Knowledge graph embedding-based industrial device event causal tracing method and system

Abstract

This invention relates to the field of industrial equipment fault diagnosis technology, specifically to a knowledge graph-embedded method and system for causal tracing of industrial equipment events. The method includes the following steps: collecting multi-source heterogeneous event data during the operation of industrial equipment; preprocessing and extracting features from the collected multi-source heterogeneous event data to obtain event feature vectors. This invention constructs a knowledge graph in the industrial equipment domain, structurally representing the complex mapping relationships between entities such as equipment type, fault type, fault cause, and fault result, providing rich semantic prior knowledge for subsequent causal inference. Compared to traditional data-driven fault diagnosis methods, the knowledge graph introduced in this invention can encode domain experts' in-depth understanding of equipment fault mechanisms, making the tracing results more accurate and interpretable.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, specifically to a knowledge graph-embedded method and system for causal tracing of industrial equipment events. Background Technology

[0002] During long-term operation, industrial equipment can experience various malfunctions due to wear, aging, and improper operation. In recent years, the rapid development of sensor and data acquisition technologies has enabled the complete recording of vast amounts of operational data generated during industrial equipment operation. This data contains rich information on equipment operating status and fault evolution patterns. How to extract the inherent patterns of equipment failures from this massive amount of data and achieve rapid and accurate fault tracing has become a research hotspot in the field of intelligent operation and maintenance of industrial equipment.

[0003] Existing methods for industrial equipment fault diagnosis mainly include rule-based expert systems, statistical anomaly detection methods, and machine learning-based fault classification methods. Rule-based expert systems require domain experts to predefine a large number of fault rules; incomplete rule coverage makes it difficult to handle novel faults, and rule maintenance costs are high. Statistical anomaly detection methods typically only detect anomalies in single variables, making it difficult to capture the correlations between multiple variables in complex equipment faults. While machine learning-based fault classification methods can automatically learn fault patterns, they require a large amount of labeled training data, which is often scarce in industrial equipment fault samples, leading to poor model training performance. Furthermore, most existing methods remain at the fault detection level, lacking in-depth analysis of the causal chain of faults and failing to answer fundamental questions such as "how did the fault occur?" and "what is the root cause?"

[0004] Causal inference technology has become a research hotspot in the field of artificial intelligence in recent years, aiming to discover causal relationships, not just correlations, between variables. Applying causal inference technology to industrial equipment fault diagnosis can help maintenance personnel understand the underlying mechanisms of fault occurrence, trace fault propagation paths, and locate the root cause of faults. However, simple causal inference methods often lack sufficient domain knowledge support when dealing with objects like industrial equipment that have complex structures and multiple relationships, resulting in insufficient accuracy and interpretability of the tracing results. Knowledge graphs, as a structured knowledge representation method, can effectively organize and integrate domain expert knowledge, providing semantically rich prior information for causal inference. Therefore, how to organically combine knowledge graphs with causal inference technology to construct an accurate and interpretable industrial equipment fault tracing system is a technical problem that urgently needs to be solved in the field of industrial intelligence. Summary of the Invention

[0005] This disclosure proposes a knowledge graph-embedded causal tracing method and system for industrial equipment events, aiming to overcome at least one of the defects in the prior art.

[0006] To achieve the above objectives, the technical solution disclosed in this invention is as follows: According to one aspect of this disclosure, a method for causal attribution of industrial equipment events embedded with a knowledge graph is provided, comprising the following steps: Collect multi-source heterogeneous event data during the operation of industrial equipment. The multi-source heterogeneous event data includes equipment operating parameter data, sensor time series data, operation log data, and environmental parameter data. Data preprocessing and feature extraction are performed on the collected multi-source heterogeneous event data to obtain event feature vectors; Based on a pre-built knowledge graph of the industrial equipment domain, graph embedding representation learning is performed on the event feature vector to obtain the event knowledge embedding vector. The knowledge graph of the industrial equipment domain includes equipment type nodes, event type nodes, cause nodes, result nodes, and edge relationships used to connect nodes. Based on event knowledge embedding vectors, causal relationships between industrial equipment events are identified through a causal inference model, generating a causal relationship map. By tracing the root cause of an event using a causal relationship diagram, the fundamental cause of the event is determined and the tracing results are output.

[0007] Furthermore, the collection of multi-source heterogeneous event data during the operation of industrial equipment specifically includes: The equipment operation monitoring system collects equipment operation parameter data, including vibration signals, temperature signals, pressure signals, and flow signals. Sensor timing data is collected by sensor nodes deployed on industrial equipment; Obtain operation log data through the industrial equipment control system; Environmental parameter data are collected through an environmental monitoring module, including ambient temperature, ambient humidity, and ambient air pressure.

[0008] Furthermore, the method for constructing the knowledge graph in the industrial equipment domain includes: Collect historical failure case data of industrial equipment, including equipment type, failure type, failure cause, failure result and failure handling record; Extract the mapping relationship between equipment type and fault type, the causal relationship between fault type and fault cause, and the propagation relationship between fault cause and fault result from the historical fault case data; Establish a set of node types, including device type nodes, component status nodes, fault type nodes, cause nodes, and result nodes; Based on the extracted mapping and causal relationships, a set of edges connecting the nodes is constructed, and each edge is assigned a relation weight calculated based on the frequency of historical cases. A graph neural network is used to embed the initially constructed knowledge graph of the industrial equipment domain into the knowledge graph to obtain a vector representation of each node in the knowledge graph of the industrial equipment domain.

