A raceway anomaly detection method and system based on deep learning

Abstract

The application relates to the technical field of raceway system thermal crown detection, and discloses a raceway thermal crown design method and system based on deep learning, which comprises the following steps: constructing an initial fault knowledge graph to form the association relationship between known thermal abnormal patterns and fault types; collecting real-time thermal image sequences of a raceway system, and performing abnormal score calculation and region positioning on the collected images through a variational autoencoder; designing a double-flow causal reasoning network to realize synchronous processing of image data flow and knowledge graph structure flow; analyzing the association strength between thermal abnormal patterns and existing nodes in a fault knowledge graph; applying a graph expansion algorithm to realize automatic updating of the fault knowledge graph; and carrying out time sequence evolution analysis to predict potential fault development trends; the double-flow causal reasoning network and the graph expansion algorithm are designed, so that the continuous accumulation and evolution capability of raceway system fault knowledge is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of raceway system thermal crown detection, more particularly, it relates to a raceway anomaly detection method and system based on deep learning. BACKGROUND

[0002] With the continuous improvement of industrial automation and intelligent manufacturing level, the safety and reliability of the raceway system as a key mechanical transmission component are widely concerned.

[0003] In the prior art, the anomaly detection method for the raceway system mainly relies on static models based on rules or traditional machine learning. This kind of method is usually trained by limited historical data before system deployment, and it is difficult to dynamically adapt to the fault evolution caused by equipment aging, working condition changes or new fault modes in the actual running process. Traditional methods generally lack the ability of continuous accumulation and updating of fault knowledge, and cannot fully mine and utilize the causal relationship contained in historical cases, resulting in limited recognition ability for new, rare and composite faults. In addition, the existing anomaly detection means are insufficient in detection accuracy and generalization ability when facing complex and variable running environment, and it is difficult to meet the actual needs of long-term stable operation and intelligent maintenance of the raceway system.

[0004] Therefore, there is an urgent need for a raceway system anomaly detection technology that can realize adaptive learning of fault knowledge, dynamic expansion and improve the recognition ability of new faults, in order to ensure the safety and reliability of its operation. SUMMARY

[0005] The present application provides a raceway anomaly detection method and system based on deep learning, which solves the technical problems of lack of fault knowledge accumulation learning ability, inability to adapt to fault mode evolution and insufficient recognition ability for new faults and composite faults in related technologies.

[0006] The present application provides a raceway anomaly detection method based on deep learning, comprising:

[0007] An initial fault knowledge graph is constructed to form the association between known thermal anomaly patterns and fault types;

[0008] Based on the initial fault knowledge graph, real-time thermal image sequences of the raceway system are collected, and the collected images are subjected to anomaly score calculation and region positioning by a variational autoencoder;

[0009] Using the anomaly score and region positioning results, a double-flow causal reasoning network is designed to realize the synchronous processing of image data flow and knowledge graph structure flow;

[0010] In combination with the detected thermal anomaly patterns, the association strength between the thermal anomaly patterns and the existing nodes in the fault knowledge graph is analyzed.

[0011] According to the correlation strength analysis result, a graph expansion algorithm is applied to realize automatic updating of the fault knowledge graph.

[0012] Based on the automatic updating result of the fault knowledge graph and the time sequence data, time sequence evolution analysis is carried out to predict potential fault development trend.

[0013] Further, the step of constructing the initial fault knowledge graph comprises:

[0014] Collecting historical raceway system fault case data;

[0015] Preprocessing the collected data, including thermal image feature standardization and fault type classification coding;

[0016] Establishing fault type nodes and thermal anomaly mode nodes;

[0017] Based on expert experience and historical data analysis, the correlation edges between nodes are established.

[0018] Further, the step of collecting the real-time thermal image sequence of the raceway system comprises:

[0019] Deploying a thermal imaging device to collect thermal image sequences of the raceway system at a fixed frequency;

[0020] Preprocessing the collected images, including noise filtering, contrast standardization and geometric correction;

[0021] Using reconstruction error to calculate the anomaly score of each image region;

[0022] Setting an adaptive threshold to mark the regions exceeding the threshold.

[0023] Further, the double-flow causal reasoning network comprises an image flow processing component, which comprises:

[0024] A spatial feature extraction unit composed of a multi-layer convolutional neural network;

[0025] A time sequence feature extraction unit composed of a long short-term memory network and a time convolution network;

[0026] A feature fusion unit for fusing spatial features and time sequence features.

[0027] Further, the double-flow causal reasoning network further comprises a knowledge graph structure flow component, which comprises:

[0028] A graph representation unit composed of a graph convolution network;

[0029] a subgraph sampling unit, configured to sample a local subgraph in the fault knowledge graph according to the detected abnormal area;

[0030] a node weighting unit, comprising an attention mechanism module, configured to weight the importance of nodes in the subgraph.

[0031] Further, the double-flow causal reasoning network further comprises a double-flow feature fusion component, which comprises:

[0032] a cross-modal attention unit, configured to align and fuse the image features and the graph features;

[0033] a causal reasoning unit, configured to infer potential fault causes based on the fused features;

[0034] a prediction output unit, configured to generate an abnormal type prediction result and a probability distribution thereof.

[0035] Further, the step of analyzing the correlation strength between the abnormal heat pattern and the existing nodes in the fault knowledge graph comprises:

[0036] calculating the similarity between the feature vector of the new abnormal heat pattern and each node in the fault knowledge graph;

[0037] using an attention mechanism-based correlation strength evaluation algorithm to quantify the correlation degree between the new abnormal heat pattern and the existing fault types;

[0038] if the highest correlation strength is lower than a set threshold, the new abnormal heat pattern is marked as a potential new fault pattern;

[0039] if the correlation strength distribution presents a multi-peak characteristic, the new abnormal heat pattern is marked as a potential complex fault pattern.

