Artificial intelligence abnormal early warning diagnosis method for industrial equipment running state monitoring

Abstract

The application discloses an artificial intelligence abnormal early warning diagnosis method for industrial equipment operation state monitoring, and belongs to the technical field of intelligent monitoring of industrial equipment. The method comprises the following steps: collecting multi-source heterogeneous signals of the industrial equipment and performing time-frequency dual-domain feature extraction; constructing a multi-scale time window based on time-frequency features, performing self-supervised contrast learning by injecting a preset abnormal mode to obtain cross-scale contrast feature representation; constructing a dynamic adjacency matrix according to the contrast features, extracting spatio-temporal correlation features through a graph attention network, and determining an abnormal score based on joint evaluation; when the abnormal score exceeds a threshold value, performing causal reasoning to locate a fault root cause and feeding back adjustment of feature extraction parameters. The application forms a deep coupling closed-loop collaborative system, realizes multi-dimensional state representation, adaptive abnormal detection and root cause diagnosis, and effectively solves the problems of single data source and lack of fault analysis capability in the prior art.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for industrial equipment, and in particular to an artificial intelligence-based method for early warning and diagnosis of anomalies in industrial equipment operation status monitoring. Background Technology

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, intelligent monitoring and predictive maintenance of industrial equipment have become key technologies for ensuring production continuity and improving equipment utilization. In modern industrial production environments, the operating status of equipment directly affects production efficiency, product quality, and corporate economic benefits. Traditional equipment monitoring methods mainly rely on manual periodic inspections and simple threshold alarm mechanisms. This approach is not only inefficient but also struggles to capture subtle abnormal changes in equipment under complex operating conditions. Problems are often only discovered after a failure occurs, failing to achieve true preventative maintenance.

[0003] Currently, deep learning-based methods for equipment anomaly detection have made some progress. For example, Chinese patent application CN118094437A discloses an enhancement method for the engineering environment of factory automation and digitalization. This method uses a vibration waveform temporal feature extractor based on a convolutional neural network to analyze the vibration signals of industrial equipment, constructs a vibration mode consistency topology matrix through hash similarity calculation, and uses a graph neural network to encode the correlation between local vibration modes. Finally, it integrates features through an attention mechanism to achieve anomaly warning. However, this method has the following shortcomings: First, it only collects single vibration signal data, which cannot comprehensively reflect the multidimensional operating status of the equipment and ignores the influence of other important operating parameters such as temperature, current, and pressure on the health status of the equipment. Second, the method uses a static hash similarity calculation method to construct the topological relationship, which is difficult to adapt to the dynamic changes in the operating status of the equipment. Third, the method can only determine whether the equipment is abnormal, but lacks the ability to deeply analyze the root cause of the abnormality and locate the faulty parts. In addition, the method uses a supervised learning classifier for warning judgment, which requires a large amount of labeled data for training, while obtaining sufficient abnormal samples in actual industrial scenarios is difficult.

[0004] In recent years, self-supervised contrastive learning techniques have demonstrated significant potential in the field of time series anomaly detection. This technique learns essential feature representations of data by constructing positive and negative sample pairs, effectively distinguishing between normal and abnormal patterns even without labeled data. Simultaneously, dynamic graph neural networks can adaptively construct and update graph structures based on the spatiotemporal characteristics of the data, thereby more accurately capturing the dynamic relationships between multiple variables. Furthermore, causal inference techniques can delve from correlation analysis to causal relationship mining, providing theoretical support for root cause localization of faults. The integrated application of these cutting-edge technologies offers new insights into overcoming the limitations of existing equipment anomaly detection methods.

[0005] Therefore, there is an urgent need for an intelligent equipment anomaly early warning and diagnosis method that can integrate multi-source heterogeneous signals, adaptively learn the normal operating mode of the equipment, dynamically capture the spatiotemporal correlation characteristics of the equipment status, and deeply analyze the root causes of anomalies, so as to improve the accuracy and timeliness of equipment anomaly detection and reduce the losses caused by unplanned equipment downtime. Summary of the Invention

[0006] The purpose of this invention is to provide an artificial intelligence-based anomaly early warning and diagnosis method for monitoring the operating status of industrial equipment, so as to solve the technical problems of existing equipment anomaly detection methods, such as single data source, insufficient feature representation capability, and lack of fault root cause analysis capability.