[0009] Furthermore, the step of identifying causal relationships between industrial equipment events based on the event knowledge embedding vector and using a causal inference model includes: The event knowledge embedding vector is input into a preset causal inference neural network model, which adopts a temporal causal network architecture based on an attention mechanism; The temporal encoder in the causal inference neural network model extracts temporal features from the event sequence to obtain temporal causal features. The causal association strength between event nodes is calculated through a causal attention layer, generating a causal adjacency matrix; Based on the causal adjacency matrix, causal path search is performed to determine the direct and indirect causal relationships between events.

[0010] Furthermore, the loss function of the causal inference neural network model includes: The causal cross-entropy loss term is used to measure the difference between the predicted causal relationship and the true causal relationship. The temporal consistency loss term is used to maintain the causal logical consistency of events in time sequence; The graph structure consistency loss term is used to maintain the distance relationship between the topological structure information in the knowledge graph of the industrial equipment domain in the embedding space.

[0011] Furthermore, the step of tracing the root cause of events based on the causal relationship graph includes: Based on the aforementioned causal relationship graph, starting from the target event node, a depth-first search-based tracing algorithm is used to trace the causal chain upstream. During the tracing process, the causal contribution of each upstream event node to the target event is calculated; The upstream events are sorted according to their causal contribution, and the event node with the highest causal contribution is determined as the root cause. Generate a source tracing results report that includes the root cause, causal propagation path, and source tracing confidence level.

[0012] Furthermore, it also includes: Based on the source tracing results, generate fault diagnosis suggestions; The source tracing results and the fault diagnosis suggestions are pushed to the industrial equipment management terminal; The source tracing results are used for feedback learning to update the edge relation weights in the knowledge graph of the industrial equipment domain.

[0013] Furthermore, the node types also include maintenance record nodes and quality indicator nodes.

[0014] Furthermore, the causal attention layer employs a multi-head self-attention mechanism, which includes multiple attention heads, each of which learns causal association patterns in different subspaces.

[0015] According to another aspect of this disclosure, a knowledge graph-embedded causal tracing system for industrial equipment events is provided, for implementing the knowledge graph-embedded causal tracing method for industrial equipment events as described above, comprising: The multi-source data acquisition module is used to collect multi-source heterogeneous event data during the operation of industrial equipment; The data preprocessing module, connected to the multi-source data acquisition module, is used to perform data preprocessing and feature extraction on the acquired multi-source heterogeneous event data to obtain event feature vectors. The graph embedding module, connected to the data preprocessing module, is used to perform graph embedding representation learning on the event feature vector based on a pre-built industrial equipment domain knowledge graph to obtain the event knowledge embedding vector. The causal inference module, connected to the graph embedding module, is used to identify causal relationships between industrial equipment events based on the event knowledge embedding vector and through a causal inference model, and generate a causal relationship graph. The source tracing analysis module, connected to the causal inference module, is used to trace the root cause of an event based on the causal relationship graph, determine the root cause of the event, and output the source tracing results. The knowledge graph construction module, connected to the graph embedding module, is used to construct the knowledge graph of the industrial equipment field based on historical fault case data, and to perform embedding learning using a graph neural network.

[0016] The beneficial effects of this invention are: This invention constructs a knowledge graph in the industrial equipment domain, structurally representing the complex mapping relationships between entities such as equipment type, fault type, fault cause, and fault result, providing rich semantic prior knowledge for subsequent causal inference. Compared to traditional data-driven fault diagnosis methods, the knowledge graph introduced in this invention can encode domain experts' in-depth understanding of equipment fault mechanisms, making the tracing results more accurate and interpretable.

[0017] Furthermore, the introduction of knowledge graphs enables causal inference models to better capture the deep causal relationships among multiple variables in industrial equipment failures. By organically combining knowledge graph embedding with causal inference neural networks, this invention achieves complementary advantages between domain knowledge and data-driven approaches, significantly improving the interpretability of results while ensuring the accuracy of tracing the source. In addition, the knowledge graph supports incremental updates; when new failure cases are collected, the knowledge graph can be easily expanded and optimized, effectively reducing system maintenance costs.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0019] Figure 1 This is a flowchart of a knowledge graph-embedded causal tracing method for industrial equipment events according to an embodiment of the present invention; Figure 2 This is a diagram illustrating the visualization and similarity analysis results of graph embedding in one embodiment of the present invention. Figure 3 This is a diagram illustrating the causal attention analysis and path strength effect in one embodiment of the present invention. Figure 4 This is a diagram illustrating the timing and delay distribution of causal propagation in one embodiment of the present invention. Figure 5 This is a diagram illustrating the causal chain and confidence evolution of the tracing results in one embodiment of the present invention. Figure 6 This is a diagram illustrating the effect of timing signal processing and spectrum analysis in one embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] In embodiments of the present invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0022] The present invention provides the following preferred embodiments: In one embodiment, reference Figures 1 to 3As shown, the knowledge graph-embedded causal tracing method for industrial equipment events of the present invention includes the following steps: Referring to step S100, collect multi-source heterogeneous event data during the operation of industrial equipment. Industrial equipment generates a large amount of data through various sensors and monitoring systems during operation, reflecting the equipment's operating status from different perspectives. Specific data collection includes: equipment operating parameter data collected by the equipment operation monitoring system, including vibration signals, temperature signals, pressure signals, and flow signals, which reflect the equipment's basic operating status and performance indicators; sensor time-series data collected by sensor nodes deployed on key components of the industrial equipment, including high-frequency sampling data such as bearing vibration acceleration, rotor speed, current, and voltage, which can capture subtle changes during equipment operation; operation log data recorded by the industrial equipment control system, including operator operation sequences, parameter adjustment records, and start-up and shutdown records, which reflect human-induced factors affecting the equipment; and environmental parameter data collected by the environmental monitoring module, including ambient temperature, ambient humidity, and ambient air pressure, which reflect the impact of the external environment on the equipment's operating status.

