Method and system for designing thermal convexity of raceway based on deep learning

Abstract

The invention relates to the technical field of raceway system thermal convexity detection, and discloses a raceway thermal convexity design method and system based on deep learning, and the method comprises the steps: constructing an initial fault knowledge graph, and forming an incidence relation between a known thermal anomaly mode and a fault type; a real-time thermal image sequence of the raceway system is collected, and abnormal score calculation and region positioning are carried out on the collected images through a variational auto-encoder; a double-flow causal reasoning network is designed, and synchronous processing of the image data flow and the knowledge graph structure flow is achieved; analyzing association strength between the thermal anomaly mode and existing nodes in the fault knowledge graph; a graph extension algorithm is applied to realize automatic updating of the fault knowledge graph; time sequence evolution analysis is carried out, and a potential fault development trend is predicted; by designing the double-flow causal reasoning network and the atlas extension algorithm, the continuous accumulation and evolution ability of the fault knowledge of the raceway system is realized.

Description

Technical Field

[0001] This invention relates to the field of raceway system thermal convexity detection technology, and more specifically, to a raceway thermal convexity design method and system based on deep learning. Background Technology

[0002] With the continuous improvement of industrial automation and intelligent manufacturing, the safety and reliability of the roller conveyor system, as a key mechanical transmission component, have received widespread attention.

[0003] In existing technologies, anomaly detection methods for raceway systems mainly rely on rule-based or traditional machine learning static models. These methods are typically trained on limited historical data before system deployment, making it difficult to dynamically adapt to the evolution of faults caused by equipment aging, changes in operating conditions, or the emergence of new fault modes during actual operation. Traditional methods generally lack the ability to continuously accumulate and update fault knowledge, failing to fully explore and utilize the causal relationships inherent in historical cases, resulting in limited ability to identify novel, rare, and complex faults. Furthermore, existing anomaly detection methods suffer from insufficient detection accuracy and generalization ability when facing complex and ever-changing operating environments, making it difficult to meet the actual needs of long-term stable operation and intelligent maintenance of raceway systems.

[0004] Therefore, there is an urgent need for a raceway system anomaly detection technology that can achieve adaptive learning of fault knowledge, dynamic expansion, and improved ability to identify new faults, so as to ensure the safety and reliability of its operation. Summary of the Invention

[0005] This invention provides a deep learning-based method and system for designing raceway thermal convexity, which solves the technical problems of traditional anomaly detection methods in related technologies, such as lack of fault knowledge accumulation and learning ability, inability to adapt to fault mode evolution, and insufficient ability to identify new and compound faults.

[0006] This invention provides a deep learning-based method for designing raceway thermal convexity, comprising: 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 two-stream causal reasoning network is designed to achieve synchronous processing of image data streams and knowledge graph structure streams; 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.

[0007] Furthermore, the step of constructing the initial fault knowledge graph includes: 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.

[0008] Furthermore, the step of acquiring the real-time thermal image sequence of the raceway system includes: 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.

[0009] Furthermore, 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.

[0010] Furthermore, the dual-stream causal reasoning network also 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.

[0011] Furthermore, the dual-stream causal inference network also 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 the anomaly type prediction results and their probability distribution.

[0012] Furthermore, the step of analyzing the correlation strength between thermal anomaly patterns and existing nodes in the fault knowledge graph includes: 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.

[0013] Furthermore, 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.

[0014] Furthermore, 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.

[0015] This invention provides a deep learning-based raceway thermal convexity design system for executing the aforementioned deep learning-based raceway thermal convexity design method, comprising: 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.

