Key equipment fault prediction method based on historical big data and fault map

Abstract

The invention provides a key equipment fault prediction method based on historical big data and a fault map. The method comprises the following steps: sorting historical fault data of equipment of a selected type based on a big data technology; performing data preprocessing on the historical fault data to obtain target data; constructing a fault map according to the target data; establishing a real-time monitoring module of the equipment according to the fault map; carrying out feature extraction on the fault map to obtain various types of structured features, and fusing the structured features with unstructured data to obtain trainable data; training an equipment fault prediction model by using an isolation forest algorithm and the trainable data; inputting the real-time monitoring data obtained by the real-time monitoring module into the trained equipment fault prediction model, and predicting the fault probability of the equipment in real time; when the fault probability exceeds a preset threshold value, the system automatically sends out an early warning signal. According to the invention, equipment faults are early warned in advance, major accidents can be effectively prevented, the downtime is reduced, and the production efficiency and the safety are improved.

Description

Technical Field

[0001] The present invention relates to the field of industrial automation and intelligence, and is a method for predicting faults of key equipment by using big data and fault map technology. Background Art

[0002] In modern industrial production, the reliability of key equipment directly affects production efficiency and safety. Traditional fault detection methods rely on regular maintenance and manual inspections, which are costly, inefficient, and difficult to detect potential faults in a timely manner. With the rapid development of information technology, especially the breakthroughs in big data and artificial intelligence technology, the traditional industrial field is undergoing unprecedented changes. Modern manufacturing is gradually shifting from a model that relies on a large number of laborers to a model that relies more on technology. Therefore, a new maintenance method is urgently needed to improve the accuracy of predictions and the timeliness of maintenance.

[0003] In this context, the present invention proposes a key equipment failure prediction method based on historical big data and failure maps; this method builds a comprehensive failure map by deeply mining and analyzing the historical failure data of the equipment, thereby realizing accurate prediction of equipment failures. This method can not only reduce unexpected downtime and maintenance costs, but also improve production efficiency and bring greater economic benefits to enterprises. Summary of the invention

[0004] In view of this, the present invention proposes a key equipment fault prediction method based on historical big data and fault maps to solve the problems of inefficient key equipment maintenance and inaccurate fault prediction in modern manufacturing industry.

[0005] The specific technical solutions of the present invention are as follows:

[0006] A key equipment failure prediction method based on historical big data and failure maps, including:

[0007] Organize historical failure data of selected types of equipment based on big data technology;

[0008] Perform data preprocessing on historical fault data to obtain target data;

[0009] Construct a fault map based on target data;

[0010] Establish a real-time monitoring module for equipment based on the fault map;

[0011] Extract features from the fault map to obtain various structured features, and fuse the structured features with unstructured data to obtain trainable data;

[0012] Training a device failure prediction model using an isolation forest algorithm and the trainable data;

[0013] The real-time monitoring data obtained by the real-time monitoring module is input into the trained equipment failure prediction model to predict the failure probability of the equipment in real time;

[0014] When the probability of failure exceeds the preset threshold, the system automatically sends out a warning signal, and the inspection personnel conduct maintenance according to the alarm signal;

[0015] The successful warning records, error warning records and equipment failure records that were not warned are synchronized to the historical failure data in real time to form new historical failure data.

[0016] Furthermore, the collating of historical fault data of the selected type of equipment based on big data technology includes: collecting historical fault data of the equipment from various sources, including sensor data, log files, fault and maintenance records, to obtain historical fault data.

[0017] Furthermore, the data preprocessing of the historical fault data to obtain the target data includes:

[0018] Extract relevant data from historical fault data, including sensor data, log file data, fault and maintenance record data;

[0019] Fusion of multimodal data including text, images, and sensor data into one text field data;

[0020] Applying cleaning techniques to clean text field data to improve data consistency and usability, the cleaning techniques include removing stop words, correcting spelling errors, and standardizing named entities;

[0021] Using natural language processing technology to perform word segmentation on the fused text field data and perform part-of-speech tagging to identify key words, including nouns and verbs;

[0022] Extract the key words and phrases related to the fault from the word segmentation results through pattern matching or machine learning methods;

[0023] Based on the extracted keywords and phrases, the text field data is classified into different fault types or equipment status categories to form target data.