[0040] Further, the steps of the graph expansion algorithm comprise:

[0041] for the identified new fault pattern, a new abnormal heat pattern node is created in the fault knowledge graph;

[0042] a unique identifier is assigned to the new node, and the detected abnormal area features are used as the node feature representation;

[0043] if the correlation strength analysis result shows that there is a new fault type, a new fault type node is created;

[0044] an association edge is established between the new node and the related existing nodes, and an initial weight is set.

[0045] Further, the steps of the time series evolution analysis comprise:

[0046] a time series graph evolution model is constructed to capture the time evolution law of the fault pattern;

[0047] Utilize the causal relationship chain in the fault knowledge graph to infer the fault propagation path;

[0048] Combine equipment operating parameters and environmental factors to predict potential fault development timelines;

[0049] Generate multi-level warning information, including short-term fault risk and long-term system health status prediction.

[0050] The application provides a raceway anomaly detection system based on deep learning, which is used to perform the raceway anomaly detection method based on deep learning.

[0051] The fault knowledge graph construction module is used to construct an initial fault knowledge graph containing the association relationship between known thermal anomaly patterns and fault types.

[0052] The thermal image anomaly detection module is used to collect real-time thermal image sequences of the raceway system, and perform anomaly score calculation and region positioning through a variational autoencoder.

[0053] The dual-flow causal reasoning module is used to design a dual-flow causal reasoning network to realize synchronous processing of image data flow and knowledge graph structure flow.

[0054] The correlation strength analysis module is used to analyze the correlation strength between the detected thermal anomaly patterns and the existing nodes in the fault knowledge graph.

[0055] The graph automatic updating module is used to automatically update the fault knowledge graph by applying a graph expansion algorithm based on the correlation strength analysis results.

[0056] The fault trend prediction module is used to carry out time series evolution analysis based on the automatic updating results of the fault knowledge graph and time series data, and predict potential fault development trends.

[0057] The application has the following advantages: through the dual-flow causal reasoning network, the abnormal features in the image space are combined and analyzed with the semantic relationship in the knowledge graph, so that the application can learn fault knowledge from historical cases, and through the graph expansion algorithm, new abnormal patterns and their causal relationships are automatically added to the knowledge graph, realizing continuous accumulation of fault knowledge.

[0058] Compared with the static anomaly detection model, the application has stronger new fault recognition capability, and by analyzing the correlation strength between new thermal anomaly patterns and existing nodes in the fault knowledge graph, the application can determine whether a new fault has occurred and include it in the knowledge system, thereby enhancing the adaptability to fault pattern evolution caused by technical iteration and environmental changes.

[0059] The correlation strength evaluation algorithm based on the attention mechanism can identify the complex fault mode with a multi-peak characteristic in the correlation strength distribution, make up for the deficiency of the traditional method in complex fault identification, and improve the understanding ability of the system to the complex fault condition.

[0060] Through the time evolution analysis, the application can predict the potential fault development trend and new fault mode, realize the change from passive response to active prevention, and provide more forward-looking support for the maintenance decision of the raceway system.

[0061] The knowledge graph structure of the application provides a visual representation of the fault type, reason, symptom and their mutual relationship, enhances the explainability, and enables the maintenance personnel to better understand the cause and development mechanism of the fault. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 is a flowchart of a raceway anomaly detection method based on deep learning in the application;

[0063] Figure 2 is a line chart of the change trend of the new fault recognition rate of different fault detection methods in different stages of system operation;

[0064] Figure 3 is a bar chart of the comparison of the composite fault recognition accuracy and fault component analysis accuracy of the traditional anomaly detection method, the single deep learning model and the method of the application;

[0065] Figure 4 is a column chart of the change of the node number of the fault knowledge graph in different stages after the operation of the application;

[0066] Figure 5 is a column chart of the comparison of the fault early warning lead time of the traditional method and the method of the embodiment under different combinations of fault types. DETAILED DESCRIPTION

[0067] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can include changes in the function and arrangement of elements discussed without departing from the scope of the disclosure. Various examples can omit, substitute, or add various procedures or components as appropriate. Also, some features described with respect to one example can be combined in other examples.

[0068] At least one embodiment of the application discloses a raceway anomaly detection method and system based on deep learning, as shown in Figure 1 , comprising:

[0069] Step 1, build an initial fault knowledge graph to form the association between known thermal anomaly patterns and fault types;

[0070] This step builds an initial knowledge graph containing the association between known thermal anomaly patterns and fault types. Specifically, it includes:

[0071] Step 1.1, collect historical raceway system fault case data;

[0072] The historical raceway system fault case data includes fault type, thermal image features, environmental parameters, and operating state information;

[0073] Step 1.2, pre-process the collected data;

[0074] Data preprocessing includes standardizing thermal image features and classifying and encoding fault types;

[0075] Further, the data preprocessing includes:

[0076] Thermal image feature standardization: subtract the mean value of the thermal image pixel value and divide by the standard deviation to convert it to a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating numerical differences under different collection devices and environmental conditions;

[0077] Thermal image pixel value preprocessing: thermal image pixel values need to be standardized and preprocessed to unify pixel values from different collection devices to the same range. For contrast standardization, an adaptive threshold method should be used to dynamically adjust the mapping parameters according to the image content. Pixel values should also be subjected to outlier detection and processing to remove obvious noise and outliers;