[0016] The beneficial effects of this invention are as follows: by combining the abnormal features of the image space with the semantic relationships in the knowledge graph through the dual-stream causal reasoning network, this invention can learn fault knowledge from historical cases, and automatically add newly discovered abnormal patterns and their causal relationships to the knowledge graph through the graph expansion algorithm, thereby realizing the continuous accumulation of fault knowledge. Compared with static anomaly detection models, this invention has a stronger ability to identify novel faults. By analyzing the correlation strength between new hot anomaly patterns and existing nodes in the fault knowledge graph, this invention can determine whether a new type of fault has occurred and incorporate it into the knowledge system, thereby enhancing its adaptability to the evolution of fault modes caused by technological iteration and environmental changes. The association strength evaluation algorithm based on the attention mechanism of the present invention can identify composite fault modes with multi-peak characteristics in the association strength distribution, making up for the shortcomings of traditional methods in composite fault identification and improving the system's ability to understand complex fault situations. Through time-series evolution analysis, this invention can predict potential fault development trends and new fault modes, realizing the transformation from passive response to proactive prevention, and providing more forward-looking support for the maintenance decision-making of the raceway system. The knowledge graph structure of this invention provides a visual representation of fault types, causes, symptoms and their interrelationships, enhancing interpretability and enabling maintenance personnel to better understand the causes and development mechanisms of faults. Attached Figure Description

[0017] Figure 1 This is a flowchart of a deep learning-based method for designing raceway thermal convexity in this invention. Figure 2 It is a line graph comparing the changing trends of novel fault identification rates of different fault detection methods at different stages of system operation; Figure 3 This is a bar chart comparing the accuracy of compound fault identification and the correctness of fault component analysis between traditional anomaly detection methods, single deep learning models, and the method of this invention. Figure 4 It is a bar chart showing the change in the number of nodes in the fault knowledge graph at different stages after the invention is implemented; Figure 5 This is a bar chart comparing the fault warning advance time of the traditional method and the method of this embodiment under different combinations of fault types. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a method and system for designing raceway thermal convexity based on deep learning, such as... Figure 1 As shown, it includes: Step 1: Construct an initial fault knowledge graph to establish the association between known thermal anomaly patterns and fault types; This step constructs an initial knowledge graph containing the associations between known thermal anomaly patterns and fault types. Specifically, it includes: Step 1.1: Collect historical data on raceway system failures; Historical raceway system failure case data includes information such as failure type, thermal image characteristics, environmental parameters, and operating status; Step 1.2: Preprocess the collected data; Data preprocessing includes standardizing thermal image features and classifying and encoding fault types; Furthermore, the data preprocessing includes: Thermal image feature standardization: Subtract the mean from the pixel values ​​of the thermal image and divide by the standard deviation to convert it into a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby eliminating numerical differences under different acquisition devices and environmental conditions; Thermal image pixel value preprocessing: Thermal image pixel values ​​need to be standardized to unify pixel values ​​from different acquisition devices to the same range. For contrast standardization, an adaptive thresholding method should be used, dynamically adjusting the mapping parameters according to the image content. Pixel values ​​should also undergo outlier detection and processing to remove obvious noise and outliers. Fault type classification coding: One-hot encoding (OHE) is used to convert text-based fault types into numerical vector representations, which facilitates subsequent data processing and model training; Fault Type Classification Data Preprocessing: Categorical data such as fault types requires encoding preprocessing. One-hot encoding is used to convert text-based fault types into numerical vector representations. For newly added fault types, a dynamic encoding mechanism should be established to ensure consistency and scalability of the encoding. The encoded vectors should undergo sparsity checks to avoid the curse of dimensionality. Time-series data alignment: Based on timestamps, thermal image sequences acquired at different time points are interpolated and resampled to ensure consistent data time intervals; Step 1.3: Establish fault type nodes and thermal anomaly mode nodes; Fault type nodes represent different categories of faults, while thermal anomaly mode nodes represent specific thermal distribution characteristics. Step 1.4: Based on expert experience and historical data analysis, establish the connection edges between nodes; The weight of an edge represents the strength of the association; The weights of associated edges need to undergo range-limiting preprocessing, restricting their values ​​to the interval [0, 1] to ensure the reasonableness of the weight values. The initial weights of newly added edges should be set based on the similarity calculation results. Weight updates should be normalized to avoid weight values ​​accumulating to excessively high or low values ​​over time. The constructed knowledge graph is stored as a graph data structure, which facilitates access and updates by subsequent algorithms.