[0024] Furthermore, constructing a fault map according to the target data includes:

[0025] Step 3.1: Extract knowledge from the target data to obtain ontology structured data, including:

[0026] Identify entities related to equipment failures in the target data based on the BERT model, including equipment components, failure types, and operating conditions;

[0027] The Bi-LSTM technology is used to capture the bidirectional dependencies of the text and the Attention mechanism is used to focus on key information for relationship extraction;

[0028] Extracting entity attributes based on graph neural networks, including component specifications and fault severity;

[0029] Step 3.2: Integrate the ontology structured data with the third-party database data and align the entities to obtain complete structured data;

[0030] Step 3.3: Construct the fault graph ontology based on the complete structured data;

[0031] Step 3.4: Perform knowledge reasoning on the fault graph through logical reasoning algorithm and graph neural network model;

[0032] Step 3.5: Complete the fault graph through the heterogeneous graph attention network;

[0033] Step 3.6: Establish an automated knowledge update mechanism so that new data can be incorporated into the fault map in a timely manner;

[0034] Step 3.7: Perform quality assessment on the fault map, including completeness assessment, consistency assessment, accuracy assessment, coverage assessment, readability assessment, and scalability assessment;

[0035] Step 3.8: Store the fault map in the form of a graph database.

[0036] Furthermore, the real-time monitoring module of the equipment is established according to the fault map, including:

[0037] Refer to the entities, entity attributes and relationships between entities in the fault map, add high-precision sensors or surveillance cameras to monitor the operating status of key equipment in real time, and obtain the equipment's temperature data, pressure data, vibration data, current data, voltage data, image data, and video data;

[0038] The data collected by high-precision sensors or surveillance cameras are transmitted to the central database in real time through the industrial Internet of Things platform, providing real-time operation data of the equipment for the equipment failure prediction model.

[0039] Furthermore, the feature extraction of the fault map includes:

[0040] By extracting the node attributes of each device in the fault map and the detailed information of each maintenance, the entity features and attribute features are extracted. The entity features include the device parts, fault types, and operating conditions, and the attribute features include the specifications, status, and maintenance history of the equipment.

[0041] By extracting the relationship between the components of the equipment and analyzing the fault propagation path between the components, the relationship features are extracted, and the features include the composition of the components and the influence between them;

[0042] According to formulas (1), (2), and (3), the degree centrality, closeness centrality, and betweenness centrality of each device or component node in the fault map are calculated to identify key components:

[0043]

[0044]

[0045]

[0046] Among them, C D (v) represents the degree centrality of the node, deg(v) represents the degree of the node, and n-1 represents the total number of nodes in the graph except node v; C C (v) represents the closeness centrality of the node, d(v,u) represents the shortest path length from node v to any other node u, ∑ u≠v d(v,u) means summing the shortest path lengths from v to all nodes u different from v; C B (v) represents the betweenness centrality of the node, s and t represent any two different nodes in the graph, σ st represents the number of all shortest paths from node s to node t, σ st (v) represents the number of shortest paths from node s to node t through node v;

[0047] Extract path and sequence features, including the order and frequency of failures, by extracting the sequence patterns of failures and analyzing the sequence of maintenance operations;

[0048] Extract time series features by extracting the timestamps of fault occurrence and maintenance operations, including the period and time distribution of fault occurrence;

[0049] Use correlation analysis and mutual information methods to select the most influential features, including the period and time distribution of fault occurrence;

[0050] Apply PCA and t-SNE dimensionality reduction techniques to reduce the number of features and improve model training efficiency;

[0051] The structured features and unstructured data are fused to obtain trainable data with extracted features. The trainable data is a structured data table, in which each column represents a feature and each row represents a sample. The structured features include entity attribute features, relationship features, centrality indicators, path and sequence features, and time series features. The unstructured data includes fault report text, sensor data, and log files.

[0052] Furthermore, the use of the isolation forest algorithm and the trainable data to train the equipment failure prediction model includes:

[0053] Merge the trainable data into a unified dataset, and split the fused dataset into a training set and a test set;

[0054] Initialize the isolation forest model and use the training set data to train the isolation forest model. The isolation forest model learns the "normal" pattern of normal data and identifies abnormal points that deviate from the "normal" pattern. Then use the trained model to calculate the anomaly score of the test set data, and use the predict method to detect anomalies on the test set data based on the calculated anomaly score. The isolation forest model marks each data point as normal (1) or abnormal (-1). Use precision, recall, and F1 score indicators to evaluate the performance of the model. Analyze the prediction results of the isolation forest model and identify the data points identified as abnormal by the model.