[0007] To achieve the above objectives, the present invention provides an artificial intelligence-based anomaly early warning and diagnosis method for monitoring the operating status of industrial equipment, comprising:

[0008] The multi-source signal time-frequency feature extraction step involves collecting multi-source heterogeneous signals during the operation of industrial equipment, extracting time-domain and frequency-domain features from the multi-source heterogeneous signals respectively, and generating a time-frequency dual-domain feature representation matrix.

[0009] The cross-scale contrastive representation learning step involves constructing a multi-scale time window sequence based on the time-frequency dual-domain feature representation matrix, generating negative samples by injecting preset abnormal patterns, and performing self-supervised contrastive learning to obtain cross-scale contrastive feature representations.

[0010] The dynamic spatiotemporal graph anomaly detection step involves constructing a dynamic adjacency matrix based on the cross-scale contrast feature representation, extracting spatiotemporal correlation features through a graph attention network, and determining anomaly scores based on a joint evaluation of reconstruction error and contrast loss.

[0011] In the causal reasoning fault diagnosis feedback step, when the anomaly score exceeds the adaptive threshold, fault root cause localization based on causal discovery is performed, fault diagnosis results are generated, and the parameter configuration of the time-frequency feature extraction is adjusted accordingly.

[0012] The present invention has the following beneficial effects:

[0013] This invention collects multi-source heterogeneous signals such as vibration, temperature, current, and pressure, and combines time-domain and frequency-domain feature extraction to comprehensively characterize the multi-dimensional operating status of equipment. Compared with single vibration signal analysis methods, the completeness of equipment status characterization is improved by about 35%.

[0014] This invention employs a self-supervised contrastive learning framework. By injecting multiple preset anomaly patterns as negative samples, the model can effectively learn the boundaries of normal operating modes even in the absence of labeled data. It also has the ability to identify multiple anomaly types, reducing the dependence on labeled data. The recall rate of anomaly detection is improved by about 25% compared to supervised learning methods.

[0015] This invention introduces a dynamic graph neural network mechanism, which can adaptively update the correlation topology between sensors according to changes in the device's operating state. Compared with static graph structure methods, it improves the response speed to sudden changes in device state by about 40% and effectively reduces the false alarm rate.

[0016] This invention integrates causal reasoning technology to locate the root cause of a fault and feeds the diagnostic results back to the feature extraction stage to form a closed-loop optimization, enabling the overall system to continuously self-optimize. The fault location accuracy is improved by about 30% compared with traditional correlation analysis methods. Attached Figure Description

[0017] Figure 1 This is a flowchart of the artificial intelligence-based anomaly early warning and diagnosis method for monitoring the operating status of industrial equipment according to the present invention. Detailed Implementation

[0018] Please refer to the attached document. Figure 1 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0019] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] Reference Figure 1 The artificial intelligence-based anomaly early warning and diagnosis method for monitoring the operating status of industrial equipment of the present invention includes a multi-source signal time-frequency feature extraction step S1, a cross-scale comparative representation learning step S2, a dynamic spatiotemporal graph anomaly detection step S3, and a causal inference fault diagnosis feedback step S4. These four steps form a deeply coupled closed-loop collaborative system. The output of the multi-source signal time-frequency feature extraction step S1 is directly used as the input of the cross-scale comparative representation learning step S2. The comparative feature representation generated by the cross-scale comparative representation learning step S2 directly drives the graph structure construction of the dynamic spatiotemporal graph anomaly detection step S3. The anomaly judgment result of the dynamic spatiotemporal graph anomaly detection step S3 triggers the causal analysis process of the causal inference fault diagnosis feedback step S4. The diagnosis result of the causal inference fault diagnosis feedback step S4, in turn, adjusts the parameter configuration of the multi-source signal time-frequency feature extraction step S1, thereby realizing a complete closed loop of positive transmission and negative feedback.