[0023] The acquisition of multi-source data is fundamental to subsequent fault analysis, and the integrity and accuracy of the data directly affect the reliability of the tracing results. In the actual system, the data acquisition module adopts a distributed architecture, with different types of data handled by corresponding acquisition units. These units synchronize and transmit data via industrial Ethernet. To ensure data timeliness, each acquisition unit collects data according to a preset sampling frequency, which is differentiated based on the equipment type and the importance of the monitored parameters. For vibration signals from critical equipment, kilohertz-level high-frequency sampling is used; for general operating parameters, hertz-level conventional sampling is used. The acquired raw data is first stored in the buffer of the edge computing nodes. After preliminary data verification and format conversion, it is uploaded to the central server for centralized processing. The edge computing nodes use high-performance embedded processors with local data preprocessing capabilities, enabling preliminary cleaning and compression before data upload, reducing network transmission burden.

[0024] Further, referring to step S200, data preprocessing and feature extraction are performed on the collected multi-source heterogeneous event data to obtain event feature vectors. Due to the complex and variable industrial environment, the collected raw data often contains noise, missing values, and outliers; direct analysis can affect the accuracy of the results. Data preprocessing mainly includes the following steps: First, data cleaning is performed, checking the quality of the collected raw data and identifying and processing missing, outliers, and duplicate values. For short-term data gaps, interpolation methods are used to fill in the gaps; for long-term data gaps, they are marked as special states and processed accordingly. For outliers that significantly exceed the normal range, a reasonable threshold range is set based on the equipment model and operating conditions, and the out-of-range data is marked as suspicious data and then flagged.

[0025] Then, data standardization is performed to unify data with different dimensions and value ranges to the same scale. Two standardization methods are used: min-max normalization and Z-score normalization, with the appropriate method selected based on the data's distribution characteristics. Z-score normalization is used for normally distributed data, while min-max normalization is used for non-normally distributed data. Feature extraction involves extracting key information reflecting the equipment's operating status and fault characteristics from the preprocessed data. Extracted features fall into two main categories: time-domain features and frequency-domain features. Time-domain features include statistical characteristics such as mean, variance, peak value, peak factor, waveform factor, impulse factor, and margin factor, which reflect the overall level and fluctuation of the signal. Frequency-domain features include dominant frequency components, spectral energy distribution, spectral centroid, and spectral variance, which reflect the frequency composition and variation patterns of the signal.

[0026] The extracted time-domain and frequency-domain features are concatenated to form a comprehensive feature vector. This comprehensive feature vector contains comprehensive information about the device's operating status, providing a data foundation for subsequent graph embedding learning. During feature concatenation, an attention mechanism is used to weight and fuse different types of features. Feature weights are automatically learned based on the contribution of each feature to fault diagnosis, improving the discriminative power of the feature representation. In practical applications, the dimension of the feature vector is typically between several hundred and several thousand. In this embodiment, a feature vector with a dimension of 512 is used, with 208 dimensions for time-domain features, 208 dimensions for frequency-domain features, and 96 dimensions for the fused comprehensive feature. The feature vector is constructed using a sparse representation method, effectively reducing storage costs and computational complexity. The sparse representation method significantly reduces storage space requirements by storing only the positions and values ​​of non-zero elements. Experiments show that compared to dense vector representation, sparse representation can reduce storage costs by approximately 70% while maintaining the effectiveness of the feature representation. Furthermore, sparse feature vectors are computationally more efficient, accelerating the subsequent graph embedding learning process. In practical deployment, this invention employs an adaptive sparsity strategy, dynamically adjusting the sparsity based on the importance of features. For important features, a high density is maintained to ensure representation accuracy; for less important features, a higher sparsity is used to save storage space. This adaptive strategy effectively reduces resource consumption while ensuring system performance.

[0027] In another embodiment, reference Figure 2 As shown, this invention provides a detailed explanation of the embedding visualization and similarity analysis of knowledge graphs in the field of industrial equipment.

[0028] The industrial equipment domain knowledge graph is a core component of this invention. It provides a structured representation of domain experts' understanding of equipment failure mechanisms, offering semantically rich prior knowledge for subsequent causal inference. Referring to steps S300 and S400, after the knowledge graph is constructed, embedding representation learning is required for the nodes in the graph, mapping each node to a low-dimensional vector space to facilitate subsequent causal inference calculations. This embodiment uses a graph attention network as the embedding learning method. The core idea of ​​a graph attention network is to use an attention mechanism to learn the importance weights of different neighboring nodes for the current node, thereby generating more accurate node embedding vectors. The embedding vector is set to 128 dimensions, which can reduce computational complexity while preserving semantic information.