[0055] Furthermore, the real-time monitoring data obtained by the real-time monitoring module is input into the trained equipment failure prediction model to predict the failure probability of the equipment in real time, including:

[0056] Perform data preprocessing on the real-time monitoring data obtained in the equipment real-time monitoring module to obtain input data suitable for model prediction;

[0057] The processed data is input into the equipment fault prediction model for fault prediction. The specific process is as follows:

[0058] The isolation forest model calculates an anomaly score S(x) for each data point x according to formula (4):

[0059]

[0060] Where d(x) is the distance from the data point x to its nearest outlier;

[0061] Then, S(x) is normalized using formula (5) to obtain an abnormality score ranging from 0 to 1, which is the probability of equipment failure:

[0062]

[0063] 10. Furthermore, when the failure probability exceeds a preset threshold, the system automatically issues an early warning signal, and the inspection personnel conduct maintenance according to the alarm signal, specifically including:

[0064] Set the alarm threshold T to determine whether the probability of equipment failure exceeds the alarm threshold T. If it exceeds, an alarm is issued:

[0065]

[0066] After the alarm is sounded, the staff will inspect and repair the equipment according to the alarm signal.

[0067] Furthermore, the successful warning records, error warning records and equipment failure records that have not been warned are synchronized in real time to the historical failure data to form new historical failure data, including:

[0068] There are two situations when the staff conducts maintenance on the equipment. One is that after the early warning system issues an early warning, the staff conducts maintenance according to the early warning information; the other is that the early warning system does not issue an early warning, and the staff conducts maintenance on the equipment after the equipment has failed. Among them, the first situation is divided into successful early warning and error early warning;

[0069] The staff will enter the successful warning records, error warning records and equipment failure records that have not been warned into the database in real time. The system will then automatically extract the sensor data of the current equipment failure and the recent log file data to form new historical failure data;

[0070] The system updates the fault map and equipment failure prediction model in real time based on new historical fault data.

[0071] The beneficial effects of the present invention are as follows: by establishing a fault prediction model based on historical big data and fault maps, the unexpected downtime of equipment can be effectively reduced, maintenance costs can be reduced, and production efficiency can be improved. At the same time, the combination of real-time monitoring and intelligent analysis makes equipment maintenance more scientific and reasonable, and can adapt to the high requirements of modern industrial production for equipment management. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention and do not constitute a limitation of the present invention, wherein:

[0073] Figure 1 This is a flow chart of a key equipment failure prediction method based on historical big data and failure maps of the present invention;

[0074] Figure 2 This is a flow chart for constructing a fault map in the present invention. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0076] Example 1

[0077] The present invention provides a key equipment failure prediction method based on historical big data and failure map. Figure 1 As shown, it mainly includes the following steps, and the present invention will be specifically described below in conjunction with the power distribution cabinet equipment:

[0078] Step 1: Organize historical fault data of power distribution cabinet equipment based on big data technology;

[0079] Step 2: Preprocess the historical fault data to obtain target data;

[0080] Step 3: Build a fault map based on historical target data;

[0081] Step 4: Establish a real-time monitoring module for the power distribution cabinet equipment according to the fault map;

[0082] Step 5: Extract features from the fault map to obtain various structured features, and fuse them with unstructured data to obtain trainable data;

[0083] Step 6: Using the isolation forest algorithm and the trainable data to train a fault prediction model for the power distribution cabinet equipment;

[0084] Step 7: Input the real-time monitoring data obtained by the real-time monitoring module into the trained equipment failure prediction model to predict the failure probability of the equipment in real time;

[0085] Step 8: When the failure probability exceeds the preset threshold, the system automatically sends out a warning signal, and the inspection personnel conduct maintenance according to the alarm signal;

[0086] Step 9: Synchronize the successful warning records, error warning records and equipment failure records that have not been warned to the historical failure data in real time to form new historical failure data.

[0087] In step 1, historical fault data from various sources of 100 distribution cabinet devices are first collected, including but not limited to sensor data (current, voltage, temperature), image data, log files, fault and maintenance records; the collected data are then classified according to type, device, time sequence or other relevant attributes for easy management and retrieval; finally, data from different sources are integrated into a unified database or data warehouse to obtain historical fault data.