[0078] Fault type classification encoding: use One-Hot Encoding (OHE) to convert text-based fault types into numerical vectors for easy data processing and model training;

[0079] Fault type classification data preprocessing: classification data such as fault types need to be encoded and preprocessed. Use One-Hot Encoding to convert text-based fault types into numerical vectors. For newly added fault types, a dynamic coding mechanism should be established to ensure consistency and scalability of the code. The encoded vector should be checked for sparsity to avoid dimension disaster problems;

[0080] Time series data alignment: interpolate and resample thermal image sequences collected at different time points based on timestamps to ensure consistent time intervals;

[0081] Step 1.3, establish fault type nodes and thermal anomaly pattern nodes;

[0082] The fault type node represents different categories of faults, and the hot anomaly pattern node represents specific heat distribution characteristics;

[0083] Step 1.4, based on expert experience and historical data analysis, establish the association edge between nodes;

[0084] The weight of the edge represents the association strength;

[0085] The association edge weight needs to be range-limited for preprocessing, limiting its value to the [0, 1] interval to ensure the rationality of the weight value. For the initial weight of the new edge, it should be set based on the similarity calculation result. When updating the weight, normalization processing should be performed to avoid the weight value accumulating too large or too small over time;

[0086] Store the constructed knowledge graph as a graph data structure for subsequent algorithm access and update.

[0087] Step 2, based on the initial fault knowledge graph, collect real-time thermal image sequences of the raceway system, and calculate abnormal scores and region positioning of the collected images through a variational autoencoder;

[0088] This step performs anomaly detection and region positioning on the collected thermal image sequences through a variational autoencoder (VAE, Variational AutoEncoder). Specifically, it includes:

[0089] Step 2.1, deploy thermal imaging equipment;

[0090] Install high-precision infrared thermal imagers at key locations of the raceway system, including around the bearing outer ring, inner ring, retainer, and other key components. The thermal imaging equipment should have high resolution (not less than 640×480 pixels), high temperature accuracy (±0.5℃), and high sampling frequency (not less than 30 seconds / second) and other performance indicators. The equipment installation location should ensure complete coverage of the thermal distribution area of the raceway system, avoiding blind spots and obstructions. At the same time, configure a data acquisition and transmission system to ensure that thermal image data can be transmitted to the central processing unit in real time for subsequent analysis.

[0091] Collect thermal image sequences of the raceway system at a fixed frequency;

[0092] Step 2.2, pre-process the collected images;

[0093] Pre-processing of the collected images includes noise filtering, standardization, and time series alignment;

[0094] Further, the pre-processing of the collected images includes:

[0095] Noise filtering: use Gaussian filtering or median filtering and other methods to remove random noise and outliers in thermal images;

[0096] Contrast normalization: linearly maps the pixel values ​​of the thermal image to the [0, 1] interval, enhancing the identifiability of thermal anomaly areas;

[0097] Geometric correction: Corrects perspective distortion and aberration in images to ensure that thermal images acquired at different times are spatially aligned;

[0098] Step 2.3: Construct and train the variational autoencoder model;

[0099] Furthermore, the variational autoencoder (VAE) model includes:

[0100] Encoder component: A sequence of neural network layers responsible for mapping the input thermal image to the latent space, including multiple convolutional layers, pooling layers, and fully connected layers;

[0101] Latent spatial components: a low-dimensional feature vector space used to represent thermal images, defined by the mean vector and the variance vector;

[0102] Decoder component: A sequence of neural network layers responsible for reconstructing the original thermal image from the latent space, including multiple deconvolutional layers and upsampling layers;

[0103] Loss function component: Composed of reconstruction loss and KL divergence loss, used to optimize model training;

[0104] Step 2.4: Calculate the anomaly score for each image region using the reconstruction error;

[0105] Furthermore, a higher anomaly score indicates a greater likelihood of an anomaly, and the formula for calculating the anomaly score is as follows:

[0106] ;

[0107] in, Representing coordinates Abnormal scores at the location, This represents the pixel value of the original thermal image at this coordinate. This represents the pixel value of the reconstructed thermal image at that coordinate. This represents the Euclidean Norm.

[0108] Step 2.5: Set an adaptive threshold, locate and mark areas exceeding the threshold, and obtain the coordinates of suspected abnormal areas.

[0109] Furthermore, the formula for calculating the adaptive threshold is:

[0110] ;

[0111] wherein, represents an adaptive threshold value, represents the mean of the abnormal score of the entire image region, represents the standard deviation of the abnormal score of the entire image region, is an adjustable sensitivity coefficient;

[0112] Adjustable sensitivity coefficient Normalization preprocessing is required to limit the value range to a reasonable interval, usually set to between 1.0 and 3.0, to avoid abnormal threshold calculation due to excessively large or small values. At system initialization, The value should be determined based on historical data statistical analysis and dynamically adjusted according to the detection effect during operation;

[0113] The specific implementation of is to calculate the arithmetic mean of the abnormal scores of all pixel positions in the entire image region, and is obtained by dividing the sum of all abnormal scores by the total number of pixels, which is used to represent the overall abnormal level of the image.

[0114] The specific implementation of is to calculate the standard deviation of the abnormal score, which is obtained by calculating the square of the difference between each abnormal score and the mean, summing up and then taking the square root of the total number of pixels, which is used to represent the dispersion degree and change amplitude of the abnormal score. The larger the value, the more uneven the abnormal distribution.

[0115] Step 3, using the abnormal score and the region positioning result, a double-flow causal reasoning network is designed to realize the synchronous processing of image data flow and knowledge graph structure flow;

[0116] This step is the core innovation of the embodiment, and a double-flow causal reasoning network capable of simultaneously processing image data flow and knowledge graph structure flow is designed.

