factor.

[0014] The present invention also provides a computer device, including a memory and a processor, where a computer program is stored in the memory, and when the processor executes the computer program, the steps of the method described in any one of the above are implemented.

[0015] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method described in any one of the above are implemented.

[0016] The intelligent sensor data anomaly recognition method provided by the present invention includes the following steps: performing multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set; performing distribution probability density mapping on the multi-level signal feature component set based on the non-parametric kernel density estimation algorithm to obtain a feature component probability distribution map; when the feature component probability distribution map indicates that the original time-series data is abnormal, extracting the abnormal region boundary of the feature component probability distribution map through an adaptive threshold segmentation algorithm to obtain a candidate abnormal event boundary sequence; performing topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal pattern relationship map; performing fault factor analysis on the intelligent sensor based on the multi-dimensional abnormal pattern relationship map to obtain the abnormal factors of the intelligent sensor, and obtaining the corresponding maintenance strategy based on the abnormal factors. Through the above technical means, the technical problem that traditional anomaly detection methods often rely on simple threshold setting or statistical models and are difficult to adapt to changing and complex actual application scenarios, especially perform poorly in the face of non-linear and non-stationary data, is solved. The method realizes the processing of the feature component probability distribution map by using the adaptive threshold segmentation algorithm, and can dynamically adjust the threshold to adapt to different data characteristics and application scenarios. This makes the method more flexible and robust in the face of different types of data anomalies, and improves the efficiency and accuracy of abnormal region boundary extraction. Description of the Drawings

[0017] Figure 1 is a schematic diagram of the steps of the intelligent sensor data anomaly recognition method in an embodiment of the present invention; Figure 2 is a structural block diagram of the intelligent sensor data anomaly recognition device in an embodiment of the present invention; Figure 3 is a schematic structural block diagram of a computer device in an embodiment of the present invention.

[0018] The implementation, functional features and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. Detailed Embodiments

[0019] In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, and are not used to limit the present invention.

[0020] As Figure 1 shown, Figure 1 is a schematic diagram of the steps of an intelligent sensor data anomaly recognition method in an embodiment of the present invention; An embodiment of the present invention provides an intelligent sensor data anomaly recognition method, including the following steps: Step S1: Perform multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set.

[0021] Specifically, for the process of performing multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set, it is first necessary to understand the basic principle and application scenario of multi-scale wavelet transform. Multi-scale wavelet transform is a signal processing technology that decomposes the original time-series data through wavelet functions of different scales to extract the feature component set at different frequencies. In this process, it is crucial to select an appropriate wavelet basis function because it directly affects the effect of feature extraction. For example, when monitoring the operating state of equipment in an industrial production environment, the data collected by the sensor usually contains various frequency components, and these components may reflect different working states or potential faults of the equipment. When specifically implemented, first input the original time-series data collected by the intelligent sensor into the multi-scale wavelet transform algorithm, and use the pre-selected wavelet basis function (such as Haar, Daubechies, etc.) to decompose it. This decomposition process will refine the time-frequency resolution of the signal layer by layer according to the set different scale parameters, and then separate the feature component set that reflects the signal characteristics at different time scales. For example, when analyzing the vibration signal of a large mechanical equipment, the low-frequency part may represent the overall operating trend of the equipment, while the high-frequency part may contain friction noise between mechanical components or other local abnormal information. In this way, we can extract a representative and distinguishable multi-level signal feature component set from the complex original data, laying a foundation for further probability density mapping and anomaly detection. This not only improves the depth of understanding of the data but also provides a richer information source for the subsequent steps, making the entire anomaly recognition system more accurate and reliable. For example, when it is found that a certain high-frequency component significantly deviates from the normal range, this may be a warning signal that a certain part inside the equipment is about to fail, which helps to take maintenance measures in time to avoid greater losses. Therefore, this step plays a crucial role in the intelligent sensor data anomaly recognition method.

[0022] Step S2: Based on the non-parametric kernel density estimation algorithm, perform distribution probability density mapping on the multi-level signal feature component set to obtain a feature component probability distribution map.

[0023] Specifically, for the process of performing distribution probability density mapping on the multi-level signal feature component set based on the non-parametric kernel density estimation algorithm to obtain the feature component probability distribution map, it is first necessary to understand the basic principle of non-parametric kernel density estimation and its application in data analysis. Non-parametric kernel density estimation is a method for estimating the probability density function of data samples. It does not require prior assumption of the data distribution form, so it can more flexibly adapt to various complex data structures. In this step, we use the multi-level signal feature component set obtained by multi-scale wavelet transform decomposition as the input and utilize the non-parametric kernel density estimation algorithm to construct the probability density mapping of these feature components. In specific implementation, first select appropriate kernel functions (such as Gaussian kernel, Epanechnikov kernel, etc.) and bandwidth parameters, and these two factors directly affect the quality of the final probability density map. Then, for each element in the feature component set, use the selected kernel function to perform weighted summation on it to generate a smooth probability density function. For example, in the application scenario of monitoring the operating state of large mechanical equipment, a multi-level signal feature component set reflecting different working states of the equipment has been obtained through multi-scale wavelet transform. Next, use the non-parametric kernel density estimation algorithm to process these feature components, and the probability density distribution map of each component can be constructed. This step not only helps us identify which feature components fluctuate within the normal range but also discovers those abnormal points that deviate from the normal range. For example, if the probability density of a certain high-frequency component is significantly higher or lower than its historical average level, it may indicate that there are potential problems in this frequency band, such as wear or looseness of mechanical components. In this way, we can obtain a more intuitive and detailed feature component probability distribution map, providing a reliable basis for the subsequent extraction of the abnormal region boundary. In addition, this probability density mapping method can also reveal the internal connections and patterns between feature components, helping to deeply analyze the working state of the equipment and providing a scientific basis for formulating effective maintenance strategies. Therefore, this step plays a key role in the entire intelligent sensor data anomaly recognition process.

[0024] Step S3, when the feature component probability distribution map indicates that the original time series data is abnormal, the adaptive threshold segmentation algorithm is used to extract the abnormal region boundary of the feature component probability distribution map to obtain the candidate abnormal event boundary sequence.

[0025] Specifically, when the probability distribution map of the feature components indicates that there are anomalies in the original time series data, the process of extracting the boundary of the abnormal region from the probability distribution map of the feature components through the adaptive threshold segmentation algorithm to obtain the candidate abnormal event boundary sequence requires first understanding the basic principle of the adaptive threshold segmentation algorithm and its application in anomaly detection. The adaptive threshold segmentation algorithm is a method of dynamically adjusting the threshold, which can automatically determine the optimal segmentation point according to the local characteristics of the data, thereby effectively identifying the abnormal region. In this step, we use the probability distribution map of the feature components generated by the non-parametric kernel density estimation algorithm as the input, and use the adaptive threshold segmentation algorithm to identify and extract the boundary of the abnormal region. Specifically, first analyze the probability distribution map of the feature components to identify those regions that deviate significantly from the normal range. These regions usually show that the probability density value is much higher or lower than its expected level, indicating that there may be abnormal situations. Then, use the adaptive threshold segmentation algorithm, which dynamically adjusts the threshold according to the local statistical characteristics of the data to ensure that it can accurately capture the true anomalies and avoid false alarms. For example, in the application scenario of monitoring the operating state of large mechanical equipment, assume that we have obtained the probability distribution map of the feature components of the equipment vibration signal through multi-scale wavelet transform and non-parametric kernel density estimation. At this time, if it is found that the probability density of a certain high-frequency component is significantly higher than the historical average level, this may mean that there are potential problems in this frequency band, such as wear or looseness of mechanical components. Through the adaptive threshold segmentation algorithm, we can accurately locate the boundary of this abnormal region and mark it as the candidate abnormal event boundary sequence. The advantage of this algorithm is that it can flexibly adjust the threshold according to the specific situation of different feature components, thereby improving the accuracy and reliability of anomaly detection. For example, for a complex signal containing multiple different frequency components, the adaptive threshold segmentation algorithm can set the optimal threshold for each frequency band respectively to ensure that all potential anomalies can be effectively identified. Finally, the obtained candidate abnormal event boundary sequence will provide basic data for further topological correlation analysis, which helps to deeply understand the abnormal pattern and formulate corresponding maintenance strategies. Therefore, this step plays a crucial role in the whole process of intelligent sensor data anomaly recognition, connecting the previous and the next steps.