[0088] In step 2, firstly, relevant data is extracted from the historical fault data; then, the multimodal data is fused, and the multimodal data including but not limited to text, image, and sensor data are fused into a text field data; among them, for the text conversion of the image, the image data is converted into a text field using image recognition technology, such as recognizing the monitoring picture of the main metering reactive power meter and giving the recognition result: "The main metering reactive power meter of No. XX is reversed"; for the text conversion of the sensor data, an automatic conversion program is designed, such as: converting "13:51, HZ-XX, 83" into "13:51, the voltage of the power supply of the No. XX closing electromagnet control is 83%, which is less than 85%, and an abnormality occurs"; then a series of cleaning techniques are applied, including but not limited to removing stop words, correcting The text field data is cleaned by correcting spelling errors and standardizing named entities to improve data consistency and availability; then the fused text is segmented using natural language processing technology, and part-of-speech tagging is performed to identify key words such as nouns and verbs; then, the fault-related keywords and phrases are extracted from the segmentation results through pattern matching or machine learning methods; finally, based on the extracted key information, the text field data is classified into different fault types (including but not limited to circuit breaker closing failure, incoming line cabinet main switch opening failure, main metering reactive power meter reversal or non-rotation failure, capacitor cabinet automatic switching failure, drawer outlet electric energy meter stoppage failure, dual power switching automatic switching failure failure) or equipment status categories to form target data.

[0089] In step three, a fault map is constructed according to the target data, and an automatic real-time updating mechanism for the fault map is established. The specific details will be described in detail in Example 2.

[0090] In step 4, by referring to the entities, entity attributes and relationships between entities in the fault map, high-precision sensors or monitoring cameras are added to monitor the operating status of key equipment in real time, including but not limited to temperature, pressure, vibration, current, voltage, images, and videos; the data collected by high-precision sensors or monitoring cameras are transmitted to the central database in real time through the industrial Internet of Things platform to provide the equipment fault prediction model with real-time operating data of the equipment; in subsequent operations, maintenance experts continuously improve the real-time monitoring module of the equipment based on the update of the fault map and the actual operating conditions.

[0091] In step 5, firstly, the entity and attribute features are extracted by extracting the node attributes of each device in the fault map and the detailed information of each maintenance; among which, the possible entities include distribution cabinet model, voltage level, current capacity, circuit breaker, capacitor cabinet power, undervoltage coil, controller fuse, capacitor compensation, closing button, closing electromagnet, opening coil, burnout, poor contact, and inability to close; then, by extracting the relationship between the various components of the equipment and analyzing the fault propagation path between the components, the relationship features are extracted, such as (opening and closing station, downlink, busbar), (ambient temperature is too high, resulting in circuit breaker failure), (current overload, resulting in controller fuse burnout); then, according to formulas (1), (2), and (3), the degree centrality, closeness centrality, and betweenness centrality of each device or component node in the fault map are calculated respectively to identify the key components:

[0092]

[0093]

[0094]

[0095] Among them, C D (v) represents the degree centrality of the node, deg(v) represents the degree of the node, and n-1 represents the total number of nodes in the graph except node v; C C (v) represents the closeness centrality of the node, d(v,u) represents the shortest path length from node v to any other node u, ∑ u≠v d(v,u) means summing the shortest path lengths from v to all nodes u different from v; C B (v) represents the betweenness centrality of the node, s and t represent any two different nodes in the graph, σ st represents the number of all shortest paths from node s to node t, σ st (v) represents the number of shortest paths from node s to node t through node v; these centrality indicators reflect the importance and connectivity of nodes in the graph.

[0096] Next, the path and sequence features are extracted by extracting the sequence pattern of fault occurrence and analyzing the sequence of maintenance operations, including the order and frequency of fault occurrence; then the time series features are extracted by extracting the timestamps of fault occurrence and maintenance operations; then the most influential features are selected using methods including but not limited to correlation analysis and mutual information, including the period and time distribution of fault occurrence; then dimensionality reduction techniques including but not limited to PCA and t-SNE are applied to reduce the number of features and improve the efficiency of model training; finally, structured features (including entity attribute features, relationship features, centrality indicators, path and sequence features, time series features) and unstructured data (such as fault report text, sensor data, log files) are fused to obtain trainable data with extracted features, and the trainable data is a structured data table, in which each column represents a feature and each row represents a sample.