[0117] Further, the double-flow causal reasoning network includes the following components:

[0118] Image flow processing component:

[0119] Spatial feature extraction unit: composed of a multi-layer convolutional neural network (CNN, Convolutional Neural Network), used to extract the spatial features of the thermal image;

[0120] Time sequence feature extraction unit: composed of a long short-term memory network (LSTM, Long Short-Term Memory) or a time convolutional network (TCN, Temporal Convolutional Network), used to extract the time sequence features of the thermal image sequence;

[0121] Feature fusion unit: for fusing spatial features and temporal features to generate comprehensive image feature vectors.

[0122] Knowledge graph structure flow component:

[0123] Graph representation unit: composed of a graph convolutional network (GCN), for representation learning of the knowledge graph;

[0124] Subgraph sampling unit: responsible for local subgraph sampling in the knowledge graph according to the detected abnormal area;

[0125] Node weighting unit: contains an attention mechanism module, which weights the importance of nodes in the subgraph and generates a structured feature vector.

[0126] Dual-flow feature fusion component:

[0127] Cross-modal attention unit: aligns and fuses image features and graph features;

[0128] Causal reasoning unit: infers potential fault causes based on fused features;

[0129] Prediction output unit: generates abnormal type prediction results and their probability distribution.

[0130] Step 4, in combination with the detected thermal anomaly pattern, analyze the correlation strength between the thermal anomaly pattern and the existing nodes in the fault knowledge graph;

[0131] This step is based on the reasoning results of step 3 to analyze the correlation between the detected new thermal anomaly pattern and the existing nodes in the fault knowledge graph. Specifically, it includes:

[0132] As shown in Figure 2 Figure 3 shows the change trend of new fault recognition rate of different fault detection methods in different stages (initial stage, middle stage, and late stage) of system operation, reflecting the technical effect that the method of the embodiment improves continuously with the improvement of knowledge accumulation ability. As shown in Figure 3, the new fault recognition rate of the method of the embodiment in the initial stage (1-3 months) is 78.6%, in the middle stage (4-6 months) is improved to 86.7%, and in the late stage (7-12 months) is 94.2%, which is improved by 20.4, 29.2 and 38.4 percentage points respectively compared with the traditional static model, fully proving the new fault adaptability of the method.

[0133] Step 4.1, calculate the similarity between the feature vector of the new thermal anomaly pattern and each node in the fault knowledge graph;

[0134] Further, before calculating the similarity, the feature vector is preprocessed:

[0135] Dimension unification: through feature mapping or dimension reduction techniques, ensure that the feature vector of the new hot anomaly pattern has the same dimension as the feature vector of the node in the fault knowledge graph;

[0136] L2 norm normalization: length normalization is performed on the feature vector, so that the Euclidean norm of all feature vectors is 1, eliminating the influence of amplitude difference and only comparing direction similarity;

[0137] Further, the similarity calculation adopts a cosine similarity function, and the calculation formula is:

[0138] ;

[0139] wherein, denotes the feature vector of the new hot anomaly pattern, denotes the feature vector of the node in the fault knowledge graph, denotes the Euclidean norm of the vector, denotes and the cosine similarity between them, the value range is , the larger the value, the more similar the two feature vectors.

[0140] Step 4.2, adopt an association strength evaluation algorithm based on attention mechanism to quantify the association degree of the new hot anomaly pattern and the existing fault type;

[0141] Further, the steps of the association strength evaluation algorithm include:

[0142] Calculate the feature similarity of the new hot anomaly pattern feature vector and each fault type node in the fault knowledge graph;

[0143] Weight the similarity based on attention weight, highlight the influence of key features;

[0144] Further, the calculation formula of the attention weight is:

[0145] ;

[0146] wherein, denotes the attention weight of the th feature, and denote the attention score of the th feature and the th feature respectively, denotes the total number of features, denotes the exponential function;

[0147] The attention score The weighted similarity is mapped to the interval [0, 1] by a nonlinear transformation function to obtain the correlation strength value. The correlation strength value is calculated by the dot product of the feature vector and the trainable attention query vector. Normalization preprocessing is required to limit the numerical range to a reasonable interval to avoid overflow due to large values in exponential operations. The softmax function is usually used for normalization to ensure that the sum of all attention weights is 1. During the calculation process, the scores with abnormally large values should be truncated;

[0148] The weighted similarity is mapped to the interval [0, 1] by a nonlinear transformation function to obtain the correlation strength value. The correlation strength value is calculated by the dot product of the feature vector and the trainable attention query vector. Normalization preprocessing is required to limit the numerical range to a reasonable interval to avoid overflow due to large values in exponential operations. The softmax function is usually used for normalization to ensure that the sum of all attention weights is 1. During the calculation process, the scores with abnormally large values should be truncated;

[0149] Further, the nonlinear transformation function is a Sigmoid function, and the calculation formula is:

[0150] ;

[0151] wherein, represents the correlation strength value, represents the weighted similarity, is a parameter that controls the slope of the function, is a parameter that controls the center point of the function, represents the base of the natural logarithm, approximately equal to 2.71828;

[0152] Range limitation preprocessing is required, usually set to between 2.0 and 10.0, the larger the value, the steeper the function curve. This parameter should be initialized based on the distribution characteristics of the training data and optimized during the model training process;

[0153] Center point calibration preprocessing is required, usually set to between -2.0 and 2.0, to adjust the center position of the function. This parameter should be set according to the distribution characteristics of the similarity data to ensure the reasonableness of the function output;

[0154] According to the distribution characteristics of the correlation strength, the fault type is determined. Single-peak distribution corresponds to a single fault, and multi-peak distribution corresponds to a composite fault.