[0026] Step S4: Perform topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph.

[0027] Specifically, in the process of obtaining the multi-dimensional abnormal pattern relationship graph through topological correlation analysis of the candidate abnormal event boundary sequence by means of a graph convolutional network, it is first necessary to understand the basic principle of the graph convolutional network (GCN) and its application in complex data structures. The graph convolutional network is a deep learning model specifically designed to process graph-structured data, capable of effectively capturing the topological relationships and feature information between nodes. In this step, we use the candidate abnormal event boundary sequence generated by the adaptive threshold segmentation algorithm as the input, and utilize the graph convolutional network to analyze the topological correlations between these boundaries, thereby constructing a multi-dimensional abnormal pattern relationship graph that reflects the relationships between different abnormal patterns. When specifically implemented, the candidate abnormal event boundary sequence is first transformed into graph-structured data, where each abnormal event is regarded as a node, and the edges between the nodes are defined according to their temporal or spatial correlations. For example, in the application scenario of monitoring the operating status of large-scale mechanical equipment, assume that we have obtained the candidate abnormal event boundary sequence of the equipment vibration signal through multi-scale wavelet transform, non-parametric kernel density estimation, and adaptive threshold segmentation algorithm. Next, the graph convolutional network is used to process these boundary sequences. First, each abnormal event boundary is regarded as a node, and connections are established based on its spatio-temporal correlations with other boundaries. Then, through multiple iterative graph convolutional operations, the feature representations of each node are gradually updated, making the similarities and differences between adjacent nodes more obvious. This step not only helps to identify which abnormal events are closely related but also reveals the deep-seated abnormal patterns hidden behind the data. For example, if two abnormal events show a high degree of correlation in both time and frequency, then they will be closely connected in the final relationship graph, indicating that they may be caused by the same underlying fault. In this way, we can obtain a multi-dimensional abnormal pattern relationship graph, which not only shows the specific positions and features of each abnormal event but also reveals their interactions and influences. Such a relationship graph is crucial for subsequent fault factor analysis because it provides rich context information to help us more accurately locate the root cause of the problem and formulate effective maintenance strategies. Therefore, this step plays a crucial role in the entire process of abnormal identification of intelligent sensor data, ensuring the complete chain from the original data to the final solution.

[0028] Step S5: Based on the multi-dimensional abnormal pattern relationship graph, perform fault factor analysis on the intelligent sensor to obtain the abnormal factors of the intelligent sensor, and obtain the corresponding maintenance strategy based on the abnormal factors.

[0029] Specifically, for the process of performing fault factor analysis on the intelligent sensor based on the multi-dimensional abnormal pattern relationship graph to obtain the abnormal factors of the intelligent sensor and obtaining the corresponding maintenance strategy based on the abnormal factors, it is first necessary to understand how to use the multi-dimensional abnormal pattern relationship graph for in-depth fault factor analysis. The multi-dimensional abnormal pattern relationship graph not only shows the specific locations and characteristics of each abnormal event, but also reveals the interactions and influences between them, which provides rich context information for identifying potential fault factors. In this step, we will determine the root causes of the anomalies by analyzing these relationships and formulate corresponding maintenance strategies accordingly. When specifically implemented, first comprehensively analyze the nodes and edges in the multi-dimensional abnormal pattern relationship graph to identify which abnormal events have significant correlations or causal relationships. For example, in the application scenario of monitoring the operating status of large mechanical equipment, assume that we have obtained the multi-dimensional abnormal pattern relationship graph of the equipment vibration signal through multi-scale wavelet transform, non-parametric kernel density estimation, adaptive threshold segmentation algorithm, and graph convolutional network. Next, by analyzing each node (i.e., candidate abnormal event) and its connecting edges in the relationship graph in detail, it can be found that abnormal events in certain specific frequency bands are highly correlated in time and space, indicating that these anomalies may be caused by the same fault source. For example, if abnormal fluctuations occur simultaneously in high-frequency and low-frequency vibration signals and show synchrony in time, this may mean that there are problems such as loose or worn mechanical components. Further, by combining the historical maintenance records and operation data of the equipment, the specific faulty components or systems can be more accurately located. Once the main abnormal factors are determined, corresponding maintenance strategies can be formulated according to their characteristics and severity. For example, for minor loosening problems, simple tightening may be sufficient; while for severe wear, damaged components may need to be replaced or a major overhaul may be required. This method of fault factor analysis based on the multi-dimensional abnormal pattern relationship graph not only improves the accuracy of fault diagnosis, but also effectively reduces unnecessary maintenance costs and improves the overall operating efficiency of the system. Therefore, this step plays a crucial role in the entire process of abnormal data recognition of intelligent sensors, ensuring the full-process optimized management from data acquisition to fault resolution. In this way, the safe and stable operation of the equipment can be better guaranteed, the service life of the equipment can be extended, and the operation cost can be reduced.

[0030] In a specific embodiment, the multi-scale wavelet transform decomposition is performed on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set, including: Perform spectrum analysis preprocessing on the original time-series data collected by the intelligent sensor to obtain a frequency-domain signal distribution map, and perform multi-resolution adaptive segmentation on the frequency-domain signal distribution map to obtain a frequency band division matrix; wherein, the frequency band division matrix includes the boundary values of the main frequency band interval and the energy distribution characteristics of the sub-frequency band interval; Perform discrete wavelet packet transform on the original time series data based on the frequency band division matrix to obtain a multi-level wavelet coefficient set, and perform denoising processing on the multi-level wavelet coefficient set through a non-linear threshold shrinkage function to obtain an optimized wavelet coefficient group; wherein, the optimized wavelet coefficient group includes a high-frequency change feature component and a low-frequency trend feature component; Extract feature components from the optimized wavelet coefficient group to obtain a multi-level signal feature component set.