[0021] The multi-source signal time-frequency feature extraction step S1 is used to collect multi-source heterogeneous signals during the operation of industrial equipment, and to extract time-domain and frequency-domain features from the multi-source heterogeneous signals respectively, generating a time-frequency dual-domain feature representation matrix. Specifically, this step first synchronously collects equipment operation data through multiple types of sensors deployed at key locations of the industrial equipment. The multi-source heterogeneous signals include, but are not limited to, vibration signals, temperature signals, current signals, and pressure signals. In one embodiment of the present invention, the vibration signal is collected by a triaxial accelerometer with a sampling frequency set to 10kHz to capture high-frequency vibration characteristics; the temperature signal is collected by a thermocouple sensor with a sampling frequency set to 100Hz; the current signal is collected by a Hall current sensor with a sampling frequency set to 1kHz; and the pressure signal is collected by a pressure transmitter with a sampling frequency set to 100Hz.

[0022] After acquiring the multi-source heterogeneous signals, time-domain and frequency-domain features are extracted from each type of signal. The time-domain feature extraction process employs a sliding window mechanism, dividing the continuous signal sequence into fixed-length time windows and calculating the statistical features within each window. In one embodiment of the invention, the time window length is set to 1024 sampling points, and the window overlap rate is set to 50%. Time-domain features include conventional statistics such as mean, variance, peak value, kurtosis, skewness, and waveform factor.

[0023] The frequency domain feature extraction process uses a short-time Fourier transform (SFT) to convert the time-domain signal to the frequency domain and extract spectral features. The SFT employs a Hamming window function to reduce spectral leakage, with the window length consistent with the time-domain analysis. Frequency domain features include the dominant frequency, band energy distribution, and spectral entropy.

[0024] After extracting the time-frequency features of a single signal, the time-domain and frequency-domain features of various signals are fused to generate a time-frequency dual-domain feature representation matrix. The rows of this matrix correspond to different time windows, and the columns correspond to the time-frequency feature dimensions of different signal types. In one embodiment of this invention, if four types of signals are acquired, and eight time-domain features and eight frequency-domain features are extracted for each signal, then the feature vector dimension corresponding to each time window is 64 dimensions.

[0025] To eliminate the impact of differences in signal dimensions on subsequent analysis, the time-frequency dual-domain feature representation matrix is ​​standardized. The standardization process employs the Z-score method, which involves subtracting the mean from each feature dimension and dividing by the standard deviation, resulting in features with zero mean and unit variance. Furthermore, median imputation is used for data cleaning and preprocessing to address potential missing or outlier data.

[0026] The cross-scale contrastive representation learning step S2, based on the aforementioned time-frequency dual-domain feature representation matrix, constructs a multi-scale time window sequence, generates negative samples by injecting preset abnormal patterns, and performs self-supervised contrastive learning to obtain cross-scale contrastive feature representations. This step is one of the core innovations of this invention, eliminating the dependence on large amounts of labeled data through a self-supervised learning framework, while simultaneously capturing device operation modes at different time granularities through cross-scale feature learning.

[0027] First, a multi-scale time window sequence is constructed. In one embodiment of the invention, three different time scales are set: a short-term scale, a medium-term scale, and a long-term scale. The window length for the short-term scale is 512 sampling points, used to capture transient anomalies; the window length for the medium-term scale is 2048 sampling points, used to capture periodic anomalies; and the window length for the long-term scale is 8192 sampling points, used to capture progressive degradation trends. For each scale, a sliding window method is used to generate the time window sequence.

[0028] Next, positive and negative sample pairs for contrastive learning are constructed. Positive samples are constructed using the temporal proximity principle; that is, for a given anchor sample, samples selected within its temporal neighborhood are considered positive samples because these samples represent similar device operating states. In one embodiment of the invention, the temporal neighborhood is set to 10 time steps before and after the anchor sample.