[0155] If the highest correlation strength is lower than the set threshold, it is marked as a potential new fault mode.

[0156] If the correlation strength distribution shows multi-peak characteristics, it is marked as a potential composite fault mode.

[0157] For example, Figure 3As shown, the performance of traditional anomaly detection methods, single deep learning models, and the method of this invention in terms of accuracy in identifying complex faults and correctness in analyzing fault components is compared, verifying the advantages of this embodiment in identifying complex faults. The method of this invention achieves an accuracy of 86.7% in identifying complex faults and a correctness of 78.3% in analyzing fault components, which are 60.0 and 60.1 percentage points higher than traditional anomaly detection methods, respectively, and 24.5 and 32.7 percentage points higher than single deep learning models, fully demonstrating the advantages of this method in identifying complex faults.

[0158] The Sigmoid function is implemented by mapping the weighted similarity to the 0-1 interval through exponential transformation and normalization. The steepness of the control function is determined by the value of the curve; the larger the value, the steeper the curve and the more sensitive it is to changes in the input. The function controls the center point position; a larger value indicates a more skewed output (1), while a smaller value indicates a more skewed output (0). In association strength assessment, this function converts continuous similarity values ​​into probabilistic association strength, facilitating subsequent fault type identification and threshold comparison.

[0159] Step 5: Based on the correlation strength analysis results, apply the graph expansion algorithm to achieve automatic updating of the fault knowledge graph;

[0160] This step is a key step in realizing the ability to accumulate and learn fault knowledge. The newly discovered thermal anomaly patterns and their causal relationships are automatically added to the fault knowledge graph through the graph expansion algorithm.

[0161] like Figure 4 As shown in Table 1, the number of nodes in the fault knowledge graph changes at different stages of system operation (initial, after 6 months, and after 12 months), intuitively reflecting the expansion capability and cumulative learning effect of the knowledge graph. The number of fault type nodes increased from 150 initially to 182 after 12 months (a growth rate of 21.3%), the number of hot anomaly pattern nodes increased from 320 to 465 (a growth rate of 45.3%), the number of related edges increased from 1200 to 1768 (a growth rate of 47.3%), and the number of high-confidence related edges increased from 450 to 713 (a growth rate of 58.4%), fully demonstrating the fault knowledge cumulative learning capability of this method.

[0162] Furthermore, the steps of the graph expansion algorithm include:

[0163] Step 5.1, Node creation and association steps;

[0164] For the novel fault modes identified in step 4, create new hot anomaly mode nodes in the fault knowledge graph;

[0165] Assign a unique identifier to the new node and use the detected abnormal region features as the node feature representation;

[0166] If the correlation strength analysis results indicate the existence of a new fault type, then a new fault type node is created.

[0167] Based on the association strength analysis results in step 4, establish association edges between the new node and related existing nodes, and set initial weights.

[0168] Step 5.2, map consistency maintenance steps;

[0169] Scan the relationship between newly added edges and existing edges to detect whether there are circular dependencies or logical contradictions;

[0170] If a contradiction is found, retain the high-confidence relationship and remove the low-confidence relationship based on the confidence level.

[0171] The weights of all edges are decayed over time, so that the weights of older associated edges gradually decrease.

[0172] Furthermore, the formula for calculating the timing decay is as follows:

[0173] ;

[0174] in, Representing an edge The updated weights Representing an edge The original weights, Indicates the attenuation coefficient. This indicates the time elapsed since the last update;

[0175] Standardization preprocessing is required to convert it into a dimensionless parameter, typically ranging from 0.001 to 0.1. This parameter should be adjusted according to the time unit (hour, day, month) to ensure the reasonableness of time-series decay. Before system deployment, the optimal value should be determined based on the timeliness analysis of historical data.

[0176] Unit standardization preprocessing is required to convert all time values ​​to a unified time unit (such as hours) to avoid attenuation calculation errors caused by inconsistent time units. For time data from different sources, a unified timestamp format should be established and time zone standardization should be performed.

[0177] Automatic graph pruning is achieved by removing edges whose weights are below a set threshold.

[0178] Step 5.3, graph structure optimization steps;

[0179] Calculate the degree centrality and closeness centrality of each node in the fault knowledge graph;

[0180] Further, the degree centrality (Degree Centrality) calculation formula is:

[0181] ;

[0182] Wherein, Degree centrality of node , degree (number of edges connected to it) of node , total number of nodes in the fault knowledge graph ;

[0183] The specific implementation is: count the number of all edges directly connected to node , including incoming edges and outgoing edges, by traversing the adjacency list or adjacency matrix of the graph.

[0184] Further, the closeness centrality calculation formula is:

[0185] ;

[0186] Wherein, Closeness centrality of node , shortest path length from node to node , total number of nodes in the fault knowledge graph ; Indicates the summation symbol;

[0187] Identify important nodes and edges, and build an index structure for the graph;

[0188] Preload frequently accessed subgraphs into the cache to improve query efficiency;

[0189] Periodically rebalance the graph structure to maintain query performance.

[0190] The specific implementation is: use Dijkstra algorithm or Floyd-Warshall algorithm to calculate the shortest path length from node to node , and return infinity if there is no path between the two nodes.