[0031] Specifically, perform spectral analysis preprocessing on the original time-series data collected by the intelligent sensor to obtain a frequency-domain signal distribution map, and perform multi-resolution adaptive segmentation on the frequency-domain signal distribution map to obtain a frequency band division matrix; wherein, the frequency band division matrix includes the boundary values of the main frequency band intervals and the energy distribution characteristics of the secondary frequency band intervals. Based on the frequency band division matrix, perform discrete wavelet packet transform on the original time-series data to obtain a multi-level wavelet coefficient set, and perform denoising processing on the multi-level wavelet coefficient set through a non-linear threshold shrinkage function to obtain an optimized wavelet coefficient group; wherein, the optimized wavelet coefficient group includes high-frequency change feature components and low-frequency trend feature components. Extract feature components from the optimized wavelet coefficient group to obtain a multi-level signal feature component set. This series of steps constitutes a complete signal processing flow, aiming to extract useful multi-level signal features from complex original time-series data. First, when performing spectral analysis preprocessing on the original time-series data collected by the intelligent sensor, we use the fast Fourier transform (FFT) or other spectral analysis methods to convert the time-domain signal into a frequency-domain signal, thereby obtaining a frequency-domain signal distribution map. This distribution map shows the energy distribution of different frequency components, providing a basis for our subsequent multi-resolution adaptive segmentation. For example, in the application scenario of monitoring the operating status of large mechanical equipment, the vibration signals collected by the sensor usually contain multiple frequency components, and these components may reflect different working states or potential faults of the equipment. By performing spectral analysis on these signals, we can identify which frequency bands have significantly higher energy than other parts, which may be early warning signals for potential problems. Next, we need to perform multi-resolution adaptive segmentation on the frequency-domain signal distribution map to obtain a frequency band division matrix. In this process, the algorithm will automatically adjust the segmentation strategy according to the energy distribution of the signal to ensure that each frequency band can accurately reflect its internal characteristics. The frequency band division matrix not only contains the boundary values of the main frequency band intervals but also records the energy distribution characteristics of the secondary frequency band intervals, which is crucial for subsequent wavelet transform. For example, in the above mechanical equipment monitoring scenario, if it is found that the energy of a certain high-frequency band is abnormally high, it indicates that there may be local faults or wear problems in this frequency band, and the frequency band division matrix can help us accurately locate and separate these key frequency bands. Based on the obtained frequency band division matrix, the next step is to perform discrete wavelet packet transform (DWPT) on the original time-series data. Discrete wavelet packet transform is a signal processing technology that can provide finer frequency resolution. It recursively decomposes different frequency bands of the signal to generate a multi-level wavelet coefficient set. These wavelet coefficients contain rich information, including both high-frequency change feature components and low-frequency trend feature components. For example, in mechanical equipment monitoring, the high-frequency components may reflect the friction noise between mechanical components or other local abnormal information, while the low-frequency components may represent the overall operating trend of the equipment. However, these wavelet coefficients are often mixed with noise and therefore need further processing.To remove noise, we use a non-linear threshold shrinkage function to denoise the multi-level wavelet coefficient set, obtaining an optimized wavelet coefficient group. The non-linear threshold shrinkage function can dynamically adjust the threshold according to the characteristics of the signal, effectively retaining the useful signal while suppressing noise. For example, in the scenario of mechanical equipment monitoring, the optimized wavelet coefficient group after denoising can more clearly display the actual operating state of the equipment, avoiding misjudgment caused by noise interference. Finally, by extracting the feature components from the optimized wavelet coefficient group, we can obtain a multi-level signal feature component set. These feature components not only contain the operating state information of the equipment at different frequencies but also can reveal potential fault patterns, providing a solid foundation for further probability density mapping, anomaly detection, and fault factor analysis. In summary, by performing multi-scale wavelet transform decomposition on the original time-series data of intelligent sensors, we can extract a multi-level feature component set from complex signals. This process not only improves the depth of understanding of the data but also provides strong support for subsequent anomaly recognition and fault diagnosis. For example, in practical applications, when an abnormal fluctuation in a specific frequency band is detected, combined with the results of multi-scale wavelet transform, we can quickly locate the root cause of the problem and take corresponding maintenance measures, thus effectively ensuring the safe and stable operation of the equipment. In this way, the entire intelligent sensor data anomaly recognition system realizes the full-process optimization management from data acquisition to fault resolution, improving the overall performance and reliability of the system.

[0032] In a specific embodiment, the distribution probability density mapping of the multi-level signal feature component set based on the non-parametric kernel density estimation algorithm to obtain a feature component probability distribution map includes: Performing a feature space orthogonal transformation on the multi-level signal feature component set to obtain a multi-dimensional feature representation matrix, and performing a kernel density sampling calculation on the multi-dimensional feature representation matrix to obtain a multi-level signal feature density sequence; Performing a kernel function expansion on the multi-level signal feature density sequence to obtain a kernel function expansion coefficient group, and performing bandwidth selection and optimization on the kernel function expansion coefficient group to obtain an adaptive density estimation sequence; Performing probability density reconstruction on the adaptive density estimation sequence through integral transform technology to obtain an initial probability distribution field, and performing boundary correction calculation on the initial probability distribution field to obtain a corrected probability density map; Performing a multi-scale decomposition operation on the corrected probability density map based on the non-parametric kernel density estimation algorithm to obtain a hierarchical density feature group, and performing a feature space mapping on the hierarchical density feature group to obtain a density feature mapping set; Constructing a probability field based on the density feature mapping set to obtain a multi-dimensional probability density field, and performing a feature fusion operation on the multi-dimensional probability density field to obtain a feature component probability distribution map.

[0033] Specifically, for the process of performing distribution probability density mapping on the multi-level signal feature component set based on the non-parametric kernel density estimation algorithm to obtain the feature component probability distribution map, it is first necessary to understand the basic principle of non-parametric kernel density estimation and its application in complex data analysis. Non-parametric kernel density estimation is a probability density estimation method that does not require assuming the form of data distribution and can flexibly adapt to various complex data structures. In this step, we take the multi-level signal feature component set as the input, and through a series of complex transformations and calculations, finally generate the feature component probability distribution map. First, perform a feature space orthogonal transformation on the multi-level signal feature component set to obtain a multi-dimensional feature representation matrix, and perform kernel density sampling calculation on the multi-dimensional feature representation matrix to obtain a multi-level signal feature density sequence. Feature space orthogonal transformation is a technique that transforms the original features into a new orthogonal coordinate system, aiming to eliminate the correlation between features and simplify subsequent calculations. For example, in the application scenario of monitoring the operating state of large mechanical equipment, assume that we have obtained the multi-level signal feature component set of the equipment vibration signal through multi-scale wavelet transform. Next, by performing a feature space orthogonal transformation on these feature components, they can be transformed into a more easily processed multi-dimensional feature representation matrix. Then, using the kernel density sampling calculation method, sample points are extracted from this matrix, and the probability density value of each sample point is estimated, thereby obtaining a multi-level signal feature density sequence. This step not only helps to simplify the data structure but also provides basic data for subsequent probability density estimation. Then, perform a kernel function expansion on the multi-level signal feature density sequence to obtain a kernel function expansion coefficient group, and perform bandwidth selection and optimization on the kernel function expansion coefficient group to obtain an adaptive density estimation sequence. Kernel function expansion is one of the core steps of kernel density estimation. It constructs a smooth probability density function by expanding the influence range of each sample point to its surrounding area. For example, in the above mechanical equipment monitoring scenario, for each sample point in the feature density sequence, a suitable kernel function (such as Gaussian kernel, Epanechnikov kernel, etc.) is used to perform weighted summation to generate a preliminary probability density estimate. However, to ensure the accuracy of this estimation result, it is also necessary to optimize the bandwidth of the kernel function. The choice of bandwidth directly affects the smoothness of the density estimation: too large a bandwidth may lead to over-smoothing and mask details, while too small a bandwidth may introduce too much noise. Therefore, an adaptive bandwidth selection algorithm is adopted to dynamically adjust the bandwidth size according to the local characteristics of the data to obtain the optimal density estimation sequence. Subsequently, perform probability density reconstruction on the adaptive density estimation sequence through integral transform technology to obtain an initial probability distribution field, and perform boundary correction calculation on the initial probability distribution field to obtain a corrected probability density map. Integral transform technology is used to convert discrete density estimation values into a continuous probability density function, enabling us to more intuitively observe the overall distribution of the data.For example, in the monitoring of mechanical equipment, the adaptive density estimation sequence obtained through the above steps reflects the probability density distribution of each characteristic component. Through integral transformation, these discrete estimated values are reconstructed into a continuous probability distribution field, so as to better display the characteristic distribution of different frequency bands. However, due to the possible existence of boundary effects or irregular shapes in the actual data, the initial probability distribution field often needs to be further corrected. The boundary correction calculation aims to adjust the edge part of the distribution field to make it more in line with the actual situation, so as to obtain the corrected probability density map. Next, based on the non-parametric kernel density estimation algorithm, a multi-scale decomposition operation is performed on the corrected probability density map to obtain a hierarchical density feature group, and a feature space mapping is performed on the hierarchical density feature group to obtain a density feature mapping set. Multi-scale decomposition is a technique for decomposing signals or data into different scale components, which helps to reveal the characteristics of data at different resolutions. For example, in the above application scenario, through the multi-scale decomposition operation, different levels of density feature groups can be extracted from the corrected probability density map. These feature groups respectively reflect the distribution of data at coarse and fine scales, providing rich information for further analysis. Then, by performing feature space mapping on these hierarchical density feature groups, they are converted into a form that is easier to understand and process, so as to obtain a density feature mapping set. This step not only improves the depth of understanding of the data, but also lays a foundation for the subsequent construction of the probability field. Finally, based on the density feature mapping set, a probability field is constructed to obtain a multi-dimensional probability density field, and a feature fusion operation is performed on the multi-dimensional probability density field to obtain a characteristic component probability distribution map. Probability field construction is the process of combining multiple univariate probability density functions into a multi-dimensional probability density field, and its purpose is to comprehensively describe the overall distribution of data. For example, in the application scenario of monitoring the operating state of large mechanical equipment, by combining the density feature mapping sets of different frequency bands, a multi-dimensional probability density field reflecting the overall operating state of the equipment can be constructed. This multi-dimensional probability density field not only shows the characteristic distribution of each frequency band, but also reveals the interaction and influence between them. In order to further improve the analysis effect, a feature fusion operation also needs to be performed on the multi-dimensional probability density field. The feature fusion operation generates a comprehensive characteristic component probability distribution map by integrating information from different dimensions. This step not only enhances the depth of understanding of the data, but also provides a reliable basis for subsequent anomaly detection and fault factor analysis. In summary, the process of obtaining the characteristic component probability distribution map by performing distribution probability density mapping on the multi-level signal characteristic component set based on the non-parametric kernel density estimation algorithm involves multiple complex steps and techniques. From the feature space orthogonal transformation to the multi-scale decomposition, and then to the probability field construction and the feature fusion operation, each step provides important support for the final result.For example, in practical applications, when abnormal fluctuations in a specific frequency band are detected, by combining the results of multi-scale wavelet transform and non-parametric kernel density estimation, we can quickly locate the root cause of the problem and take corresponding maintenance measures, thereby effectively ensuring the safe and stable operation of the equipment. In this way, the entire intelligent sensor data anomaly recognition system realizes the full-process optimized management from data acquisition to fault resolution, improving the overall performance and reliability of the system. This method not only enhances the depth of understanding of the data but also provides strong support for subsequent anomaly detection and fault diagnosis, ensuring the efficient operation and long-term stability of the system.