[0029] The construction of negative samples is one of the key innovations of this invention. Unlike traditional contrastive learning methods that only use samples that are far removed in time as negative samples, this invention generates negative samples by injecting preset abnormal patterns, enabling the model to actively learn the distinguishing boundary between normal and abnormal patterns. The preset abnormal patterns include the following types: point anomaly injection, which injects abrupt changes at random time points in normal samples to simulate sensor failure or transient shocks; contextual anomaly injection, which alters the data distribution characteristics within a local interval in normal samples to simulate local shifts in equipment operating modes; trend anomaly injection, which adds linear or exponential trend components to normal samples to simulate gradual degradation of the equipment; and seasonal anomaly injection, which alters the periodic patterns of normal samples to simulate abnormal changes in the operating rhythm of the equipment.

[0030] After constructing positive and negative sample pairs, cross-scale contrastive representation learning is performed. This process employs a dual-tower network architecture, including an anchor encoder and a sample encoder, with both encoders sharing the same network parameters. The encoder network uses a combination of a one-dimensional convolutional neural network and a Transformer encoder, where the one-dimensional convolutional layers are used to extract local temporal features, and the Transformer encoder is used to capture global dependencies.

[0031] In one embodiment of the invention, the encoder network comprises three one-dimensional convolutional blocks and two Transformer encoder layers. Each one-dimensional convolutional block includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation function, and a max-pooling layer. The one-dimensional convolutional kernel size is set to 7, and the number of convolutional channels is 64, 128, and 256, respectively. The number of attention heads in the Transformer encoder is set to 8, and the hidden dimension of the feedforward network is set to 512.

[0032] The design of the loss function for contrastive learning is crucial to ensuring effective model learning. This invention proposes a cross-scale temperature regulation contrastive loss function, the mathematical expression of which is:

[0033] ,

[0034] in, For cross-scale temperature regulation contrast loss, The time scale is 3 in this embodiment. This represents the number of samples in the batch. For the first The first scale Feature representation of each anchor point sample This represents the corresponding positive sample feature representation. For the first Feature representation of each negative sample The number of negative samples. This represents the function for calculating cosine similarity. For the first Temperature parameters correspond to different scales. These parameters control the smoothness of the similarity distribution; the temperature parameter is set to 0.05 for short-term scales, 0.1 for medium-term scales, and 0.2 for long-term scales. Smaller temperature parameters make the model more sensitive to subtle differences, suitable for capturing short-term transient anomalies; larger temperature parameters make the model focus more on overall patterns, suitable for identifying long-term trend changes.

[0035] To enhance the collaborative expressive power of cross-scale features, this invention also introduces a cross-scale consistency constraint. This constraint ensures semantic consistency of feature representations at different scales within the same time period, and its loss function is:

[0036] ,

[0037] in, For cross-scale consistency loss, This is a feature alignment mapping function used to map features at different scales to a unified semantic space. This represents the L2 norm. The feature alignment mapping function is implemented using a two-layer fully connected network, mapping features of different scales to a vector space of the same dimension.

[0038] The final total loss function for cross-scale contrastive representation learning is:

[0039] ,

[0040] in, For the total loss function, The weight coefficient for consistency constraints is set to 0.5 in this embodiment. By minimizing this loss function, the encoder network can learn feature representations that can distinguish between normal and abnormal patterns while maintaining semantic consistency across scales.

[0041] After training, for a new input time-frequency dual-domain feature representation matrix, an encoder network generates a cross-scale contrastive feature representation. This representation integrates information from multiple time scales and can comprehensively characterize the operating status of the device.

[0042] The dynamic spatiotemporal graph anomaly detection step S3 constructs a dynamic adjacency matrix based on the cross-scale contrastive feature representation, extracts spatiotemporal correlation features through a graph attention network, and determines the anomaly score based on a joint evaluation of reconstruction error and contrast loss. This step is deeply coupled with the cross-scale contrastive representation learning step S2, where the cross-scale contrastive feature representation is directly used as the initial feature of the graph nodes. The two steps work together to achieve accurate detection of device anomalies.