[0020] Step 2: Based on the initial fault knowledge graph, collect real-time thermal image sequences of the raceway system, and use a variational autoencoder to calculate anomaly scores and locate regions in the collected images. This step uses a variational autoencoder (VAE) to perform anomaly detection and region localization on the acquired thermal image sequence. Specifically, it includes: Step 2.1, Deploy thermal imaging equipment; High-precision infrared thermal imagers should be installed at key locations in the raceway system, including around critical components such as the bearing outer ring, inner ring, and cage. The thermal imaging equipment should possess performance specifications including high resolution (no less than 640×480 pixels), high temperature accuracy (±0.5℃), and a high sampling frequency (no less than 30 seconds / sample). The equipment installation location should ensure complete coverage of the thermal distribution area of ​​the raceway system, avoiding blind spots and obstructions. A data acquisition and transmission system should also be configured to ensure that thermal image data can be transmitted to the central processing unit in real time for subsequent analysis.

[0021] Acquire thermal image sequences of the raceway system at a fixed frequency; Step 2.2: Preprocess the acquired images; Preprocessing of the acquired images includes noise filtering, normalization, and temporal alignment; Furthermore, the preprocessing of the acquired images includes: Noise filtering: Use methods such as Gaussian filtering or median filtering to remove random noise and outliers from thermal images; Contrast normalization: linearly maps the pixel values ​​of the thermal image to the [0, 1] interval, enhancing the identifiability of thermal anomaly areas; Geometric correction: Corrects perspective distortion and aberration in images to ensure that thermal images acquired at different times are spatially aligned; Step 2.3: Construct and train the variational autoencoder model; Furthermore, the variational autoencoder (VAE) model includes: 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; Latent spatial components: a low-dimensional feature vector space used to represent thermal images, defined by the mean vector and the variance vector; 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; Loss function component: Composed of reconstruction loss and KL divergence loss, used to optimize model training; Step 2.4: Calculate the anomaly score for each image region using the reconstruction error; Furthermore, a higher anomaly score indicates a greater likelihood of an anomaly, and the formula for calculating the anomaly score is as follows: ; 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. Step 2.5: Set an adaptive threshold, locate and mark areas exceeding the threshold, and obtain the coordinates of suspected abnormal areas.

[0022] Furthermore, the formula for calculating the adaptive threshold is: ; in, Indicates an adaptive threshold. This represents the mean of the anomaly scores for the entire image region. The standard deviation of the anomaly score for the entire image region. The sensitivity coefficient is adjustable. Adjustable sensitivity coefficient Normalization preprocessing is required to limit its value range to a reasonable range, typically between 1.0 and 3.0, to avoid abnormal threshold calculations due to excessively large or small values. This is done during system initialization. The value should be determined based on historical data statistical analysis and dynamically adjusted according to the detection results during operation; The specific implementation is as follows: calculate the arithmetic mean of the anomaly scores of all pixel locations in the entire image region, which is obtained by summing all anomaly scores and dividing by the total number of pixels, and is used to characterize the overall anomaly level of the image.

[0023] The specific implementation is as follows: calculate the standard deviation of the outlier scores by summing the squares of the differences between each outlier score and the mean, dividing by the total number of pixels, and then taking the square root. This standard deviation is used to characterize the dispersion and variation of the outlier scores. The larger the value, the more uneven the outlier distribution.

[0024] Step 3: Using the anomaly scores and region localization results, design a two-stream causal reasoning network to achieve synchronous processing of image data streams and knowledge graph structure streams; This step is the core innovation of this implementation method, which designs a dual-stream causal reasoning network that can simultaneously process image data streams and knowledge graph structure streams.

[0025] Furthermore, the dual-stream causal reasoning network includes the following components: Image stream processing components: Spatial feature extraction unit: Composed of a multi-layer convolutional neural network (CNN), used to extract spatial features from thermal images; Temporal feature extraction unit: Composed of a Long Short-Term Memory (LSTM) network or a Temporal Convolutional Network (TCN), used to extract temporal features from thermal image sequences; Feature fusion unit: used to fuse spatial features and temporal features to generate a comprehensive image feature vector.

[0026] Knowledge graph structure flow component: Graph representation unit: Composed of a graph convolutional network (GCN), used for learning representations of knowledge graphs; Subgraph Sampling Unit: Responsible for performing local subgraph sampling in the knowledge graph based on detected abnormal regions; Node weighting unit: contains an attention mechanism module, which weights the importance of nodes in the subgraph to generate structured feature vectors.