[0034] In a specific embodiment, the extraction of the abnormal region boundary of the feature component probability distribution map through the adaptive threshold segmentation algorithm to obtain the candidate abnormal event boundary sequence includes: Performing gradient calculation and direction analysis on the feature component probability distribution map to obtain a probability distribution gradient field, and performing local extreme value detection on the probability distribution gradient field to obtain an initial threshold candidate set; Performing probability distribution region segmentation on the initial threshold candidate set through the adaptive threshold segmentation algorithm to obtain an initial segmentation region group, and calculating the boundary curvature of the initial segmentation region group to obtain a region boundary feature sequence; Performing boundary expansion on the region boundary feature sequence through a preset region growing technique to obtain an extended boundary contour set, and performing morphological processing on the extended boundary contour set to obtain an optimized boundary sequence; Performing topological structure analysis on the optimized boundary sequence to obtain a boundary topological relationship graph, and performing region merging calculation on the boundary topological relationship graph to obtain a candidate abnormal region set; Performing temporal correlation analysis on the candidate abnormal region set to obtain a temporal correlation feature group, and performing boundary sequence screening and fusion on the temporal correlation feature group to obtain a candidate abnormal event boundary sequence; wherein, the candidate abnormal event boundary sequence includes temporal abnormal interval boundary markers, abnormal event topological contour features, abnormal region density gradient vectors, and abnormal boundary curvature feature descriptors.

[0035] Specifically, in the process of extracting the boundary of the abnormal region from the probability distribution map of feature components through the adaptive threshold segmentation algorithm to obtain the candidate abnormal event boundary sequence, it is first necessary to understand the basic principle of the adaptive threshold segmentation algorithm and its application in complex data analysis. Adaptive threshold segmentation is a technique that automatically adjusts the threshold according to the local characteristics of an image or data field, and can effectively identify abnormal regions in the data. In this process, we use the probability distribution map of feature components as the input, and through a series of complex calculations and transformations, finally generate the candidate abnormal event boundary sequence. At the beginning, gradient calculation and direction analysis are performed on the probability distribution map of feature components to obtain the probability distribution gradient field, and local extreme value detection is performed on the probability distribution gradient field to obtain the initial threshold candidate set. Gradient calculation and direction analysis are important steps to identify the trend of data changes, which helps us understand the change rate and direction of different parts of the data. For example, in the application scenario of large mechanical equipment monitoring, assume that we have obtained the probability distribution map of the feature components of the equipment operating state. By performing gradient calculation on this map, we can obtain a probability distribution gradient field that reflects the change rate of the feature density in each frequency band. Then, using the local extreme value detection technology, find those points with significant change characteristics from this gradient field. These points may represent the boundaries of potential abnormal regions, thus forming the initial threshold candidate set. This process not only reveals the key change points in the data but also provides a basis for subsequent threshold segmentation. Next, probability distribution region segmentation is performed on the initial threshold candidate set through the adaptive threshold segmentation algorithm to obtain the initial segmentation region group, and boundary curvature calculation is performed on the initial segmentation region group to obtain the region boundary feature sequence. The adaptive threshold segmentation algorithm dynamically adjusts the threshold according to the local characteristics of the data, making the segmentation result more accurate. For example, in the above application scenario, for each initial threshold candidate point, use an appropriate adaptive threshold segmentation method to divide the surrounding area into different parts to form the initial segmentation region group. Then, calculate the curvature of the boundaries of these segmented regions to quantify their geometric shape characteristics and obtain the region boundary feature sequence. This step not only helps to distinguish normal and abnormal regions but also provides the necessary geometric information for subsequent boundary expansion. Subsequently, boundary expansion is performed on the region boundary feature sequence through the preset region growing technique to obtain the extended boundary contour set, and morphological processing is performed on the extended boundary contour set to obtain the optimized boundary sequence. The region growing technique is a method based on seed points that gradually expands, aiming to connect similar regions and eliminate isolated noise points. For example, in mechanical equipment monitoring, by selecting some representative boundary feature points as seed points and using the region growing technique to gradually expand the boundary until the preset conditions are met, the extended boundary contour set is formed.However, in order to ensure the authenticity and integrity of the boundaries, morphological operations such as opening and closing need to be performed on these extended boundaries to remove small interfering objects and fill holes, thereby obtaining an optimized boundary sequence. This step not only improves the accuracy of the boundaries but also lays a foundation for subsequent topological structure analysis. Then, topological structure analysis is performed on the optimized boundary sequence to obtain a boundary topological relationship graph, and regional merging calculations are performed on the boundary topological relationship graph to obtain a candidate abnormal region set. Topological structure analysis is used to study the spatial relationships between objects, which helps us understand the connection methods and interactions between different regions. For example, in the above application scenario, by performing topological structure analysis on the optimized boundary sequence, a boundary topological relationship graph describing the relationships between regions can be constructed. Then, using the regional merging calculation method, those regions that are physically connected or functionally related are merged to form a candidate abnormal region set. This step not only helps to simplify the data structure but also improves the depth of understanding of abnormal phenomena. Finally, temporal correlation analysis is performed on the candidate abnormal region set to obtain a temporal correlation feature group, and boundary sequence screening and fusion are performed on the temporal correlation feature group to obtain a candidate abnormal event boundary sequence. Temporal correlation analysis is used to study the relationships between data at different time points, which helps us discover potential abnormal patterns and trends. For example, in the application scenario of monitoring the operating status of large mechanical equipment, by performing temporal correlation analysis on the candidate abnormal region set, those abnormal patterns that continuously appear in the time dimension can be found to form a temporal correlation feature group. Then, by performing boundary sequence screening and fusion on these feature groups, irrelevant information is removed, and the most valuable part is retained, ultimately obtaining a candidate abnormal event boundary sequence. This sequence includes important information such as temporal abnormal interval boundary markers, abnormal event topological contour features, abnormal region density gradient vectors, and abnormal boundary curvature feature descriptors, which together constitute a data set that comprehensively describes abnormal events. This method not only improves the recognition accuracy of abnormal phenomena but also provides strong support for subsequent fault diagnosis and maintenance strategy formulation. In summary, the process of obtaining a candidate abnormal event boundary sequence by extracting the abnormal region boundary from the probability distribution map of feature components through an adaptive threshold segmentation algorithm involves multiple complex steps and techniques. From gradient calculation and direction analysis to topological structure analysis and then to temporal correlation analysis, each step provides important support for the final result. For example, in practical applications, when an abnormal fluctuation is detected within a specific time period, combined with the results of multi-scale wavelet transform and non-parametric kernel density estimation, we can quickly locate the root cause of the problem and take corresponding maintenance measures, thereby effectively ensuring the safe and stable operation of the equipment. In this way, the entire intelligent sensor data abnormal recognition system realizes the full-process optimization management from data acquisition to fault resolution, improving the overall performance and reliability of the system.This method not only improves the recognition accuracy of abnormal phenomena, but also provides strong support for subsequent fault diagnosis and maintenance strategy formulation, ensuring the efficient operation and long-term stability of the system.