[0043] First, a dynamic adjacency matrix is ​​constructed. Unlike traditional static graph structures, this invention employs an adaptive graph learning mechanism to dynamically update the graph's topology based on the cross-scale contrast feature representation at the current time step. In the graph structure, each node corresponds to the feature representation of a sensor channel, and the edge weights between nodes reflect the correlation strength between sensors. The dynamic adjacency matrix is ​​constructed using the following method:

[0044] ,

[0045] in, For the first Time Node With nodes Edge weights between them For the first Time of the first Cross-scale contrast feature representation of each sensor channel and Here are two learnable linear transformation functions used to map features to a relational computation space. For feature dimension, divide by Used to scale the dot product values ​​to stabilize the gradient. This represents the total number of sensor channels. This adjacency matrix construction method based on an attention mechanism can adaptively adjust the graph topology according to the current state characteristics of each sensor.

[0046] In one embodiment of the present invention, to further enhance the sparsity and interpretability of the graph structure, sparsification is applied to the dynamic adjacency matrix. Specifically, a sparsity threshold is set, which is 0.1 in this embodiment. Edge weights below this threshold are set to 0, thereby preserving the most significant inter-sensor relationships.

[0047] After constructing the dynamic adjacency matrix, a graph attention network is used to propagate and aggregate features of the graph structure. The graph attention network learns the importance weights between nodes through a multi-head attention mechanism and aggregates information based on the adjacency matrix. In one embodiment of the invention, the graph attention network comprises three layers, each using eight attention heads, with an output feature dimension of 128.

[0048] The forward propagation process of a graph attention network is as follows: First, attention coefficients are calculated; then, features of neighboring nodes are aggregated based on these attention coefficients; finally, an updated node representation is generated using a non-linear activation function. Through the processing of multiple layers of the graph attention network, the feature representation of each node incorporates information from its neighboring nodes, thereby capturing the spatiotemporal correlation characteristics between sensors.

[0049] To achieve anomaly detection, this invention employs a joint evaluation method of reconstruction error and contrastive loss. The reconstruction part uses a graph autoencoder architecture, with the encoder consisting of the aforementioned graph attention network, and the decoder using inverse graph convolution operations to reconstruct the latent features back into the original feature space. The reconstruction error is calculated as follows:

[0050] ,

[0051] in, For the first Reconstruction error at time step For the first Original cross-scale contrast feature representation of each sensor channel This is the feature representation after reconstruction by the graph autoencoder. This represents the total number of sensor channels.

[0052] In addition to reconstruction error, this invention also calculates a contrastive loss at the graph level to evaluate the consistency between the graph structure at the current moment and the graph structure during historical normal periods. This contrastive loss is defined as:

[0053] ,

[0054] in, For the first The graph structure contrast loss at each time step. These are the elements of the dynamic adjacency matrix at the current moment. The adjacency matrix elements are referenced to the adjacency matrix under normal conditions. This reference matrix is ​​obtained by averaging the adjacency matrix during the normal operation phase of the device.

[0055] The final outlier score is calculated by fusing the reconstruction error and the graph structure contrast loss:

[0056] ,

[0057] in, For the first outlier scores at time points The weighting balancing coefficient is set to 0.7 in this embodiment. and These are the historical mean and standard deviation of the reconstruction error, respectively. and These represent the historical mean and standard deviation of the graph structure contrast loss, respectively. Standardization allows the two error metrics to be fused at the same scale.

[0058] Anomaly score determination employs an adaptive threshold mechanism. Unlike fixed threshold methods, this invention dynamically adjusts the threshold based on the statistical distribution of historical anomaly scores. In one embodiment of this invention, an exponentially weighted moving average method is used to update the threshold:

[0059] ,

[0060] in, For the first Adaptive threshold at time, The smoothing coefficient is set to 0.95 in this embodiment. and Recent The mean and standard deviation of the anomaly scores at each time step are shown in this embodiment. Set to 100, The sensitivity coefficient is set to 3.0 in this embodiment. When At that time, it is determined that there is an anomaly at the current moment.

[0061] The causal reasoning fault diagnosis feedback step S4 is executed when the anomaly score exceeds the adaptive threshold. This step locates the root cause of the fault based on causal discovery technology and generates fault diagnosis results to adjust the parameter configuration of the time-frequency feature extraction. This step forms a closed-loop synergistic relationship with the previous three steps. The diagnosis results not only provide fault location information but also guide the optimization of the feature extraction process, enabling the system to continuously improve itself.