[0029] To verify the effectiveness of embedding learning, this invention employs t-SNE dimensionality reduction technology to map high-dimensional embedding vectors onto a two-dimensional plane for visualization. The visualization results reveal that semantically similar nodes cluster together on the two-dimensional plane, forming distinct clusters. For example, nodes related to component states, such as bearing status nodes and rotor status nodes, cluster in one area, while fault type nodes, such as abnormal vibration and temperature rise, cluster in another. Cause nodes, such as poor lubrication and wear, cluster in yet another new area. This clustering effect demonstrates that embedding learning successfully captures the semantic similarity between nodes. In the visualization, different types of nodes are identified with different colors and shapes; for example, equipment type nodes are represented by green circles, component status nodes by orange squares, fault type nodes by red triangles, and cause nodes by purple rhombuses. The distinction between colors and shapes allows for a direct observation of the distribution patterns of various node types.

[0030] Furthermore, this invention calculates the cosine similarity between different nodes to construct a node similarity heatmap. The values ​​in the heatmap represent the cosine similarity between the embedding vectors of two corresponding nodes; the closer the value is to one, the closer the semantic relationship between the two nodes. For example, the similarity between the bearing state node and the poor lubrication node reaches 0.82, indicating a close correlation in their fault mechanisms. The similarity between the rotor state node and the vibration anomaly node reaches 0.68, reflecting that rotor imbalance is a common cause of vibration anomalies. The similarity between the vibration anomaly node and the temperature rise node reaches 0.79, indicating that these two fault modes often occur simultaneously. This similarity analysis can verify the rationality of edge relationships in the knowledge graph and also discover potential correlations. The heatmap's color blocks gradually transition from red to green, with red representing high similarity and green representing low similarity. This color coding method makes the interpretation of the similarity matrix more intuitive.

[0031] Nodes with semantic similarity are close in vector space. For example, the embedding vectors of a bearing wear node and a poor lubrication node are close because they are closely related in terms of failure mechanism. Through this embedding learning method, the structured knowledge in the knowledge graph is transformed into dense vector representations, which facilitates subsequent causal inference calculations. The training process of embedding learning adopts a contrastive learning method, which makes the embedding vectors of similar nodes as close as possible and the embedding vectors of dissimilar nodes as far apart as possible, thereby enhancing the discriminative ability of the embedding representation. Contrastive learning uses a triple loss function, where a triple consists of an anchor node, a positive example node, and a negative example node. The training objective is to make the distance between the anchor node and the positive example node less than the distance between the anchor node and the negative example node. During training, the triple selection strategy has a significant impact on model performance. This invention adopts a hard negative example mining strategy, that is, it prioritizes selecting negative example nodes that are semantically similar to the anchor node for training. This strategy enables the model to learn a more refined discrimination boundary. Experiments show that the hard negative example mining strategy can improve the discrimination accuracy of the embedding representation by about 15%.

[0032] In another embodiment, reference Figure 3 As shown, this invention provides a detailed explanation of causal attention analysis and path strength.

[0033] Referring to step S400, the causal inference model is the core module for fault tracing in this invention. Based on the knowledge graph embedding vectors, it uses deep learning technology to identify causal relationships between industrial equipment events and generate a causal relationship graph. The overall architecture of the model includes three main components: a temporal encoder, a causal attention layer, and a causal path finder. Referring to step S500, the root cause of the event is further traced based on the causal relationship graph to determine the fundamental cause of the event and output the tracing results.

[0034] The temporal encoder is responsible for extracting temporal features from the event sequence. It employs a bidirectional long short-term memory (LSTM) network architecture, capable of simultaneously capturing both forward and backward dependencies in the event sequence. The network consists of two layers of bidirectional LSM, each containing 256 hidden units, with a dropout ratio set to 0.2 to prevent overfitting. The input to the temporal encoder is the event knowledge embedding vector, obtained by embedding the event feature vector into a knowledge graph. The output is the temporal representation vector of the event sequence. This temporal representation vector encodes the evolutionary pattern of the event over time, providing temporal contextual information for subsequent causal analysis.

[0035] The causal attention layer is the core innovation of the causal inference model. It learns the strength of causal relationships between event nodes through an attention mechanism. Unlike ordinary self-attention mechanisms, the causal attention layer introduces causal constraints when calculating attention scores, allowing only attention from earlier events to later events. This causal constraint aligns with the physical laws of fault propagation in industrial equipment. The causal attention layer employs a multi-head attention mechanism, containing eight parallel attention heads, each learning causal relationship patterns in a different subspace. This invention uses a 3D surface plot to illustrate the distribution of causal attention weights among event node pairs. The plot shows that attention weights are high near the diagonal, indicating the strongest correlation between the same event and itself; conversely, attention weights gradually decrease away from the diagonal along the time axis, consistent with the physical laws of causal propagation. In the 3D surface plot, the surface height represents the magnitude of the attention weights, with red indicating high-weight regions and blue indicating low-weight regions. By rotating the viewpoint, the distribution pattern of attention weights can be observed from different angles.