[0036] In a specific embodiment, the topological correlation analysis of the candidate abnormal event boundary sequence by the graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph includes: Performing multi-dimensional feature encoding transformation on the candidate abnormal event boundary sequence to obtain a boundary event feature vector group, and constructing a graph structure for the boundary event feature vector group to obtain an initial abnormal event relationship graph; Performing Laplacian matrix decomposition on the initial abnormal event relationship graph through a preset spectral domain transformation technique to obtain a spectral feature descriptor set, and performing multi-level convolutional operations on the spectral feature descriptor set to obtain a hierarchical graph convolutional feature group; Through the graph convolutional network, based on the hierarchical graph convolutional feature group, perform cross-domain correlation mapping on a preset abnormal event to obtain an abnormal pattern correlation matrix, and perform graph attention calculation on the abnormal pattern correlation matrix to obtain a weighted abnormal relationship network; Performing dynamic graph evolution analysis on the weighted abnormal relationship network to obtain a set of abnormal propagation paths, and performing multi-scale fusion processing on the set of abnormal propagation paths to obtain a multi-level abnormal relationship topology graph; Performing spatio-temporal correlation embedding on the multi-level abnormal relationship topology graph to obtain an abnormal event correlation embedding vector space, and performing abnormal pattern extraction on the abnormal event correlation embedding vector space to obtain a multi-dimensional abnormal pattern relationship graph; wherein, the multi-dimensional abnormal pattern relationship graph includes abnormal event type correlation metrics and abnormal pattern evolution characteristics.

[0037] Specifically, in the process of performing topological correlation analysis on the candidate abnormal event boundary sequence through the graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph, it is first necessary to understand the basic principle of the graph convolutional network (GCN) and its application in complex data analysis. The graph convolutional network is a deep learning model specifically designed to process graph-structured data, which can effectively capture the topological relationships and feature information between nodes. In this process, we use the candidate abnormal event boundary sequence as the input, and through a series of complex calculations and transformations, finally generate a multi-dimensional abnormal pattern relationship graph. First, perform multi-dimensional feature encoding transformation on the candidate abnormal event boundary sequence to obtain a boundary event feature vector group, and construct a graph structure for the boundary event feature vector group to obtain an initial abnormal event relationship graph. Multi-dimensional feature encoding transformation is a key step in converting the original data into a form suitable for processing by the graph convolutional network. For example, in the application scenario of large mechanical equipment monitoring, assume that we have obtained the candidate abnormal event boundary sequence of the equipment operation status. By performing multi-dimensional feature encoding transformation on these boundary sequences, multi-dimensional features such as time, space, and frequency of each abnormal event can be extracted to form a boundary event feature vector group. Then, based on these feature vectors, use graph structure construction technology to represent each abnormal event as a node in the graph, and define edges according to their spatio-temporal correlation, thereby constructing an initial abnormal event relationship graph. This step not only helps to convert the original data into a form suitable for processing by the graph convolutional network, but also provides a basis for subsequent graph spectrum feature extraction. Next, perform Laplacian matrix decomposition on the initial abnormal event relationship graph through a preset spectral domain transformation technology to obtain a graph spectrum feature representation set, and perform multi-level convolutional operations on the graph spectrum feature representation set to obtain a hierarchical graph convolutional feature group. The spectral domain transformation technology is an important part of the graph convolutional network. It maps graph-structured data to the frequency domain space through Laplacian matrix decomposition, thereby achieving efficient feature extraction. For example, in the above application scenario, by performing Laplacian matrix decomposition on the initial abnormal event relationship graph, its corresponding graph spectrum feature representation set can be obtained. Then, using the multi-level convolutional operation method, gradually extract graph convolutional features at different levels to form a hierarchical graph convolutional feature group. This step not only improves the depth of understanding of graph-structured data, but also provides rich feature information for subsequent cross-domain correlation mapping. Subsequently, through the graph convolutional network, perform cross-domain correlation mapping on the preset abnormal events based on the hierarchical graph convolutional feature group to obtain an abnormal pattern correlation matrix, and perform graph attention calculation on the abnormal pattern correlation matrix to obtain a weighted abnormal relationship network. Cross-domain correlation mapping aims to discover the potential connections between different abnormal events, while the graph attention mechanism further enhances the accuracy of this correlation. For example, in mechanical equipment monitoring, by performing cross-domain correlation mapping on the hierarchical graph convolutional feature group, it can be identified which abnormal events have significant correlations or causal relationships, forming an abnormal pattern correlation matrix.Then, using the graph attention calculation method, the weights are dynamically adjusted according to the importance of the nodes to obtain a weighted abnormal relationship network. This step not only reveals the internal connections between abnormal events but also lays the foundation for subsequent dynamic graph evolution analysis. Next, dynamic graph evolution analysis is performed on the weighted abnormal relationship network to obtain a set of abnormal propagation paths, and multi-scale fusion processing is performed on the set of abnormal propagation paths to obtain a multi-level abnormal relationship topology graph. Dynamic graph evolution analysis is used to study the propagation paths and influence ranges of abnormal events over time. For example, in the above application scenario, by performing dynamic graph evolution analysis on the weighted abnormal relationship network, the propagation paths of abnormal events can be traced, and those abnormal patterns that continuously appear in time and space can be identified to form a set of abnormal propagation paths. Then, using the multi-scale fusion processing method, the abnormal propagation paths at different scales are integrated to obtain a multi-level abnormal relationship topology graph. This step not only improves the depth of understanding of abnormal phenomena but also provides necessary information for subsequent spatio-temporal correlation embedding. Finally, spatio-temporal correlation embedding is performed on the multi-level abnormal relationship topology graph to obtain an abnormal event correlation embedding vector space, and abnormal pattern extraction is performed on the abnormal event correlation embedding vector space to obtain a multi-dimensional abnormal pattern relationship graph; where the multi-dimensional abnormal pattern relationship graph includes abnormal event type correlation metrics and abnormal pattern evolution characteristics. Spatio-temporal correlation embedding aims to map the relationship graph of abnormal events into a low-dimensional vector space for better understanding and analysis. For example, in the application scenario of monitoring the operating status of large mechanical equipment, by performing spatio-temporal correlation embedding on the multi-level abnormal relationship topology graph, each abnormal event and its mutual relationship can be represented as points in a vector space to form an abnormal event correlation embedding vector space. Then, by performing abnormal pattern extraction on this vector space, different types of abnormal events and their evolution characteristics can be identified, and finally, a multi-dimensional abnormal pattern relationship graph can be obtained. This graph not only shows the specific positions and characteristics of each abnormal event but also reveals their interactions and influences, providing strong support for fault factor analysis and maintenance strategy formulation. In summary, the process of obtaining a multi-dimensional abnormal pattern relationship graph through topological correlation analysis of the candidate abnormal event boundary sequence by a graph convolutional network involves multiple complex steps and technologies. From multi-dimensional feature encoding conversion to dynamic graph evolution analysis and then to spatio-temporal correlation embedding, each step provides important support for the final result. For example, in practical applications, when abnormal fluctuations are detected within a specific time period, combined with the results of multi-scale wavelet transform, non-parametric kernel density estimation, and adaptive threshold segmentation algorithms, we can quickly locate the root cause of the problem and take corresponding maintenance measures, thus effectively ensuring the safe and stable operation of the equipment. In this way, the entire intelligent sensor data abnormal recognition system realizes the full-process optimization management from data acquisition to fault resolution, improving the overall performance and reliability of the system.This method not only improves the recognition accuracy of abnormal phenomena, but also provides strong support for subsequent fault diagnosis and maintenance strategy formulation, ensuring the efficient operation and long-term stability of the system. In addition, through in-depth analysis of the graph convolutional network, deep abnormal patterns hidden behind the data can be discovered, further enhancing the intelligent level of the system.