[0097] In step six, the trainable data is first fused into a unified dataset, for example, the feature vectors of text, image and sensor data are concatenated together to form a comprehensive feature vector; the fused dataset is then split into a training set and a test set, usually in a certain ratio, such as 80% training set and 20% test set; the isolation forest model is then initialized; the isolation forest model is then trained using the training set data, and the model will learn the "normal" patterns of normal data and identify abnormal points that deviate from these patterns; the trained model is then used to calculate the anomaly score of the test set data, and based on the calculated anomaly score, the predict method is used to perform anomaly detection on the test set data, and the model will mark each data point as normal (1) or abnormal (-1); the performance of the model is then evaluated using metrics including but not limited to precision, recall and F1 score; finally, the model's prediction results are analyzed to identify which data points are identified as abnormal by the model, which may be signs of equipment failure.

[0098] In step 7, the real-time monitoring data obtained in the equipment real-time monitoring module is first preprocessed to obtain input data suitable for model prediction; then an anomaly score S(x) is calculated for each data point x according to formula (4):

[0099]

[0100] Where d(x) is the distance from the data point x to its nearest outlier;

[0101] Then, S(x) is normalized using formula (5) to obtain an abnormality score ranging from 0 to 1, which is the probability of equipment failure:

[0102]

[0103] In step eight, the alarm threshold T is finally set to determine whether the probability of equipment failure exceeds the alarm threshold T. If it exceeds, an alarm is issued:

[0104]

[0105] After the alarm is issued, the staff will inspect and repair the equipment according to the alarm signal. The alarm signals may include: the main switch of the C20 incoming cabinet is abnormally opened, which may be due to poor contact of the opening button or burnt-out of the opening coil; the main reactive meter of H53 is reversed or not rotating, which may be due to high capacitor compensation power factor or wrong phase sequence of the transformer incoming line.

[0106] In step nine, the staff will enter the successful warning records, error warning records and equipment failure records that have not been warned into the database in real time. The system will then automatically extract the sensor data and recent log file data at the time of the current equipment failure to form new historical failure data; the system will update the fault map and equipment failure prediction model in real time based on the new historical failure data.

[0107] According to the above technical solution, the expected results are as follows:

[0108] Reduce unexpected downtime: Assuming that without an early warning system, the distribution cabinet equipment will experience two unexpected downtimes per month on average, after using the fault prediction method, the number of unexpected downtimes is expected to be reduced to once a month;

[0109] Reduce maintenance costs: Calculate the average maintenance cost of each unexpected downtime, assuming it is 10,000 yuan. By reducing the number of downtimes, the monthly maintenance cost of 100 machines can be saved by 1 million yuan;

[0110] Improve production efficiency: Calculate the production loss caused by downtime. Assuming that the production loss is 100 units per downtime, after using the fault prediction method, the monthly increase in production is expected to be 100 units.

[0111] Reduce major repeated failures by 30%, reduce failures with a repair time of more than 60 minutes by 20% year-on-year, and reduce single failure processing time by 30%;

[0112] Accurate prediction of spare parts inventory reduces costs by 15% and equipment maintenance hours by 30%;

[0113] Real-time monitoring of equipment safety door signals and status reduces safety risks for personnel maintenance.

[0114] Example 2

[0115] Different from the systematic description of the entire invention in Example 1, Example 2 mainly describes the specific steps of constructing the fault map in step 3 of Example 1; the present invention proposes a method for constructing a fault map, a knowledge map specifically for key equipment fault prediction, such as Figure 2 As shown, it mainly includes the following steps:

[0116] Step 3.1: Extract knowledge from the target data to obtain ontology structured data;

[0117] Step 3.2: Integrate the ontology structured data with the third-party database data and align the entities to obtain complete structured data;

[0118] Step 3.3: Construct the fault graph ontology based on the complete structured data;

[0119] Step 3.4: Perform knowledge reasoning on the fault graph through logical reasoning algorithm and graph neural network model;

[0120] Step 3.5: Complete the fault graph through the heterogeneous graph attention network;

[0121] Step 3.6: Establish an automated knowledge update mechanism so that new data can be incorporated into the fault map in a timely manner;

[0122] Step 3.7: Perform quality assessment on the fault map, including but not limited to completeness assessment, consistency assessment, accuracy assessment, coverage assessment, readability assessment, and scalability assessment;

[0123] Step 3.8: Store the fault graph in the form of a graph database to ensure efficient query and analysis.