[0036] This invention uses a causal path finder to extract specific causal paths from a causal adjacency matrix. A causal path is a causal chain connecting a source event and a target event, reflecting the propagation path of a fault in the device. Each causal path has a corresponding path strength, which is calculated by combining the causal weights of each edge on the path. This invention uses a dotted-line graph to illustrate the strength distribution of different causal paths. The graph shows that the causal path from E1 to E3 has the highest strength, reaching 0.89, indicating that the path of poor lubrication directly leading to excessive vibration is the most important fault propagation channel. The causal path from E2 to E4 has a strength of 0.76, indicating that the path of temperature rise leading to bearing damage is a secondary but still important fault evolution path. The causal path from E1 to E4 has a strength of 0.72, indicating that poor lubrication can also indirectly lead to excessive vibration through other intermediate events. The causal path from E3 to E6 has a strength of 0.58, indicating that the path of bearing wear leading to lubrication system failure is relatively less important. In the dotted-line graph, each data point is marked with a dot, and the data points are connected by straight lines. The thickness and color intensity of the connecting lines represent the magnitude of the path strength.

[0037] The design of the loss function in the causal inference model is crucial to ensuring its generalization ability. In this embodiment, the loss function consists of three parts: a causal cross-entropy loss term, which measures the difference between the predicted causal relationship and the true causal relationship, calculated by comparing the predicted causal adjacency matrix with the true causal adjacency matrix constructed based on historical failure cases; a temporal consistency loss term, which maintains the causal logical consistency of events in time, ensuring that the direction of the causal relationship is consistent with the temporal direction; and a causal transitivity loss term, which maintains the transitivity of the causal relationship, i.e., if A causes B, and B causes C, then A should be able to cause C. The overall goal of the loss function is to minimize the weighted sum of the three loss terms while satisfying the mathematical constraints of the causal relationship. The model is trained using the backpropagation algorithm and the Adam optimizer, with a learning rate of 0.001, a batch size of 32, and a total of 500 training epochs.

[0038] In another embodiment, reference Figure 4 As shown, this invention provides a detailed explanation of the temporal analysis and delay distribution of causal propagation.

[0039] Industrial equipment faults exhibit time delays during propagation, reflecting the physical process of fault propagation within the equipment. This invention captures the temporal characteristics of fault propagation through time-series analysis and quantifies the statistical distribution of propagation delays.

[0040] This invention collects time-series data on fault propagation through simulation experiments. During the experiment, a specific type of fault is injected into the device under test, and the occurrence time of subsequent fault events is recorded. By analyzing the time-series data of multiple causal links, the time delay characteristics of fault propagation can be obtained. The time-series graph clearly shows that the intensity of the fault event initially increases and then decreases over time, due to the gradual attenuation of fault energy during propagation. Different types of fault events exhibit different time-series characteristics: the lubrication failure event shows a gentle rise and fall curve, with the peak occurring approximately 18 seconds after fault injection; the temperature rise event has a faster response speed, with the peak occurring approximately 28 seconds after fault injection; the vibration exceeding limit event shows a multi-peak time-series curve, reflecting the complexity of the fault propagation path. In the time-series graph, event intensity is represented by a filled curve, with different events distinguished by different colors. The filled area below the curve represents the time integral of the event intensity.

[0041] This invention further analyzes the statistical distribution characteristics of causal propagation delay. By collecting propagation delay data from a large number of failure cases, a cumulative distribution curve of the delay is fitted. The median delay of eight seconds can be read from the distribution curve, indicating that 50% of causal propagation delays are below eight seconds. The shape of the distribution curve shows that most causal propagation delays are concentrated in the range of two to twenty seconds, which is consistent with the physical characteristics of industrial equipment. The shaded area next to the distribution curve represents a 95% confidence interval, reflecting the uncertainty range of the delay distribution. At the tail end of the delay distribution, the curve gradually flattens out, indicating the existence of a small number of special cases with long delays. These special cases may be due to special operating conditions of the equipment or changes in the fault propagation path. By analyzing the delay distribution, the search depth of the tracing algorithm can be optimized, reducing the computational load while ensuring the accuracy of tracing.

[0042] The causal path finder is responsible for extracting specific causal paths from the causal adjacency matrix. A causal path is a causal chain connecting a source event and a target event, reflecting the propagation path of a fault within the device. This embodiment employs a depth-first search-based shortest path search method, starting from the target event and searching backwards for all possible causal paths. During the search process, the total weight of the paths is calculated based on the weight values ​​in the causal adjacency matrix, prioritizing paths with higher weights. The final output causal paths are presented in list form, with each path containing the source event, intermediate events, target event, and path weight. The time delay between adjacent events is also marked during the tracing process. As shown in the figure, the delay from poor lubrication to temperature rise is eight seconds, and the delay from temperature rise to excessive vibration is ten seconds. This time delay data is of significant value for understanding the fault propagation mechanism.

[0043] In another embodiment, reference Figure 5 As shown, this invention provides a detailed explanation of the causal chain analysis and confidence evolution of the tracing results.

[0044] Source tracing analysis is the core function of this invention. Based on the causal relationship graph, it uses a source tracing algorithm to determine the root cause of equipment failure. The source tracing algorithm adopts a ranking method based on causal contribution. It first takes the target failure event as the starting point for source tracing and extracts all upstream causal chains from the causal relationship graph.