[0027] Dual-stream feature fusion component: Cross-modal attention unit: Enables alignment and fusion of image features and spectral features; Causal reasoning unit: Infers potential causes of failure based on fused features; Prediction output unit: Generates the prediction results of anomaly types and their probability distributions.

[0028] Step 4: Combine the detected thermal anomaly patterns to analyze the correlation strength between the thermal anomaly patterns and existing nodes in the fault knowledge graph. This step, based on the reasoning results of step 3, analyzes the association between the detected new thermal anomaly patterns and existing nodes in the fault knowledge graph. Specifically, it includes: like Figure 2 As shown in Figure 3, the novel fault identification rate of different fault detection methods changes at different stages of system operation (initial stage, middle stage, and late stage), reflecting the technical effect of the method in this embodiment as its knowledge accumulation capability improves. As shown in Figure 3, the novel fault identification rate of the method in this embodiment is 78.6% in the initial stage (1-3 months), increases to 86.7% in the middle stage (4-6 months), and reaches 94.2% in the late stage (7-12 months), which is 20.4, 29.2, and 38.4 percentage points higher than the traditional static model, respectively, fully demonstrating the novel fault adaptability of this method.

[0029] Step 4.1: Calculate the similarity between the feature vector of the new thermal anomaly pattern and each node in the fault knowledge graph; Furthermore, the feature vectors are preprocessed before calculating similarity: Dimensional unification: Through feature mapping or dimensionality reduction techniques, ensure that the feature vectors of new hot anomaly patterns have the same dimension as the feature vectors of nodes in the fault knowledge graph; L2 norm normalization: Normalizes the length of the eigenvectors so that the Euclidean norm of all eigenvectors is 1, eliminating the influence of amplitude differences and comparing only directional similarity. Furthermore, the similarity calculation uses the cosine similarity function, and its calculation formula is as follows: ; in, The feature vector representing the new thermal anomaly pattern, This represents the feature vector of a node in a fault knowledge graph. The Euclidean norm of a vector. express and The cosine similarity between them ranges from 1 to 10. The larger the value, the more similar the two feature vectors are.

[0030] Step 4.2: Use an attention-based correlation strength evaluation algorithm to quantify the correlation between new thermal anomaly patterns and existing fault types; Furthermore, the steps of the association strength evaluation algorithm include: Calculate the feature similarity between the feature vector of the new thermal anomaly pattern and the feature similarity of each fault type node in the fault knowledge graph; Similarity is weighted based on attention weights to highlight the impact of key features; Furthermore, the formula for calculating the attention weight is as follows: ; in, Indicates the first Attention weights for each feature and They represent the first The first feature and the first Attention scores for each feature Represents the total number of features. Represents an exponential function; Attention score The score is calculated by the dot product of the feature vector and the trainable attention query vector. It requires normalization preprocessing to limit its numerical range to a reasonable interval, avoiding exponential overflow due to excessively large values. Normalization is typically performed using the softmax function to ensure that the sum of all attention weights is 1. During the calculation, abnormally large scores should be truncated. By applying a nonlinear transformation function to map the weighted similarity to the interval [0, 1], the association strength value is obtained; Furthermore, the nonlinear transformation function is the Sigmoid function, and its calculation formula is: ; in, Indicates the correlation strength value. Indicates weighted similarity. It is a parameter that controls the slope of the function. It is the center point parameter of the function. The base of the natural logarithm is approximately 2.71828; Preprocessing with range limitations is required, typically set 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 model training. Center point calibration preprocessing is required, with a value typically ranging from -2.0 to 2.0, used 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; The fault type can be determined based on the characteristics of the correlation strength distribution: a single-peaked distribution corresponds to a single fault, while a multi-peaked distribution corresponds to a compound fault. If the highest correlation strength is lower than the 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.

[0031] like Figure 3 As 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.

[0032] 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.

[0033] Step 5: Based on the correlation strength analysis results, apply the graph expansion algorithm to achieve automatic updating of the fault knowledge graph; 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.

[0034] 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.

[0035] Furthermore, the steps of the graph expansion algorithm include: Step 5.1, Node creation and association steps; For the novel fault modes identified in step 4, 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. 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.