[0124] In step 3.1, knowledge extraction from data includes the following three aspects:

[0125] 1. Identify entities related to equipment failure in the target data based on the BERT model, including but not limited to equipment components, failure types, and operating conditions;

[0126] Second, the Bi-LSTM technology is used to capture the bidirectional dependency of the text and the Attention mechanism is used to focus on key information for relationship extraction;

[0127] 3. Extracting the attributes of entities (e.g., component specifications, severity of faults) based on graph neural networks (GNNs);

[0128] After the above three steps, structured ontology data can be obtained.

[0129] In step 3.2, if there is a third-party database (such as a historical fault database of a certain device in different factories), the structured ontology data can be integrated and entity aligned with the third-party database data to obtain complete structured data. This operation can enrich the data content and provide a more complete data foundation for subsequent work.

[0130] In step 3.3, unlike the ontology construction of knowledge graphs such as language large models, the data range of equipment failures is bounded and easier to structure. Based on this feature, a top-down approach is used to construct the ontology of the fault graph based on several entities, several attributes corresponding to the entities, and entity relationships.

[0131] In step 3.4, logical reasoning algorithms and graph neural network models are used to discover implicit relationships between knowledge and enrich and improve the content of the fault map. The logical reasoning algorithm can be selected from the following algorithms according to the actual application scenario, including but not limited to the FOIL rule learning algorithm, the AMIE rule mining algorithm, the PRA path sorting algorithm, and the probabilistic relationship model based on the Bayesian network. These algorithms can be used alone or in combination to improve the accuracy and efficiency of equipment fault prediction.

[0132] In step 3.5, the KGHAN model is first constructed based on the heterogeneous graph attention network (HAN) model. On the basis of the HAN model, the fault knowledge structure information and fault graph structure information are added to fully learn the deep expression of fault entities and fault relations. The KGHAN model is then used to improve the accuracy of the fault entity concept completion task, that is, to predict the type label of the fault entity. Finally, a specific fault relationship is selected by calculating the similarity between fault entities to achieve fault relationship link completion.

[0133] In step 3.6, an automated update mechanism is established to update the knowledge graph in real time with the new data immediately after receiving it, ensuring that the new data can be incorporated into the knowledge graph in a timely manner;

[0134] In step 3.7, an automated evaluation mechanism is established. Once a change in the fault map is detected, the quality of the fault map is immediately evaluated. The new fault map is evaluated through methods including but not limited to completeness assessment, consistency assessment, accuracy assessment, coverage assessment, readability assessment, and scalability assessment, and an evaluation report and adjustment suggestions are given.

[0135] In step 3.8, the fault graph is stored in the form of a graph database. You can choose Neo4j, which is a high-performance NoSQL graph database that supports the Cypher query language and is suitable for processing complex relational data to ensure efficient query and analysis.

[0136] The present invention adds a multimodal data fusion module to data preprocessing, which fuses data including but not limited to text, image, and sensor data into a text field, thereby providing more comprehensive data support for fault map construction; the BERT model is used to identify entities related to equipment failures. The BERT model captures the contextual information of the text through bidirectional encoding, which can better understand the text, making the model more accurate in identifying named entities related to equipment failures; the Bi-LSTM technology is used to capture the bidirectional dependency of the text and the Attention mechanism is used to focus on key information. The two technologies are combined to improve the accuracy of relationship extraction; the attributes of the entity are extracted through graph neural networks (GNNs), which can be more accurate. The graph structure of the fault map can be well processed; the fault map can be completed by the heterogeneous graph attention network, which can improve the embedding representation quality of fault entities and fault relationships, and improve the accuracy of fault diagnosis and prediction; the implicit relationship between knowledge can be discovered through logical reasoning algorithms and graph neural network models, enriching and improving the content of the fault map; by establishing an automated knowledge update mechanism, it can ensure that new data can be included in the fault map in a timely manner and the fault map can be updated; the fault prediction model is trained by the isolation forest algorithm, and its high sensitivity to outliers in the data can be used to train a fault prediction model with higher accuracy; by establishing a real-time entry mechanism for new data, real-time updating of the fault map and fault prediction model can be achieved.