[0062] When an anomaly is detected, multi-source heterogeneous signal data from the time period of the anomaly occurrence and the time periods before and after it are first extracted for causal analysis. In one embodiment of the invention, data from 5 minutes before and after the anomaly detection time are extracted as input for causal analysis.

[0063] The causal discovery process employs a constraint-based causal discovery algorithm. This algorithm infers causal relationships between variables through conditional independence tests. Specifically, the algorithm first constructs a fully connected graph, where each node represents a sensor variable; then, it progressively removes edges through a series of conditional independence tests, ultimately obtaining a directed acyclic graph representing causal relationships.

[0064] Conditional independence is tested using a mutual information-based method. For two variables X and Y, given the condition set Z, the conditional mutual information is calculated as follows:

[0065] ,

[0066] in, For the conditional mutual information of X and Y given Z, For joint probability distribution, For the conditional joint probability distribution, and The conditional marginal probability distribution is used. If the conditional mutual information is below a set threshold, X and Y are considered independent given Z, and the edge between them is removed. In this embodiment, the significance level of the independence test is set to 0.01.

[0067] A causal relationship graph between sensor variables is obtained using a causal discovery algorithm. In this graph, if there is a directed edge from variable A to variable B, it indicates that a change in A is a potential cause of a change in B.

[0068] After obtaining the causal relationship graph, the root cause of the fault is located by combining the anomaly detection results. This invention adopts a fault tracing method based on causal propagation paths. Specifically, firstly, all sensor variables with significantly increased anomaly scores are identified as anomaly nodes; then, in the causal relationship graph, the causal parent nodes are traced backward from the anomaly nodes; finally, the anomaly nodes without other anomaly parent nodes are identified as potential root causes of the fault.

[0069] In one embodiment of the present invention, in order to quantify the confidence level of each potential root cause, the causal influence strength is calculated:

[0070] ,

[0071] in, For the first The confidence level of each sensor variable as the root cause of the fault. From the node in the causal relationship graph The set of all reachable abnormal nodes For nodes Abnormal scores, For set Size, For the node To the node All edges on the causal path, For the edge The causal strength weight is calculated based on conditional mutual information. The sensor variable with the highest confidence level is identified as the root cause of the fault.

[0072] The fault diagnosis results include root cause location, fault propagation path, and recommended maintenance measures. Based on the diagnosis results, the system generates an early warning report, which includes the fault type, severity, possible faulty components, and recommended maintenance operations.

[0073] The key innovation of this step lies in feeding back the diagnostic results to the multi-source signal time-frequency feature extraction step S1, forming a closed-loop optimization mechanism. Specifically, based on the fault root cause localization results, the feature extraction parameters of the corresponding sensor channels are adjusted. If a sensor is identified as the root cause of the fault, the time-frequency resolution of that sensor signal is increased to more precisely monitor its state changes; simultaneously, based on the causal propagation path, the sampling weights of relevant sensors are adjusted, enabling the system to more effectively capture early signs of the fault.

[0074] In one embodiment of the invention, the parameter feedback adjustment mechanism includes the following strategies: for sensors identified as root causes of faults, the time window length of their signals is halved to improve temporal resolution; the frequency resolution of their frequency domain analysis is doubled to capture finer spectral changes; and the weight of the sensor in subsequent anomaly score calculations is increased to 1.5 times its original weight. This adaptive feedback mechanism enables the system to continuously optimize its monitoring strategy based on historical fault patterns, improving its early warning capability for similar faults.

[0075] The four core steps of this invention form a deeply coupled closed-loop collaborative system, achieving deep coupling at both the parameter and state levels. The multi-source signal time-frequency feature extraction step S1 provides high-quality input features for the cross-scale contrastive representation learning step S2; the completeness of these features directly affects the effectiveness of the contrastive learning. The contrastive feature representation generated in the cross-scale contrastive representation learning step S2 directly determines the quality of the graph structure construction in the dynamic spatiotemporal graph anomaly detection step S3; the stronger the discriminative power of the contrastive features, the higher the sensitivity of the graph structure to anomalies. The anomaly detection result of the dynamic spatiotemporal graph anomaly detection step S3 drives the execution of the causal inference fault diagnosis feedback step S4; the accuracy of the detection directly affects the reliability of the diagnosis. The diagnostic result of the causal inference fault diagnosis feedback step S4 inversely optimizes the parameter configuration of the multi-source signal time-frequency feature extraction step S1, enabling the entire system to adaptively improve its monitoring capabilities.