[0036] Step 5.2, map consistency maintenance steps; Scan the relationship between newly added edges and existing edges to detect whether there are circular dependencies or logical contradictions; If a contradiction is found, retain the high-confidence relationship and remove the low-confidence relationship based on the confidence level. The weights of all edges are decayed over time, so that the weights of older associated edges gradually decrease. Furthermore, the formula for calculating the timing decay is as follows: ; 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; 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. 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. Automatic graph pruning is achieved by removing edges whose weights are below a set threshold.

[0037] Step 5.3, graph structure optimization steps; Calculate the degree centrality and proximity centrality indices of each node in the fault knowledge graph; Furthermore, the formula for calculating degree centrality is as follows: ; in, Represents a node Degree centrality, Represents a node The degree (the number of edges connected to it). This represents the total number of nodes in the fault knowledge graph; The specific implementation is as follows: statistics and nodes The number of all directly connected edges, including incoming and outgoing edges, is counted by traversing the graph's adjacency list or adjacency matrix.

[0038] Furthermore, the formula for calculating proximity centrality is: ; in, Represents a node Proximity centrality Represents a node To the node The shortest path length, This represents the total number of nodes in the fault knowledge graph; Indicates the summation symbol; Identify important nodes and edges, and construct the index structure of the graph; Preload frequently accessed subgraphs into the cache to improve query efficiency; Regularly rebalance the graph structure to maintain query performance.

[0039] The specific implementation involves using either Dijkstra's algorithm or the Floyd-Warshall algorithm to compute nodes. To the node The shortest path length is returned, and if no path exists between the two nodes, infinity is returned.

[0040] Step 5.4, Incremental learning and update steps; Extract feature representations of newly added nodes and edges; Furthermore, the extraction of the feature representation includes a data preprocessing step: Feature normalization: Normalize different types of features (such as thermal anomaly pattern features, fault type features, associated edge weights, etc.) to a uniform numerical range; Heterogeneous data fusion: Encode and unify the features of thermal images (continuous data) and fault types (discrete data) to form hybrid features that can be used for training; Time-weighted: Introduces a time decay factor into the feature representation, making more recent data have a greater impact on the model; Feature representation preprocessing: All feature representations need to undergo consistency checks 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 the curse of dimensionality. Use these features for incremental training to update the parameters of the two-stream causal inference network; The structure of the graph convolutional network (GCN) is dynamically adjusted according to changes in the graph size. Validate the updated model to ensure that its performance is not lower than that of the original model.

[0041] Step 6: Based on the automatic update results of the fault knowledge graph and time series data, conduct time series evolution analysis to predict potential fault development trends; This step predicts fault development trends based on the updated fault knowledge graph and historical raceway system fault case data. Specifically, it includes prediction and analysis using a time-series evolution model.

[0042] like Figure 5 As shown, the comparison of fault warning lead time between the traditional method and the method of this embodiment under different fault type combinations highlights the advantage of this method in terms of timely warning. The average warning lead time for compound faults reaches 31.4 hours, which is 25.2 hours higher than the 6.2 hours of the traditional method. Specifically, the warning lead time for the combination of "lubricant degradation + surface microcracks" reaches 32.7 hours, and the warning lead time for the combination of "surface contamination + lubricant degradation" reaches 36.2 hours. This provides sufficient preparation time for the maintenance of the bearing raceway system and effectively prevents potential equipment damage and safety accidents.

[0043] Furthermore, the time-series graph evolution model includes the following components: Time-series feature extraction component: responsible for extracting time-series features from historical raceway system failure case data to capture the evolution patterns of failure modes; Furthermore, the historical raceway system failure case data needs to be preprocessed before the time series feature extraction: Time window standardization: Truncate or pad fault case data of different lengths to a uniform time window length; Time series data preprocessing: Time series data requires window standardization preprocessing to truncate or pad fault case data of different lengths to a uniform time window length. For seasonal data, seasonal adjustments should be made to remove the influence of periodic variations. When extracting multi-scale time features, data at different time granularities should be aligned. Seasonal adjustment: Remove seasonal factors from the time series, such as periodic changes in equipment heat load; Multi-scale time feature extraction: Simultaneously extract features from different time granularities (hours, days, months) to comprehensively capture short-term fluctuations and long-term trends; Causal chain analysis component: Path tracing unit: Locates and traces possible fault propagation paths in the fault knowledge graph; Influence assessment unit: Calculates the influence of each node in the causal chain; Propagation prediction unit: Predicts the possible direction and speed of fault propagation in the system.