[0191] Step 5.4, incremental learning update step;

[0192] Extract the feature representation of the newly added nodes and edges;

[0193] ​Further, the extraction of feature representation includes a data preprocessing step:

[0194] Feature normalization: different types of features (such as thermal anomaly pattern features, fault type features, and correlation edge weights) are normalized to a unified numerical range;

[0195] Heterogeneous data fusion: encode and uniformly represent thermal image features (continuous data) and fault types (discrete data) to form hybrid features for training;

[0196] Time weighting: introduce a time decay factor for feature representation, so that recent data has a greater impact on the model;

[0197] Feature representation preprocessing: all feature representations need to be checked for consistency during extraction to ensure that features from different sources have the same numerical range and distribution characteristics. For newly added features, a feature selection mechanism should be established to avoid feature redundancy and dimension disaster problems;

[0198] Incremental training using these features to update the parameters of the dual-flow causal reasoning network;

[0199] According to the change of graph scale, dynamically adjust the structure of graph convolutional network (GCN, Graph Convolutional Network);

[0200] Verify the updated model to ensure that the performance is not lower than that before the update.

[0201] Step 6, based on the automatic update results of the fault knowledge graph and the time series data, carry out time series evolution analysis and predict the potential fault development trend;

[0202] This step is based on the updated fault knowledge graph and historical raceway system fault case data to predict the development trend of the fault. Specifically, it uses a time series graph evolution model for prediction and analysis.

[0203] As shown in Figure 5 , the comparison of fault warning advance time between traditional method and the method of the embodiment under different fault type combinations is shown, which highlights the advantages of the method in terms of early warning timeliness. The average early warning advance of the method for composite faults reaches 31.4 hours, which is 25.2 hours higher than the 6.2 hours of the traditional method. The early warning advance of the "lubricant degradation + surface microcracks" combination reaches 32.7 hours, and the early warning advance of the "surface contamination + lubricant degradation" combination reaches 36.2 hours, providing sufficient preparation time for bearing raceway system maintenance and effectively preventing potential equipment damage and safety accidents.

[0204] Further, the time series graph evolution model includes the following components:

[0205] Time series feature extraction component: responsible for extracting time series features from historical raceway system failure case data, capturing the evolution of failure patterns;

[0206] Further, the time series feature extraction requires pre-processing of historical raceway system failure case data:

[0207] Time window normalization: truncate or pad different length of failure case data to a uniform time window length;

[0208] Time series data preprocessing: time series data needs to be window normalized, truncating or padding different length of failure case data to a uniform time window length. For seasonal data, seasonal adjustment should be performed to remove the influence of periodic changes. For multi-scale time feature extraction, different time granularity data should be aligned;

[0209] Seasonal adjustment: remove seasonal factors in time series, such as periodic changes in equipment thermal load;

[0210] Multi-scale time feature extraction: extract features from different time granularities (hour, day, month) at the same time, fully capture short-term fluctuations and long-term trends;

[0211] Causal chain analysis component:

[0212] Path tracking unit: locate and track possible failure propagation paths in the failure knowledge graph;

[0213] Influence evaluation unit: calculate the influence size of each node in the causal chain;

[0214] Propagation prediction unit: predict the possible propagation direction and speed of failure in the system.

[0215] Multi-factor fusion component:

[0216] Parameter integration unit: integrate device operating parameters and environmental factors into the prediction model;

[0217] Further, the parameter integration involves pre-processing of heterogeneous data:

[0218] Dimensionless: convert different dimensional parameters (such as temperature, pressure, speed, etc.) to [0, 1] interval through min-max normalization;

[0219] Device operating parameters and environmental factors preprocessing: Dimensionless preprocessing is required for device operating parameters such as temperature, pressure, and speed. All parameters are converted to the [0, 1] interval through min-max normalization. Environmental humidity and other percentage parameters need to be converted to decimal form. The weight coefficients of different parameters should be determined based on expert knowledge and historical data analysis to ensure that each parameter contributes reasonably to the prediction model;

[0220] Weight calibration: Based on expert knowledge and historical data analysis, appropriate weight coefficients are assigned to different parameters;

[0221] Missing value processing: Missing data in device operating parameters and environmental factors is handled through interpolation or average value filling methods;

[0222] Weight adjustment unit: Dynamically adjust the weight of different factors in prediction according to their importance;

[0223] Timeline generation unit: Generate a detailed timeline of potential fault development.

[0224] Multi-level early warning component:

[0225] Short-term risk assessment unit: Assess the probability and severity of recent failures;

[0226] Long-term trend analysis unit: Predict the changing trend of the system's long-term health status;

[0227] Early warning information formatting unit: Generate customized early warning information according to different user roles.

[0228] A deep learning-based raceway anomaly detection system for executing the above deep learning-based raceway anomaly detection method, comprising:

[0229] Fault knowledge graph construction module for constructing an initial fault knowledge graph containing known thermal anomaly patterns and fault type association relationships;

[0230] Thermal image anomaly detection module for collecting real-time thermal image sequences of the raceway system, and performing anomaly score calculation and region positioning through a variational autoencoder;

[0231] Dual-flow causal reasoning module for designing a dual-flow causal reasoning network to realize synchronous processing of image data flow and knowledge graph structure flow;

[0232] Correlation strength analysis module for analyzing the correlation strength between detected thermal anomaly patterns and existing nodes in the fault knowledge graph;

[0233] Graph automatic update module for automatically updating the fault knowledge graph based on the correlation strength analysis results and applying a graph expansion algorithm;

[0234] a fault trend prediction module, configured to carry out time series evolution analysis based on the automatic update result of the fault knowledge graph and the time series data, and predict potential fault development trends.