[0038] In a specific embodiment, the multi-dimensional feature encoding conversion of the candidate abnormal event boundary sequence to obtain a boundary event feature vector group includes: Performing time-frequency domain decomposition and reconstruction on the candidate abnormal event boundary sequence to obtain a multi-domain boundary feature matrix, and performing non-linear projection conversion on the multi-domain boundary feature matrix to obtain a dimensionality-reduced feature space; Performing local sensitive hashing encoding on the dimensionality-reduced feature space to obtain a feature hashing encoding set, and performing multi-scale redundancy analysis on the feature hashing encoding set to obtain a compressed feature encoding table; Performing high-order relationship extraction on the compressed feature encoding table through tensor decomposition technology to obtain a multi-modal feature interaction tensor, and performing nuclear norm minimization processing on the multi-modal feature interaction tensor to obtain a sparse feature table collection; Performing manifold learning and embedding on the sparse feature table collection to obtain a low-dimensional embedding space, and performing similarity measurement and clustering on the low-dimensional embedding space to obtain a boundary event feature vector group; wherein, the boundary event feature vector group includes abnormal event category encoding and spatio-temporal correlation strength characterization.

[0039] Specifically, in the process of converting the candidate abnormal event boundary sequence into a multi-dimensional feature code to obtain a boundary event feature vector group, it is first necessary to clarify the principle of each step and its role in the overall process. This process starts with time-frequency domain decomposition and reconstruction, and finally generates a boundary event feature vector group containing abnormal event category coding and spatiotemporal correlation strength representation through a series of complex transformations and calculations. First, the candidate abnormal event boundary sequence is decomposed and reconstructed in the time-frequency domain to obtain a multi-domain boundary feature matrix, and the multi-domain boundary feature matrix is nonlinearly projected to obtain a reduced-dimensional feature space. This step aims to extract the different frequency components hidden in the original data and their laws of change over time, thereby providing a rich information basis for subsequent analysis. For example, in the application scenario of monitoring the operating status of large-scale mechanical equipment, by decomposing and reconstructing the data collected by the equipment sensor in the time-frequency domain, a multi-domain boundary feature matrix reflecting the vibration mode of the equipment under different working conditions can be obtained. Then, these high-dimensional data are mapped to a low-dimensional space using nonlinear projection transformation methods, such as principal component analysis (PCA) or t-distributed neighbor embedding algorithm (t-SNE), to achieve data dimensionality reduction. This process not only retains the main features of the original data, but also reduces the complexity of subsequent calculations, providing a more concise and effective input for subsequent steps. Then, the reduced-dimensional feature space is locally sensitive hashed to obtain a feature hash code set, and the feature hash code set is subjected to multi-scale redundancy analysis to obtain a compressed feature code table. Locally sensitive hash coding is an efficient data indexing technology that can quickly retrieve similar data points. In this process, we first convert the reduced-dimensional feature space into a binary coding form through local sensitive hash coding to form a feature hash code set. Subsequently, in order to further reduce data redundancy and improve storage efficiency, these feature hash code sets are subjected to multi-scale redundancy analysis. This step can be implemented in a variety of ways, such as identifying and removing duplicate information based on statistical methods or machine learning models, thereby obtaining a more compact compressed feature code table. This not only reduces data storage requirements, but also speeds up subsequent processing steps. Next, the compressed feature code table is subjected to high-order relationship extraction through tensor decomposition technology to obtain a multimodal feature interaction tensor, and the multimodal feature interaction tensor is subjected to nuclear norm minimization processing to obtain a sparse feature representation set. Tensor decomposition technology is an important means of processing multi-dimensional data. It can mine potential structural information from complex high-dimensional data. In this application scenario, by applying tensor decomposition technology to the compressed feature encoding table, we can reveal the deep correlation between different modal features and construct a multi-modal feature interaction tensor. In order to simplify this tensor representation and make it easier to understand and use, we need to minimize the nuclear norm to obtain a sparse feature representation set.This step helps to highlight those features that best represent the essential characteristics of the data, while filtering out the influence of noise and other irrelevant factors. Finally, manifold learning and embedding are performed on the sparse feature table collection to obtain a low-dimensional embedding space, and similarity measurement and clustering are performed on the low-dimensional embedding space to obtain a set of boundary event feature vectors; wherein, the set of boundary event feature vectors includes abnormal event category codes and spatio-temporal correlation strength characterizations. Manifold learning is a technique for discovering the intrinsic geometric structure of data. Through this method, we can map the sparse feature table collection into a lower-dimensional space, that is, a low-dimensional embedding space. In this new representation, similar data points will be clustered together, while different data points will maintain a certain distance. For the above application scenarios, this means that abnormal events with similar fault patterns will be close to each other in the embedding space, while different fault types will be effectively distinguished. Based on such a low-dimensional embedding space, we can further perform similarity measurement and clustering analysis, such as using the K-means clustering algorithm or hierarchical clustering method, to identify the specific categories of each abnormal event and the spatio-temporal correlation strength between them, and finally form a set of boundary event feature vectors. Such a set of feature vectors not only contains rich information about the types of abnormal events, but also reflects the complex spatio-temporal relationships between them, providing an important basis for subsequent fault diagnosis and maintenance strategy formulation. In summary, through the process of multi-dimensional feature encoding conversion of the candidate abnormal event boundary sequence, from time-frequency domain decomposition and reconstruction to manifold learning and embedding, each step is closely connected, jointly constructing a comprehensive and detailed analysis framework. For example, in actual operation, when we face a set of monitoring data from large-scale mechanical equipment, through the above steps, we can accurately locate the possible problems of the equipment and understand how these problems evolve over time and space distribution. This method not only improves the detection accuracy of abnormal phenomena, but also makes it possible to deeply explore the reasons behind them, making preventive maintenance more scientific and efficient. In addition, the whole process emphasizes the importance of data analysis, and through the combination of a variety of advanced algorithms and technologies, it realizes the effective transformation from raw data to valuable information.

[0040] In a specific embodiment, the fault factor analysis of the intelligent sensor based on the multi-dimensional abnormal pattern relationship graph to obtain the abnormal factors of the intelligent sensor includes: Perform abnormal event topological feature decomposition on the multi-dimensional abnormal pattern relationship graph to obtain a set of abnormal event topological feature vectors, and perform multi-scale singular value decomposition on the set of abnormal event topological feature vectors to obtain a sensor abnormal feature matrix; Perform multi-dimensional feature space mapping on the preset sensor states based on the sensor anomaly feature matrix to obtain a sensor state feature space representation, and perform Gaussian mixture clustering analysis on the sensor state feature space representation to obtain a set of sensor state clustering clusters; Perform dynamic Bayesian network modeling on the set of sensor state clustering clusters to obtain a sensor state transition probability matrix, and perform hidden Markov model parameter estimation on the sensor state transition probability matrix to obtain a sensor state hidden variable sequence; Perform multivariate statistical process control analysis on the sensor state hidden variable sequence to obtain a sensor fault control chart, and perform fault feature extraction and quantification on the sensor fault control chart to obtain a sensor anomaly factor.