[0044] Multi-factor fusion component: Parameter integration unit: Integrates equipment operating parameters and environmental factors into the prediction model; Furthermore, the parameter integration involves the preprocessing of heterogeneous data: Dimensional unification: Parameters with different dimensions (such as temperature, pressure, speed, etc.) are transformed into the [0, 1] interval through min-max normalization; Preprocessing of equipment operating parameters and environmental factors: Equipment operating parameters such as temperature, pressure, and speed need to be preprocessed to unify dimensions, transforming all parameters to the [0, 1] interval through min-max normalization. Percentage parameters such as ambient humidity need to be converted to decimal form. The weighting coefficients of different parameters should be determined based on expert knowledge and historical data analysis to ensure that the contribution of each parameter in the prediction model is reasonable; Weight calibration: Based on domain expert knowledge and historical data analysis, assign reasonable weight coefficients to different parameters; Missing value handling: Missing data in equipment operating parameters and environmental factors are handled through methods such as interpolation or averaging. Weighting adjustment unit: Dynamically adjusts the weight of different factors in the prediction based on their importance; Timeline generation unit: Generates a detailed timeline of the development of potential faults.

[0045] Multi-level early warning components: Short-term risk assessment unit: Assess the probability and severity of near-term failures; Long-term trend analysis unit: Predicts the changing trend of the system's long-term health status; Warning message formatting unit: Generates customized warning messages based on different user roles.

[0046] A deep learning-based raceway thermal convexity design system is used to execute the aforementioned deep learning-based raceway thermal convexity design method, comprising: 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.

[0047] Here, this implementation provides an application example: This embodiment focuses on the raceway system of high-speed train bearings, addressing the thermal convexity problem caused by uneven temperature distribution during long-term operation. High-speed train bearings generate significant heat during operation; uneven temperature distribution leads to geometric deformation of the raceway surface, affecting bearing life and train operation safety. This embodiment utilizes an intelligent system with knowledge accumulation and learning capabilities to achieve early detection and prevention of bearing raceway thermal convexity problems.

[0048] The initial fault knowledge graph contains 150 fault type nodes and 320 thermal anomaly pattern nodes. Fault types include raceway surface wear, uneven deformation of bearing steel balls, lubricant degradation, and surface microcracks. Thermal anomaly patterns include various characteristic temperature distribution patterns. Through expert experience and historical fault case analysis, approximately 1200 association edges were established, of which 450 are 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.

[0049] Twelve high-precision infrared thermal imagers were installed on the train bearing system, acquiring thermal images every 30 seconds with a resolution of 640×480 pixels. The acquired images underwent preprocessing steps such as noise filtering, contrast normalization, and geometric correction. Taking an abnormal bearing temperature rise event as an example, the variational autoencoder model calculated an anomaly score of 0.78 (with a set adaptive threshold of 0.65), successfully locating the hotspot area concentrated on the right side of the inner ring of the raceway.

[0050] For the detected abnormal regions, the image stream processing component of the dual-stream causal inference network extracts spatial features and temporal features over the past 24 hours through its spatial feature extraction unit and temporal feature extraction unit, respectively. The feature fusion unit then fuses these two types of features to obtain a 372-dimensional feature vector. Simultaneously, the knowledge graph structure stream component samples local subgraphs related to the bearing from the fault knowledge graph through its subgraph sampling unit. These subgraphs contain 27 fault type nodes and 52 thermal anomaly pattern nodes, and their importance is weighted by the node weighting unit. The cross-modal attention unit of the dual-stream feature fusion component aligns and fuses the image features with the graph features, and the causal inference unit infers the cause of the fault based on the fused features.