[0235] Here, the embodiment provides an application example:

[0236] The embodiment takes the bearing raceway system of a high-speed train as an application object, and solves the problem of thermal crown caused by uneven temperature distribution in the long-term operation process. A large amount of heat is generated in the operation process of the bearing of the high-speed train, and uneven temperature distribution will cause geometric deformation of the raceway surface, affecting the service life of the bearing and the safety of train operation. The embodiment realizes early discovery and prevention of the thermal crown problem of the bearing raceway by constructing an intelligent system with knowledge accumulation and learning ability.

[0237] The initial fault knowledge graph contains 150 fault type nodes and 320 thermal anomaly mode nodes. The fault types include raceway surface wear, uneven deformation of bearing steel ball, lubricant degradation, surface micro-cracks, etc. The thermal anomaly modes include various characteristic temperature distribution patterns. Through expert experience and historical fault case analysis, about 1200 association edges are established, of which there are 450 high-confidence association edges (weight > 0.8). The constructed knowledge graph is stored in a distributed database in the form of a graph data structure.

[0238] Twelve high-precision infrared thermographs are installed on the train bearing system, and a thermal image is collected every 30 seconds with a resolution of 640x480 pixels. The collected images are preprocessed through noise filtering, contrast standardization, and geometric correction. Taking a bearing abnormal temperature rise event as an example, the anomaly score calculated by the variational autoencoder model reaches 0.78 (the adaptive threshold is set to 0.65), successfully locating the coordinate position where the hot spot area is concentrated in the right side area of the inner ring of the raceway.

[0239] For the detected abnormal area, the image stream processing component of the dual-stream causal reasoning network extracts the spatial features and the near-24-hour time series features of the area through its spatial feature extraction unit and time series feature extraction unit, respectively. The feature fusion unit fuses these two features to obtain a 372-dimensional feature vector. At the same time, the knowledge graph structure stream component samples a local subgraph related to the bearing from the fault knowledge graph through the subgraph sampling unit, which contains 27 fault type nodes and 52 thermal anomaly mode nodes, and performs importance weighting through the node weighting unit. The cross-modal attention unit of the dual-stream feature fusion component aligns and fuses the image features and the graph features, and the causal reasoning unit infers the fault cause based on the fused features.

[0240] After dimension unification and L2 norm normalization preprocessing of the feature vector, the system uses the cosine similarity function to calculate the similarity between the new hot anomaly pattern and each node in the fault knowledge graph. Through the correlation strength evaluation algorithm based on the attention mechanism, the correlation degree between the new hot anomaly pattern and the existing fault type is quantified. The highest correlation strength value is 0.53, which is lower than the set threshold value 0.65, and the system judges it as a potential new fault pattern. The correlation strength distribution shows unimodal characteristics rather than multimodal characteristics, so it does not belong to a composite fault pattern.

[0241] The system creates a new node for the recognized new hot anomaly pattern in the fault knowledge graph and assigns a unique identifier "HATM_20230524_013", using the detected abnormal area features as node feature representation. After three weeks of operation observation and data accumulation, this anomaly pattern has established a high-strength association (0.82) with the "lubricant carbonization deposition" fault type, and the system has created a new fault association edge. The system also maintains the consistency of the graph, removes several low-confidence association edges, and recalculates the degree centrality and closeness centrality indicators of the nodes, optimizing the graph structure. Through incremental learning updates, the system updates the parameters of the bi-flow causal reasoning network.

[0242] Based on the updated fault knowledge graph and historical raceway system fault case data, the time series evolution model predicts the development trend of this type of fault. The time series feature extraction component extracts time series features from historical data, normalizes the data with a time window, and adjusts for seasonality. The causal chain analysis component locates the possible fault propagation path in the fault knowledge graph through the path tracking unit and finds that "lubricant carbonization deposition" may cause "raceway surface wear" to intensify. The multi-factor fusion component combines parameters such as bearing temperature, speed, and environmental humidity to predict the fault development timeline. Finally, the multi-level warning component generates warning information, suggesting that bearing inspection and lubricant replacement should be performed within the next 31.4 hours to avoid more serious faults.

[0243] This embodiment is applied to the bearing system of 120 EMUs in a section of Beijing-Shanghai High-speed Railway, and the technical effect is formed after one year of operation:

[0244] The initial fault knowledge graph node statistics of the system are shown in Table 1:

[0245] Table 1: Initial fault knowledge graph node statistics of the system

[0246]

[0247] The comparison results of the identification performance of different anomaly detection methods for new faults are shown in Table 2:

[0248] Table 2: Comparison of identification performance of different anomaly detection methods for new faults

[0249]

[0250] The comparison results of the new fault recognition rate of the embodiment and the traditional method in different operation stages are shown in Table 3:

[0251] Table 3: Comparison of new fault recognition rate (%) of the embodiment and the traditional method in different operation stages

[0252]

[0253] The comparison of the recognition performance of different methods on composite faults is shown in Table 4:

[0254] Table 4: Comparison of recognition performance of different methods on composite faults

[0255]

[0256] The statistical results of the composite fault early warning lead time are shown in Table 5:

[0257] Table 5: Composite fault early warning lead time statistics (hours)

[0258]

[0259] From the above data, it can be seen that the double-flow causal reasoning network and the graph expansion algorithm of the embodiment improve the adaptability to new faults and the recognition ability of composite faults. Especially in the later stage of system operation, with the continuous expansion and optimization of the fault knowledge graph, the new fault recognition rate reaches 94.2%, which is 38.4 percentage points higher than that of the traditional method. The recognition accuracy of composite faults reaches 86.7%, and the early warning can be given 31.4 hours in advance, which provides sufficient preparation time for bearing raceway system maintenance, effectively prevents potential equipment damage and safety accidents.

[0260] The embodiments of the present application are described above, but the embodiments are not limited to the specific implementation described above, and the specific implementation described above is only illustrative, not restrictive, and those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the embodiments.