[0045] This invention uses a causal chain diagram to illustrate the complete causal propagation path of the tracing results. The diagram shows that excessive vibration, as the final failure event, is caused by two direct causes: bearing wear and temperature rise. Further tracing upstream, the direct cause of bearing wear is poor lubrication, and the direct cause of temperature rise is grease oxidation. Poor lubrication, as the root cause, contributes the most, reaching 92%. Grease oxidation contributes 65%, ambient humidity contributes 42%, and excessive load contributes 55%. The diagram also shows that ambient humidity and excessive load are auxiliary factors with relatively low contributions, but this does not mean they are unimportant. Under certain conditions, these auxiliary factors can become decisive factors. For example, in high humidity environments, the severity of poor lubrication is amplified; under overload conditions, the rate of bearing wear is significantly accelerated.

[0046] Source tracing confidence is a crucial indicator of the reliability of source tracing results. This invention continuously improves source tracing confidence through an iterative optimization algorithm. The confidence evolution curve shows that the method of this invention achieves a high confidence level in the initial stage, which is further improved to above 0.95 through iterative optimization. Compared with traditional methods, this invention maintains a higher confidence level throughout the iteration process, indicating that the introduction of knowledge graphs can indeed improve the reliability of source tracing. The initial confidence level of traditional methods is 0.75, and it can only be improved to 0.71 in the end. This is because traditional methods lack the support of domain knowledge and are prone to getting trapped in local optima during iteration. The confidence curve of expert experience lies between that of this invention and traditional methods, verifying the effectiveness of this invention. The shaded area of ​​the confidence curve represents the range of confidence fluctuation. The shaded area of ​​this invention's method is significantly narrower than that of the traditional method, indicating that this invention's method has better stability.

[0047] The interpretability of the tracing results is a significant advantage of this invention. Unlike black-box machine learning models, the tracing results of this invention clearly demonstrate the complete causal chain from cause to effect. Maintenance personnel can analyze step-by-step along this causal chain to understand the mechanism of the fault. When the system outputs "poor bearing lubrication" as the root cause, maintenance personnel can view the causal chain: poor bearing lubrication leads to increased bearing temperature, increased bearing temperature accelerates lubricant oxidation, oxidized lubricant further exacerbates bearing wear, and bearing wear ultimately leads to excessive vibration. This interpretable tracing result helps maintenance personnel quickly locate the root cause of the problem and develop targeted solutions. Furthermore, the tracing results can be linked with a knowledge graph to view the contextual information of the cause within the knowledge graph, understanding the relevant equipment type, fault type, and handling methods.

[0048] In another embodiment, reference Figure 6As shown, this invention provides a detailed description of time-series signal processing and spectrum analysis.

[0049] Sensor data generated during the operation of industrial equipment contains a large amount of noise, making it difficult to extract effective fault features by directly analyzing the raw signals. This invention employs signal processing techniques to denoise and extract features from the raw signals, providing high-quality input data for subsequent causal inference.

[0050] Raw sensor signals often contain irrelevant components such as environmental noise and electromagnetic interference. This invention employs a Savitzky-Golay filter to denoise the raw signal. This filter effectively removes high-frequency noise while preserving signal edge information. The Savitzky-Golay filter is a digital filter based on local polynomial regression, which estimates the smoothed signal value by performing polynomial fitting on the data within a sliding window. Compared to traditional Gaussian filters, the Savitzky-Golay filter better preserves the high-frequency components and abrupt changes in the signal. The comparison of the signals before and after denoising shows that the raw signal exhibits obvious spike-like fluctuations, mainly caused by measurement noise. After denoising, the signal becomes smoother but still retains the main characteristics and abrupt changes of the original signal; for example, signal abrupt changes around 20 and 50 seconds are well preserved. The denoised signal is more conducive to subsequent feature extraction and fault diagnosis. In the figure, the thin gray line represents the raw signal, the thick blue line represents the denoised signal, and the light blue area below the signal curve represents the signal filling effect.

[0051] Spectrum analysis is an important method for extracting fault characteristics. This invention uses Fast Fourier Transform (FFT) to convert the time-domain signal to the frequency domain, obtaining the signal's spectral distribution. From the spectrum, two main frequency components can be clearly identified: a low-frequency component (0.05 Hz) corresponds to the slow changing trend of the equipment; this frequency component is related to the equipment's operating cycle, for example, the rotational speed of rotating equipment determines the fundamental frequency. A high-frequency component (0.125 Hz) corresponds to the equipment's rapid oscillation mode; this frequency component is related to the equipment's vibration characteristics, such as gear meshing frequency or bearing rolling frequency. By analyzing the changes in the spectral components, it is possible to determine whether the equipment experiences abnormal vibration. For example, when a bearing wears, the energy of the high-frequency component increases significantly; when the rotor is unbalanced, a significant peak is generated at the fundamental frequency. In the spectrum, logarithmic coordinates are used to display the amplitude over a wide dynamic range, and red dashed lines mark the positions of the two main frequency components.