[0051] After preprocessing the feature vectors by unifying their dimensions and normalizing them to L2 norm, the system uses a cosine similarity function to calculate the similarity between the new hot anomaly pattern and each node in the fault knowledge graph. An attention-based association strength evaluation algorithm quantifies the degree of association between the new hot anomaly pattern and existing fault types. The highest association strength value is 0.53, which is below the set threshold of 0.65, and the system identifies it as a potential new fault pattern. The association strength distribution exhibits a unimodal characteristic rather than a multimodal characteristic, therefore it does not belong to a composite fault pattern.

[0052] The system creates new nodes in the fault knowledge graph for newly identified thermal anomaly patterns and assigns them a unique identifier, "HATM_20230524_013," using the detected anomaly region features as node feature representations. After three weeks of operation and observation, this anomaly pattern established a high-strength association (0.82) with the "lubricant carbonization deposition" fault type, and the system created new fault association edges. The system also performed consistency maintenance on the graph, removing several low-confidence association edges and recalculating the degree centrality and proximity centrality indices of the nodes, thus optimizing the graph structure. Through incremental learning updates, the system updated the parameters of the two-stream causal inference network.

[0053] Based on the updated fault knowledge graph and historical raceway system fault case data, the time-series evolution model predicted the development trend of this type of fault. The time-series feature extraction component extracted time-series features from historical data and performed time window standardization and seasonal adjustments. The causal chain analysis component, through the path tracing unit, located possible fault propagation paths in the fault knowledge graph and found that "lubricant carbonization deposition" might exacerbate "raceway surface wear." The multi-factor fusion component combined parameters such as bearing temperature, speed, and ambient humidity to predict the fault development timeline. Finally, the multi-level early warning component generated warning information, recommending bearing inspection and lubricant replacement within the next 31.4 hours to prevent more serious faults.

[0054] This embodiment was applied to the bearing system of 120 EMU train sets on a section of the Beijing-Shanghai High-Speed ​​Railway, and after one year of operation, the following technical results were achieved: The statistical results of the initial fault knowledge graph nodes of the system are shown in Table 1: Table 1: Statistics of Nodes in the System Initial Fault Knowledge Graph

[0055] Table 2 shows the comparison results of the identification performance of different anomaly detection methods for novel faults: Table 2: Comparison of the identification performance of different anomaly detection methods for novel faults

[0056] The comparison results of the novel fault identification rate of this embodiment with the traditional method at different operating stages are shown in Table 3: Table 3: Novel Fault Identification Rate (%) of this Embodiment and Traditional Methods at Different Operating Stages

[0057] Table 4 shows a comparison of the performance of different methods in identifying complex faults: Table 4: Comparison of the identification performance of different methods for compound faults

[0058] The statistical results of the advance warning of composite faults are shown in Table 5: Table 5: Statistics on the Lead Time for Complex Fault Warnings (in hours)

[0059] The data above demonstrates that the dual-stream causal reasoning network and graph expansion algorithm in this embodiment improve the adaptability to novel faults and the ability to identify complex faults. Particularly in the later stages of system operation, with the continuous expansion and optimization of the fault knowledge graph, the identification rate of novel faults reached 94.2%, an improvement of 38.4 percentage points compared to traditional methods. The accuracy rate for identifying complex faults reached 86.7%, and it can issue warnings an average of 31.4 hours in advance, providing ample preparation time for the maintenance of the bearing raceway system and effectively preventing potential equipment damage and safety accidents.

[0060] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A deep learning-based method for designing raceway thermal convexity, 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. 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.

2. The deep learning-based method for designing raceway thermal convexity 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 deep learning-based method for designing raceway thermal convexity according to claim 1, characterized in that, The steps for acquiring real-time thermal image sequences of the raceway system 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.

4. The deep learning-based method for designing raceway thermal convexity 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.

5. The deep learning-based method for designing raceway thermal convexity 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.

6. The deep learning-based method for designing raceway thermal convexity according to claim 1, characterized in that, 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 the anomaly type prediction results and their probability distribution.

7. The deep learning-based method for designing raceway thermal convexity 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.

8. The deep learning-based method for designing raceway thermal convexity 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.

9. The deep learning-based method for designing raceway thermal convexity according to claim 1, characterized in that, 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.

10. A deep learning-based raceway thermal convexity design system, characterized in that, A method for implementing a deep learning-based raceway thermal convexity design according to any one of claims 1-9 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