[0041] Specifically, for the process of performing fault factor analysis on intelligent sensors based on the multi-dimensional anomaly pattern relationship graph to obtain the anomaly factors of intelligent sensors, it is first necessary to understand the basic principles of the multi-dimensional anomaly pattern relationship graph and its application in complex data analysis. This process involves a series of complex transformations and calculations, and finally identifies the specific factors that cause sensor anomalies. The following is a detailed explanation of each step of this process and its actual application scenarios. First, perform anomaly event topological feature decomposition on the multi-dimensional anomaly pattern relationship graph to obtain an anomaly event topological feature vector set, and perform multi-scale singular value decomposition on the anomaly event topological feature vector set to obtain a sensor anomaly feature matrix. Anomaly event topological feature decomposition aims to extract features that reflect the topological relationships between different anomaly events from the multi-dimensional anomaly pattern relationship graph. For example, in the application scenario of monitoring the operating status of large mechanical equipment, assume that we have obtained the multi-dimensional anomaly pattern relationship graph of the equipment vibration signal. By performing topological feature decomposition on this graph, a feature vector set containing the topological relationships between various anomaly events can be obtained. Then, using the multi-scale singular value decomposition (SVD) method, these feature vectors are further decomposed into components at different scales, thus generating a sensor anomaly feature matrix. This step not only helps to reveal the deep-seated correlations between anomaly events but also provides basic data for subsequent state mapping. Next, perform multi-dimensional feature space mapping on the preset sensor states based on the sensor anomaly feature matrix to obtain a sensor state feature space representation, and perform Gaussian mixture clustering analysis on the sensor state feature space representation to obtain a set of sensor state clustering clusters. Multi-dimensional feature space mapping is to transform the anomaly features of the sensor into a new feature space to better understand and analyze its state. For example, in the above application scenario, by performing multi-dimensional feature space mapping on the sensor anomaly feature matrix, the different states of the sensor can be represented as points in a feature space, forming a sensor state feature space representation. Then, using the Gaussian mixture clustering analysis method, these feature points are divided into several clustering cluster sets according to their distribution. This step not only helps to identify the typical patterns shown by the sensor in different states but also provides support for further state transition probability modeling. Subsequently, perform dynamic Bayesian network modeling on the set of sensor state clustering clusters to obtain a sensor state transition probability matrix, and perform hidden Markov model parameter estimation on the sensor state transition probability matrix to obtain a sequence of sensor state hidden variables. Dynamic Bayesian network is a probabilistic graphical model used to model time series data and can capture the dynamic characteristics of system states changing over time. For example, in the above application scenario, by performing dynamic Bayesian network modeling on the set of sensor state clustering clusters, a transition probability matrix describing the evolution of the sensor state over time can be constructed. Then, using the hidden Markov model (HMM) parameter estimation method, a sequence of hidden state variables is extracted from this transition probability matrix.This step not only reveals the changing pattern of the sensor state but also provides an important basis for subsequent fault control analysis. Next, perform multivariate statistical process control analysis on the sequence of the hidden variables of the sensor state to obtain a sensor fault control chart, and extract and quantify the fault characteristics of the sensor fault control chart to obtain sensor anomaly factors. Multivariate statistical process control (MSPC) is a technique used to monitor and detect abnormal changes in the production process, which can help identify potential fault modes. For example, in the above application scenario, by performing multivariate statistical process control analysis on the sequence of the hidden variables of the sensor state, a fault control chart reflecting the operating state of the sensor can be generated. Then, using the method of fault feature extraction and quantification, those features deviating from the normal range are identified from this control chart, thereby obtaining the sensor anomaly factors. This step not only improves the recognition accuracy of fault phenomena but also provides a scientific basis for formulating effective maintenance strategies. To sum up, the process of fault factor analysis for intelligent sensors based on the multi-dimensional abnormal pattern relationship diagram involves multiple complex steps and techniques. From the decomposition of the topological features of abnormal events to the dynamic Bayesian network modeling, and then to the multivariate statistical process control analysis, each step provides important support for the final result. For example, in actual operation, when we face a set of monitoring data from large-scale mechanical equipment, through the above steps, we can accurately locate the possible problems of the equipment and understand how these problems evolve over time and space distribution. Specifically, in the application scenario of monitoring the operating state of large-scale mechanical equipment, assume that the vibration sensor data of a certain device shows abnormal fluctuations. Through the analysis of the multi-dimensional abnormal pattern relationship diagram, we find that the energy in certain frequency bands is significantly higher than other parts, indicating that there may be local faults or wear problems. Further, through the decomposition of the topological features and multi-scale singular value decomposition of these abnormal events, we obtain the sensor anomaly feature matrix. Then, through multi-dimensional feature space mapping and Gaussian mixture clustering analysis, we divide the different states of the sensor into several typical clustering clusters. Next, using dynamic Bayesian network modeling and hidden Markov model parameter estimation, we construct the sensor state transition probability matrix and extract the sequence of hidden state variables from it. Finally, through the multivariate statistical process control analysis of the sequence of hidden variables, we generate a fault control chart and extract specific fault characteristics from it, such as the abnormal fluctuations of high-frequency vibration signals, indicating that there are problems of looseness or wear in mechanical components. This detailed fault factor analysis not only improves the detection accuracy of abnormal phenomena but also makes it possible to deeply explore the reasons behind them, thus making preventive maintenance more scientific and efficient. The whole process emphasizes the importance of data analysis, and through the combination of a variety of advanced algorithms and techniques, the effective transformation from raw data to valuable information is achieved.

[0042] The above describes the intelligent sensor data anomaly recognition method in the embodiments of the present invention. Next, the intelligent sensor data anomaly recognition device in the embodiments of the present invention will be described. Please refer to Figure 2 , one embodiment of the intelligent sensor data anomaly recognition device in the embodiments of the present invention includes: A decomposition module 21, configured to perform multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set; A mapping module 22, configured to perform distribution probability density mapping on the multi-level signal feature component set based on a non-parametric kernel density estimation algorithm to obtain a feature component probability distribution map; An extraction module 23, configured to, when the feature component probability distribution map indicates that the original time-series data is abnormal, extract the boundary of the abnormal region from the feature component probability distribution map through an adaptive threshold segmentation algorithm to obtain a candidate abnormal event boundary sequence; A first analysis module 24, configured to perform topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph; A second analysis module 25, configured to perform fault factor analysis on the intelligent sensor based on the multi-dimensional abnormal pattern relationship graph to obtain the abnormal factors of the intelligent sensor, and obtain corresponding maintenance strategies based on the abnormal factors.

[0043] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to that described in the above method embodiment, and details will not be elaborated here.

[0044] Refer to Figure 3 , the embodiments of the present invention also provide a computer device, and its internal structure can be as Figure 3 shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface, and a database connected through a system bus. Among them, the processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0045] Those skilled in the art can understand that Figure 3 the structure shown in

[0046] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0047] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium provided by the present invention and used in the embodiments can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0048] It should be noted that in this article, the terms "include", "comprise", or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, apparatus, article, or method including a series of elements not only includes those elements but also includes other elements not explicitly listed, or further includes elements inherent to such a process, apparatus, article, or method. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, apparatus, article, or method including that element.

[0049] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structural or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, are equally included in the patent protection scope of the present invention.

Claims

1. An intelligent sensor data anomaly recognition method, characterized in that, Including the following steps: Perform multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set; Based on the non-parametric kernel density estimation algorithm, perform distribution probability density mapping on the multi-level signal feature component set to obtain a feature component probability distribution map; When the feature component probability distribution map indicates that the original time-series data is abnormal, use the adaptive threshold segmentation algorithm to extract the abnormal region boundary of the feature component probability distribution map to obtain a candidate abnormal event boundary sequence; Perform topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph; Based on the multi-dimensional abnormal pattern relationship graph, perform fault factor analysis on the intelligent sensor to obtain the abnormal factors of the intelligent sensor, and based on the abnormal factors, obtain the corresponding maintenance strategy.

2. The intelligent sensor data anomaly recognition method according to claim 1, wherein The performing multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set includes: Perform spectrum analysis preprocessing on the original time-series data collected by the intelligent sensor to obtain a frequency-domain signal distribution map, and perform multi-resolution adaptive segmentation on the frequency-domain signal distribution map to obtain a frequency band division matrix; wherein, the frequency band division matrix includes the main frequency band interval boundary value and the secondary frequency band interval energy distribution characteristics; Based on the frequency band division matrix, perform discrete wavelet packet transform on the original time-series data to obtain a multi-level wavelet coefficient set, and perform denoising processing on the multi-level wavelet coefficient set through a non-linear threshold shrinkage function to obtain an optimized wavelet coefficient group; wherein, the optimized wavelet coefficient group includes high-frequency change feature components and low-frequency trend feature components; Perform feature component extraction on the optimized wavelet coefficient group to obtain a multi-level signal feature component set.