[0052] The timing signal processing and spectrum analysis method of this invention has broad application prospects in industrial equipment condition monitoring. By monitoring vibration signals during equipment operation in real time, abnormal equipment conditions can be detected early, preventing the occurrence and development of faults. The system adopts an edge computing architecture, completing real-time signal acquisition and processing at the device end, and only uploading the analysis results to the central server, greatly reducing network transmission load and response latency. In actual deployment, the sampling frequency is set to 2 kHz, and the data window length for each processing is 4,096 sampling points, equivalent to approximately two seconds of data. After processing, the system extracts spectral features and compares them with the spectral benchmark of the healthy state. If the deviation exceeds a preset threshold, a fault alarm is triggered and the source tracing analysis process is initiated. The latency of the entire processing process is controlled within 500 milliseconds, meeting industrial real-time requirements. The system also supports fault early warning functionality, establishing an equipment fault prediction model through the analysis of historical fault data. This model can predict the probability of equipment failure in the future based on current operating status data. When the predicted probability exceeds a preset threshold, the system will issue an early warning to help maintenance personnel take preventive maintenance measures and effectively avoid unplanned downtime. The core of the fault early warning function is establishing a mapping relationship between equipment operating status and the probability of fault occurrence. This invention employs a deep learning-based prediction model. This model takes the temporal features extracted by the temporal encoder and the causal correlation strength output by the causal inference module as input, and outputs the probability of equipment failure within the next 24 hours. The model training adopts a transfer learning strategy, first pre-training on a large-scale industrial equipment dataset, and then fine-tuning on specific equipment. This strategy effectively solves the problem of scarce industrial equipment fault samples. The system also provides rich data visualization functions to help maintenance personnel intuitively understand the equipment's operating status and fault conditions. The data visualization module supports the display of real-time and historical data, and users can access the system's monitoring interface anytime, anywhere through a browser or mobile device. The visualization interface adopts a responsive design, adapting to different screen sizes to ensure a good user experience. The system also supports custom alarm rules, allowing users to set alarm thresholds and alarm methods according to actual needs. Alarm methods include multiple channels such as SMS, email, and WeChat, ensuring that alarm information is delivered to relevant personnel in a timely manner. The system also provides alarm history query and statistical functions to help users analyze the patterns and trends of alarm occurrences. The system's data storage employs a distributed architecture, with time-series data stored in a time-series database, relational data stored in a graph database, and alarm records stored in a key-value database. This distributed storage architecture ensures high availability and scalability; the failure of a single storage node will not affect the normal operation of the entire system. The system also supports automatic data backup and recovery, effectively preventing data loss. The system's security mechanism adopts a multi-layered design, including four levels: network security, host security, application security, and data security.The network security layer employs firewalls and intrusion detection systems to protect the system from external attacks. The host security layer uses access control and permission management to ensure that only authorized users can access system resources. The application security layer uses input validation and security auditing to prevent application-layer attacks and data leaks. The data security layer uses encrypted storage and transmission to ensure data confidentiality and integrity. In terms of system operation and maintenance, this invention provides comprehensive monitoring and logging functions. The system monitors the operating status of each module in real time and automatically triggers alarms when an anomaly is detected. Log records cover all aspects of system operation, user operations, and fault handling, supporting log query, statistics, and export functions. Through log analysis, the causes of faults can be traced, providing data support for system optimization. This invention also supports integration with other enterprise information systems, interacting with equipment management systems, work order systems, and asset management systems through standardized data interfaces to achieve information sharing and business collaboration. This integration method avoids data silos and improves the overall informatization level of the enterprise. The system also supports secondary development, providing rich application programming interfaces (APIs) to facilitate user-defined functional expansion and customization according to actual needs. In summary, this invention achieves intelligent fault tracing for industrial equipment through the deep integration of knowledge graphs and causal inference. The system boasts excellent scalability, maintainability, and ease of use, adaptable to the application needs of enterprises across different industries and sizes. With the development of the Industrial Internet and the advancement of intelligent manufacturing strategies, this invention will be applied and promoted in more fields. The technical solution of this invention has undergone thorough experimental verification and has achieved excellent application results in multiple industrial scenarios. Experimental results show that compared to traditional methods, the accuracy of fault tracing is improved by approximately 20%, and the speed is improved by approximately 30%, creating significant economic benefits for enterprises. In the future, this invention will continue to deepen the research on the integration of knowledge graphs and causal inference technologies, exploring cutting-edge directions such as multimodal data fusion and cross-device collaborative fault tracing, promoting the development of intelligent operation and maintenance technology for industrial equipment.

[0053] Although the present invention has been specifically described above with reference to preferred embodiments, it should be understood that the present invention is not limited to the embodiments described above. Various modifications and variations can be made by those skilled in the art without departing from the spirit of the present invention, and such modifications and variations should fall within the scope defined by the appended claims and their equivalents.

Claims

1. A method for causal tracing of industrial equipment events using knowledge graph embedding, characterized in that, Includes the following steps: Collect multi-source heterogeneous event data during the operation of industrial equipment. The multi-source heterogeneous event data includes equipment operating parameter data, sensor timing data, operation log data, and environmental parameter data. The collected multi-source heterogeneous event data is preprocessed and features are extracted to obtain event feature vectors; Based on a pre-built knowledge graph of the industrial equipment domain, graph embedding representation learning is performed on the event feature vector to obtain event knowledge embedding vector, wherein the industrial equipment domain knowledge graph includes equipment type nodes, event type nodes, cause nodes, result nodes, and edge relationships for connecting the nodes; Based on the event knowledge embedding vector, the causal relationships between industrial equipment events are identified through a causal inference model, and a causal relationship map is generated. Based on the causal relationship graph, the root cause of the event is traced to determine the fundamental cause of the event and the tracing results are output.