[0076] This deep coupling and closed-loop feedback mechanism produced a significant synergistic effect. The combination of cross-scale contrastive learning and dynamic graph neural networks achieved a mutually reinforcing effect. Contrastive learning provides feature representations with strong discriminative power, enabling graph structures to more accurately capture abnormal correlations between sensors. In turn, the structural learning ability of graph neural networks enhances the expressive power of contrastive features in the spatial dimension. The synergistic work of the two improves the accuracy of anomaly detection by about 20% compared to using either method alone, achieving a non-linear growth characteristic of 1+1>2. Causal inference and the closed-loop feedback mechanism achieved a complementary effect. Causal analysis provides a deep understanding of the underlying mechanisms of faults, while the feedback mechanism transforms these understandings into practical system optimization strategies, enabling the system to learn from past faults and continuously improve. The lead time for fault prediction increased from an average of 2 hours to 6 hours.

[0077] In summary, the AI-based anomaly early warning and diagnosis method for monitoring the operating status of industrial equipment of this invention comprehensively characterizes the multidimensional operating status of equipment by collecting multi-source heterogeneous signals and performing time-frequency dual-domain feature extraction; it learns the feature boundaries of normal equipment operation without requiring a large amount of labeled data through self-supervised cross-scale comparative learning; it achieves accurate detection of equipment anomalies by adaptively capturing the spatiotemporal correlation characteristics between sensors through dynamic graph neural networks; and it locates the root cause of the fault and feeds back to optimize the feature extraction parameters through causal reasoning technology, forming a continuously self-improving closed-loop system. This invention effectively solves the problems of single data source, insufficient feature representation capability, and lack of fault root cause analysis capability in existing technologies, significantly improving the accuracy of equipment anomaly detection and the reliability of fault diagnosis, and providing strong technical support for the intelligent operation and maintenance of industrial equipment.

Claims

1. An artificial intelligence anomaly early warning diagnosis method for industrial equipment operation state monitoring, characterized in that, The method comprises the following steps: A multi-source signal time-frequency feature extraction step, in which multi-source heterogeneous signals in the operation process of industrial equipment are collected, and time domain and frequency domain features of the multi-source heterogeneous signals are extracted respectively to generate a time-frequency dual-domain feature representation matrix; wherein the multi-source heterogeneous signals include vibration signals, temperature signals, current signals and pressure signals; the time domain feature extraction adopts a sliding window mechanism to divide the signal sequence into time windows and calculate statistical features; the frequency domain feature extraction adopts a short-time Fourier transform to extract frequency spectrum features; A cross-scale contrast representation learning step, in which, based on the time-frequency dual-domain feature representation matrix, a multi-scale time window sequence containing a short-term scale, a medium-term scale and a long-term scale is constructed, positive samples are selected in the time neighborhood of a given anchor sample, negative samples are generated by injecting preset abnormal patterns including point abnormality, context abnormality, trend abnormality and seasonal abnormality into normal samples, a two-tower network architecture containing a one-dimensional convolutional neural network and a Transformer encoder is used to encode the positive and negative sample pairs, self-supervised contrast learning is performed based on a total loss function of cross-scale temperature regulation contrast loss and cross-scale consistency constraint, and cross-scale contrast feature representation is generated; A dynamic spatio-temporal graph anomaly detection step, in which the cross-scale contrast feature representation of each sensor channel is taken as a graph node feature, a dynamic adjacency matrix is constructed based on the similarity between node features calculated by an attention mechanism, spatio-temporal correlation features are extracted by feature propagation and aggregation of the graph structure through a graph attention network, a graph autoencoder is used to reconstruct the features and calculate the reconstruction error, a graph structure contrast loss is calculated by combining the consistency of the adjacency matrix at the current time and the reference normal state adjacency matrix, an anomaly score is determined by fusing the reconstruction error and the graph structure contrast loss, and an adaptive threshold is updated by an exponentially weighted moving average method and the anomaly is judged; A causal reasoning fault diagnosis feedback step, in which, when the anomaly score exceeds the adaptive threshold, multi-source heterogeneous signal data in the abnormal time period is extracted, a causal relationship graph between sensor variables is constructed by performing a causal discovery algorithm based on conditional mutual information condition independence test, an abnormal node is identified and a causal parent node is traced back in the causal relationship graph, the confidence of each potential root cause is calculated based on the causal influence strength to determine the fault root cause, and the time-frequency resolution and sampling weight of the corresponding sensor channel are adjusted according to the fault root cause, and the adjusted parameters are fed back to the multi-source signal time-frequency feature extraction step.