Claims

1. A deep learning-based method for detecting raceway anomalies, characterized in that, include: Construct an initial fault knowledge graph to establish the association between known thermal anomaly patterns and fault types; Based on the initial fault knowledge graph, real-time thermal image sequences of the raceway system are collected, and anomaly scores and region localization are performed on the collected images through a variational autoencoder. By utilizing anomaly scores and region localization results, a dual-stream causal inference network is designed to achieve simultaneous processing of image data streams and knowledge graph structure streams. The dual-stream causal inference network comprises the following key components: an image data stream branch responsible for extracting spatial-temporal features from thermal images, including using convolutional neural networks to extract spatial features and long short-term memory networks to extract temporal features; a knowledge graph structure stream branch performing representation learning and feature extraction on the fault knowledge graph through graph convolutional networks; and features from the two branches are aligned and fused through a cross-modal attention mechanism to ultimately achieve inference analysis of fault causes. The dual-stream causal inference network further includes a dual-stream feature fusion component, which includes: Cross-modal attention units enable the alignment and fusion of image features and spectral features; The causal reasoning unit infers potential causes of failure based on fused features; The prediction output unit generates thermal anomaly patterns and their probability distributions. Based on the detected thermal anomaly patterns, analyze the correlation strength between the thermal anomaly patterns and existing nodes in the fault knowledge graph; Based on the correlation strength analysis results, a graph expansion algorithm is applied to achieve automatic updating of the fault knowledge graph; Based on the automatic update results of the fault knowledge graph and time-series data, time-series evolution analysis is carried out to predict potential fault development trends; the steps of acquiring real-time thermal image sequences of the raceway system and calculating anomaly scores and locating regions in the acquired images using a variational autoencoder include: Deploy thermal imaging equipment to acquire thermal image sequences of the raceway system at a fixed frequency; The acquired images are preprocessed, including noise filtering, contrast normalization, and geometric correction. Calculate the anomaly score for each image region using the reconstruction error; Set an adaptive threshold and mark the location of areas that exceed the threshold; The steps of the time-series evolution analysis include: Construct a time-series evolution model to capture the temporal evolution patterns of failure modes; Infer fault propagation paths by utilizing causal relationship chains in fault knowledge graphs; By combining equipment operating parameters and environmental factors, the potential timeline of failure development can be predicted; Generate multi-level early warning information, including short-term failure risk and long-term system health status prediction.

2. The method for detecting raceway anomalies based on deep learning according to claim 1, characterized in that, The steps for constructing the initial fault knowledge graph include: Collect historical data on raceway system failures; The collected data is preprocessed, including thermal image feature standardization and fault type classification coding; Establish fault type nodes and thermal anomaly mode nodes; Based on expert experience and historical data analysis, establish the connection edges between nodes.

3. The method for detecting raceway anomalies based on deep learning according to claim 1, characterized in that, The dual-stream causal inference network includes an image stream processing component, which includes: The spatial feature extraction unit is composed of a multi-layer convolutional neural network; The temporal feature extraction unit consists of a long short-term memory network and a temporal convolutional network. The feature fusion unit is used to fuse spatial features and temporal features.

4. The method for detecting raceway anomalies based on deep learning according to claim 1, characterized in that, The dual-stream causal reasoning network further includes a knowledge graph structure flow component, which includes: The graph representation unit is composed of graph convolutional networks; The subgraph sampling unit is responsible for performing local subgraph sampling in the fault knowledge graph based on the detected abnormal regions; The node weighting unit, which includes an attention mechanism module, weights the importance of nodes in the subgraph.

5. The method for detecting raceway anomalies based on deep learning according to claim 1, characterized in that, The steps for analyzing the correlation strength between thermal anomaly patterns and existing nodes in the fault knowledge graph include: Calculate the similarity between the feature vector of the new thermal anomaly pattern and each node in the fault knowledge graph; An attention-based correlation strength evaluation algorithm is used to quantify the correlation between new thermal anomaly patterns and existing fault types; If the highest correlation strength is lower than a set threshold, it is marked as a potential new failure mode; If the correlation strength distribution exhibits multi-peak characteristics, it is marked as a potential composite failure mode.

6. The method for detecting raceway anomalies based on deep learning according to claim 1, characterized in that, The steps of the graph expansion algorithm include: For newly identified fault modes, create new hot anomaly mode nodes in the fault knowledge graph; Assign a unique identifier to the new node and use the detected abnormal region features as the node feature representation; If the correlation strength analysis results indicate the existence of a new fault type, then a new fault type node is created. Establish connections between the new node and related existing nodes, and set initial weights.

7. A deep learning-based raceway anomaly detection system, characterized in that, A method for performing a deep learning-based rollway anomaly detection according to any one of claims 1-6 includes: The fault knowledge graph construction module is used to construct an initial fault knowledge graph containing the association between known thermal anomaly patterns and fault types. The thermal image anomaly detection module is used to acquire real-time thermal image sequences of the raceway system and to calculate anomaly scores and locate areas using a variational autoencoder. The dual-stream causal reasoning module is used to design a dual-stream causal reasoning network to achieve synchronous processing of image data streams and knowledge graph structure streams; The correlation strength analysis module is used to analyze the correlation strength between the detected thermal anomaly patterns and existing nodes in the fault knowledge graph; The automatic graph update module is used to automatically update the fault knowledge graph based on the correlation strength analysis results and by applying the graph expansion algorithm. The fault trend prediction module is used to conduct time-series evolution analysis based on the automatic update results of the fault knowledge graph and time-series data, and to predict potential fault development trends.

Images