2. The knowledge graph-embedded causal tracing method for industrial equipment events according to claim 1, characterized in that, The collection of multi-source heterogeneous event data during the operation of industrial equipment specifically includes: The equipment operation monitoring system collects equipment operation parameter data, including vibration signals, temperature signals, pressure signals, and flow signals. Sensor timing data is collected by sensor nodes deployed on industrial equipment; Obtain operation log data through the industrial equipment control system; Environmental parameter data are collected through an environmental monitoring module, including ambient temperature, ambient humidity, and ambient air pressure.

3. The knowledge graph-embedded causal tracing method for industrial equipment events according to claim 1, characterized in that, The method for constructing the knowledge graph in the industrial equipment field includes: Collect historical failure case data of industrial equipment, including equipment type, failure type, failure cause, failure result and failure handling record; Extract the mapping relationship between equipment type and fault type, the causal relationship between fault type and fault cause, and the propagation relationship between fault cause and fault result from the historical fault case data; Establish a set of node types, including device type nodes, component status nodes, fault type nodes, cause nodes, and result nodes; Based on the extracted mapping and causal relationships, a set of edges connecting the nodes is constructed, and each edge is assigned a relation weight calculated based on the frequency of historical cases. A graph neural network is used to embed the initially constructed knowledge graph of the industrial equipment domain into the knowledge graph to obtain a vector representation of each node in the knowledge graph of the industrial equipment domain.

4. The knowledge graph-embedded causal tracing method for industrial equipment events according to claim 3, characterized in that, The step of identifying causal relationships between industrial equipment events based on the event knowledge embedding vector and using a causal inference model includes: The event knowledge embedding vector is input into a preset causal inference neural network model, which adopts a temporal causal network architecture based on an attention mechanism; The temporal encoder in the causal inference neural network model extracts temporal features from the event sequence to obtain temporal causal features. The causal association strength between event nodes is calculated through a causal attention layer, generating a causal adjacency matrix; Based on the causal adjacency matrix, causal path search is performed to determine the direct and indirect causal relationships between events.

5. The knowledge graph-embedded causal tracing method for industrial equipment events according to claim 4, characterized in that, The loss function of the causal inference neural network model includes: The causal cross-entropy loss term is used to measure the difference between the predicted causal relationship and the true causal relationship. The temporal consistency loss term is used to maintain the causal logical consistency of events in time sequence; The graph structure consistency loss term is used to maintain the distance relationship between the topological structure information in the knowledge graph of the industrial equipment domain in the embedding space.

6. The knowledge graph-embedded causal tracing method for industrial equipment events according to claim 1, characterized in that, The process of tracing the root cause of events based on the causal relationship graph includes: Based on the aforementioned causal relationship graph, starting from the target event node, a depth-first search-based tracing algorithm is used to trace the causal chain upstream. During the tracing process, the causal contribution of each upstream event node to the target event is calculated; The upstream events are sorted according to their causal contribution, and the event node with the highest causal contribution is determined as the root cause. Generate a source tracing results report that includes the root cause, causal propagation path, and source tracing confidence level.

7. The knowledge graph-embedded causal tracing method for industrial equipment events according to claim 1, characterized in that, Also includes: Based on the source tracing results, generate fault diagnosis suggestions; The source tracing results and the fault diagnosis suggestions are pushed to the industrial equipment management terminal; The source tracing results are used for feedback learning to update the edge relation weights in the knowledge graph of the industrial equipment domain.

8. The knowledge graph-embedded causal tracing method for industrial equipment events according to claim 3, characterized in that, The node types also include maintenance record nodes and quality indicator nodes.

9. The knowledge graph-embedded causal tracing method for industrial equipment events according to claim 4, characterized in that, The causal attention layer employs a multi-head self-attention mechanism, which includes multiple attention heads, each of which learns causal association patterns in different subspaces.

10. A knowledge graph-embedded causal tracing system for industrial equipment events, used to implement the knowledge graph-embedded causal tracing method for industrial equipment events as described in any one of claims 1 to 9, characterized in that, include: The multi-source data acquisition module is used to collect multi-source heterogeneous event data during the operation of industrial equipment; The data preprocessing module, connected to the multi-source data acquisition module, is used to perform data preprocessing and feature extraction on the acquired multi-source heterogeneous event data to obtain event feature vectors. The graph embedding module, connected to the data preprocessing module, is used to perform graph embedding representation learning on the event feature vector based on a pre-built industrial equipment domain knowledge graph to obtain the event knowledge embedding vector. The causal inference module, connected to the graph embedding module, is used to identify causal relationships between industrial equipment events based on the event knowledge embedding vector and through a causal inference model, and generate a causal relationship graph. The source tracing analysis module, connected to the causal inference module, is used to trace the root cause of an event based on the causal relationship graph, determine the root cause of the event, and output the source tracing results. The knowledge graph construction module, connected to the graph embedding module, is used to construct the knowledge graph of the industrial equipment field based on historical fault case data, and to perform embedding learning using a graph neural network.

Images