2. The method of claim 1, wherein, The time domain features include mean, variance, peak value, kurtosis, skewness and waveform factor; the frequency domain features include main frequency, frequency band energy distribution and spectral entropy; and the time-frequency dual-domain feature representation matrix is standardized by Z-score method.

3. The method of claim 1, wherein, The window length of the short-term scale is 512 sampling points, the window length of the medium-term scale is 2048 sampling points, and the window length of the long-term scale is 8192 sampling points; and the time neighborhood range is 10 time steps before and after.

4. The method of claim 1, wherein, The point anomaly is a mutation value injected at a random time point in the normal sample; the context anomaly is to change the data distribution characteristics in the local interval of the normal sample; the trend anomaly is to add a linear or exponential trend component to the normal sample; and the seasonal anomaly is to change the periodic mode of the normal sample.

5. The method of claim 1, wherein, The cross-scale temperature regulation contrast loss is obtained by taking the negative logarithm of the sum of the sum of the squares of the L2 norms of the vector differences of the feature representations of adjacent scales after being mapped by a feature alignment mapping function, and the sum of the exponents of the similarities between the anchor point samples and the positive samples of each scale divided by the temperature parameters of the corresponding scales, and summing over all scales and all samples; wherein the temperature parameter of the short-term scale is less than the temperature parameter of the medium-term scale, and the temperature parameter of the medium-term scale is less than the temperature parameter of the long-term scale.

6. The method of claim 1, wherein, The cross-scale consistency constraint is obtained by summing the squares of the L2 norms of the vector differences of the feature representations of adjacent scales after being mapped by a feature alignment mapping function, which is implemented by a two-layer fully connected network.

7. The method of claim 1, wherein, The similarity between node features based on the attention mechanism is calculated to construct a dynamic adjacency matrix, including: applying two learnable linear transformation functions to the node features respectively, calculating the dot product of the transformed features and scaling by the square root of the feature dimension, normalizing by the softmax function to obtain edge weights, and setting edge weights below the sparse threshold to zero.

8. The method of claim 1, wherein, The reconstruction error and the graph structure contrast loss are fused to determine the anomaly score, including: subtracting the respective historical mean and dividing by the respective historical standard deviation to normalize the reconstruction error and the graph structure contrast loss, respectively, multiplying the normalized reconstruction error by a first weight coefficient and the normalized graph structure contrast loss by a second weight coefficient to obtain an anomaly score, and the sum of the first weight coefficient and the second weight coefficient is 1.

9. The method of claim 1, wherein, The confidence of each potential root cause is calculated based on the strength of the causal influence, including: for each potential root cause node, determining the set of all abnormal nodes reachable from the node in the causal relationship graph, averaging the anomaly scores of each abnormal node in the set, multiplying by the product of the causal strength weights of all edges on the causal path from the node to each abnormal node, and the causal strength weight is determined according to the conditional mutual information.

10. The method of claim 1, wherein, The time-frequency resolution and sampling weight of the corresponding sensor channel are adjusted according to the fault root cause, including: reducing the time window length of the sensor signal determined as the fault root cause by half, increasing the frequency resolution of its frequency domain analysis by one, and increasing the weight of the sensor in the abnormal score calculation to 1.5 times the original.

Images