3. The intelligent sensor data anomaly recognition method according to claim 1, wherein The performing distribution probability density mapping on the multi-level signal feature component set based on the non-parametric kernel density estimation algorithm to obtain a feature component probability distribution map includes: Perform feature space orthogonal transformation on the multi-level signal feature component set to obtain a multi-dimensional feature representation matrix, and perform kernel density sampling calculation on the multi-dimensional feature representation matrix to obtain a multi-level signal feature density sequence; Perform kernel function expansion on the multi-level signal feature density sequence to obtain a kernel function expansion coefficient group, and perform bandwidth selection and optimization on the kernel function expansion coefficient group to obtain an adaptive density estimation sequence; Perform probability density reconstruction on the adaptive density estimation sequence through integral transform technology to obtain an initial probability distribution field, and perform boundary correction calculation on the initial probability distribution field to obtain a corrected probability density map; Based on the non-parametric kernel density estimation algorithm, perform multi-scale decomposition operation on the corrected probability density map to obtain a hierarchical density feature group, and perform feature space mapping on the hierarchical density feature group to obtain a density feature mapping set; Based on the density feature mapping set, perform probability field construction to obtain a multi-dimensional probability density field, and perform feature fusion operation on the multi-dimensional probability density field to obtain a feature component probability distribution map.

4. The intelligent sensor data anomaly recognition method according to claim 1, wherein Performing abnormal region boundary extraction on the probability distribution map of the feature components through an adaptive threshold segmentation algorithm to obtain a candidate abnormal event boundary sequence, including: Calculating the gradient and direction analysis of the probability distribution map of the feature components to obtain a probability distribution gradient field, and performing local extreme value detection on the probability distribution gradient field to obtain an initial threshold candidate set; Performing probability distribution region segmentation on the initial threshold candidate set through an adaptive threshold segmentation algorithm to obtain an initial segmentation region group, and calculating the boundary curvature of the initial segmentation region group to obtain a region boundary feature sequence; Performing boundary expansion on the region boundary feature sequence through a preset region growing technique to obtain an extended boundary contour set, and performing morphological processing on the extended boundary contour set to obtain an optimized boundary sequence; Performing topological structure analysis on the optimized boundary sequence to obtain a boundary topological relationship graph, and performing region merging calculation on the boundary topological relationship graph to obtain a candidate abnormal region set; Performing temporal correlation analysis on the candidate abnormal region set to obtain a temporal correlation feature group, and performing boundary sequence screening and fusion on the temporal correlation feature group to obtain a candidate abnormal event boundary sequence; wherein, the candidate abnormal event boundary sequence includes temporal abnormal interval boundary markers, abnormal event topological contour features, abnormal region density gradient vectors, and abnormal boundary curvature feature descriptors.

5. The intelligent sensor data anomaly recognition method according to claim 1, wherein, Performing topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph, including: Performing multi-dimensional feature encoding conversion on the candidate abnormal event boundary sequence to obtain a boundary event feature vector group, and constructing a graph structure for the boundary event feature vector group to obtain an initial abnormal event relationship graph; Performing Laplacian matrix decomposition on the initial abnormal event relationship graph through a preset spectral domain transformation technique to obtain a spectral feature descriptor set, and performing multi-level convolutional operations on the spectral feature descriptor set to obtain a hierarchical graph convolutional feature group; Through a graph convolutional network, performing cross-domain correlation mapping on a preset abnormal event based on the hierarchical graph convolutional feature group to obtain an abnormal pattern correlation matrix, and performing graph attention calculation on the abnormal pattern correlation matrix to obtain a weighted abnormal relationship network; Performing dynamic graph evolution analysis on the weighted abnormal relationship network to obtain a set of abnormal propagation paths, and performing multi-scale fusion processing on the set of abnormal propagation paths to obtain a multi-level abnormal relationship topology graph; Performing spatio-temporal correlation embedding on the multi-level abnormal relationship topology graph to obtain an abnormal event correlation embedding vector space, and performing abnormal pattern extraction on the abnormal event correlation embedding vector space to obtain a multi-dimensional abnormal pattern relationship graph; wherein, the multi-dimensional abnormal pattern relationship graph includes abnormal event type correlation metrics and abnormal pattern evolution features.

6. The intelligent sensor data anomaly recognition method according to claim 5, wherein, Performing multi-dimensional feature encoding conversion on the candidate abnormal event boundary sequence to obtain a boundary event feature vector group, including: Perform time-frequency domain decomposition and reconstruction on the candidate abnormal event boundary sequence to obtain a multi-domain boundary feature matrix, and perform non-linear projection transformation on the multi-domain boundary feature matrix to obtain a reduced-dimensional feature space; Perform local sensitive hashing coding on the reduced-dimensional feature space to obtain a feature hashing coding set, and perform multi-scale redundancy analysis on the feature hashing coding set to obtain a compressed feature coding table; Extract high-order relationships from the compressed feature coding table through tensor decomposition technology to obtain a multi-modal feature interaction tensor, and perform nuclear norm minimization processing on the multi-modal feature interaction tensor to obtain a sparse feature table collection; Perform manifold learning and embedding on the sparse feature table collection to obtain a low-dimensional embedding space, and perform similarity measurement and clustering on the low-dimensional embedding space to obtain a boundary event feature vector group; wherein, the boundary event feature vector group includes abnormal event category coding and spatio-temporal correlation strength characterization.

7. The intelligent sensor data anomaly recognition method according to claim 1, wherein Perform fault factor analysis on the intelligent sensor based on the multi-dimensional abnormal pattern relationship graph to obtain abnormal factors of the intelligent sensor, including: Perform abnormal event topological feature decomposition on the multi-dimensional abnormal pattern relationship graph to obtain an abnormal event topological feature vector set, and perform multi-scale singular value decomposition on the abnormal event topological feature vector set to obtain a sensor abnormal feature matrix; Perform multi-dimensional feature space mapping on a preset sensor state based on the sensor abnormal feature matrix to obtain a sensor state feature space representation, and perform Gaussian mixture clustering analysis on the sensor state feature space representation to obtain a sensor state clustering cluster set; Perform dynamic Bayesian network modeling on the sensor state clustering cluster set to obtain a sensor state transition probability matrix, and perform hidden Markov model parameter estimation on the sensor state transition probability matrix to obtain a sensor state hidden variable sequence; Perform multi-variable statistical process control analysis on the sensor state hidden variable sequence to obtain a sensor fault control chart, and perform fault feature extraction and quantification on the sensor fault control chart to obtain sensor abnormal factors.

8. An intelligent sensor data anomaly recognition device, characterized in that, Including: A decomposition module for performing multi-scale wavelet transform decomposition on the original time-series data of the intelligent sensor to obtain a multi-level signal feature component set; A mapping module for performing distribution probability density mapping on the multi-level signal feature component set based on a non-parametric kernel density estimation algorithm to obtain a feature component probability distribution map; An extraction module for, when the feature component probability distribution map indicates that the original time-series data is abnormal, extracting the boundary of the abnormal region from the feature component probability distribution map through an adaptive threshold segmentation algorithm to obtain a candidate abnormal event boundary sequence; A first analysis module for performing topological correlation analysis on the candidate abnormal event boundary sequence through a graph convolutional network to obtain a multi-dimensional abnormal pattern relationship graph; A second analysis module for performing fault factor analysis on the intelligent sensor based on the multi-dimensional abnormal pattern relationship graph to obtain abnormal factors of the intelligent sensor, and obtaining corresponding maintenance strategies based on the abnormal factors.

9. A computer device, comprising a memory and a processor, wherein a computer program is stored in the memory, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

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

Images