[0137] The beneficial effects of the present invention are as follows: by establishing a fault map that is more suitable for equipment fault prediction, early warning and fault diagnosis of key equipment failures can be achieved more accurately, the reliability and safety of equipment can be improved, the service life of equipment can be extended, the cost of manual maintenance can be reduced, production efficiency can be enhanced, and decision support can be provided, thereby bringing significant economic and operational benefits to the enterprise; in addition, the method ensures the continuous updating and high quality of the knowledge map through automated knowledge updating and quality assessment mechanisms, so that the fault prediction model can adapt to new data and environmental changes and maintain prediction accuracy and reliability.

[0138] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A key equipment failure prediction method based on historical big data and failure maps, characterized in that: The following steps are involved: Organize historical failure data of selected types of equipment based on big data technology; Perform data preprocessing on historical fault data to obtain target data; Construct a fault map based on target data; Establish a real-time monitoring module for equipment based on the fault map; Extract features from the fault map to obtain various structured features, and fuse the structured features with unstructured data to obtain trainable data; Training a device failure prediction model using an isolation forest algorithm and the trainable data; The real-time monitoring data obtained by the real-time monitoring module is input into the trained equipment failure prediction model to predict the failure probability of the equipment in real time; When the probability of failure exceeds the preset threshold, the system automatically sends out a warning signal, and the inspection personnel conduct maintenance according to the alarm signal; The successful warning records, error warning records and equipment failure records that were not warned are synchronized to the historical failure data in real time to form new historical failure data.

2. The key equipment failure prediction method based on historical big data and failure map according to claim 1 is characterized in that: The method of collating historical fault data of selected types of equipment based on big data technology includes: collecting historical fault data of equipment from various sources, including sensor data, log files, fault and maintenance records, to obtain historical fault data.

3. The key equipment failure prediction method based on historical big data and failure map according to claim 1 is characterized in that: The data preprocessing of the historical fault data to obtain the target data includes: Extract relevant data from historical fault data, including sensor data, log file data, fault and maintenance record data; Fusion of multimodal data including text, images, and sensor data into one text field data; Applying cleaning techniques to clean text field data to improve data consistency and usability, the cleaning techniques include removing stop words, correcting spelling errors, and standardizing named entities; Using natural language processing technology to perform word segmentation on the fused text field data and perform part-of-speech tagging to identify key words, including nouns and verbs; Extract the key words and phrases related to the fault from the word segmentation results through pattern matching or machine learning methods; Based on the extracted keywords and phrases, the text field data is classified into different fault types or equipment status categories to form target data.

4. The key equipment failure prediction method based on historical big data and failure map according to claim 1 is characterized in that: The constructing of a fault map according to the target data includes: Step 3.1: Extract knowledge from the target data to obtain ontology structured data, including: Based on the BERT model, entities related to equipment failure in the target data are identified, including equipment components, failure types, and operating conditions; The Bi-LSTM technology is used to capture the bidirectional dependencies of the text and the Attention mechanism is used to focus on key information for relationship extraction; Extracting entity attributes based on graph neural networks, including component specifications and fault severity; Step 3.2: Integrate the ontology structured data with the third-party database data and align the entities to obtain complete structured data; Step 3.3: Construct the fault graph ontology based on the complete structured data; Step 3.4: Perform knowledge reasoning on the fault graph through logical reasoning algorithm and graph neural network model; Step 3.5: Complete the fault graph through the heterogeneous graph attention network; Step 3.6: Establish an automated knowledge update mechanism so that new data can be incorporated into the fault map in a timely manner; Step 3.7: Perform quality assessment on the fault map, including completeness assessment, consistency assessment, accuracy assessment, coverage assessment, readability assessment, and scalability assessment; Step 3.8: Store the fault map in the form of a graph database.

5. The key equipment failure prediction method based on historical big data and failure map according to claim 4 is characterized in that: The real-time monitoring module of the equipment is established according to the fault map, including: Refer to the entities, entity attributes and relationships between entities in the fault map, add high-precision sensors or surveillance cameras to monitor the operating status of key equipment in real time, and obtain the equipment's temperature data, pressure data, vibration data, current data, voltage data, image data, and video data; The data collected by high-precision sensors or surveillance cameras are transmitted to the central database in real time through the industrial Internet of Things platform, providing real-time operation data of the equipment for the equipment failure prediction model.

6. The key equipment failure prediction method based on historical big data and failure map according to claim 4 is characterized in that: The feature extraction of the fault spectrum includes: By extracting the node attributes of each device in the fault map and the detailed information of each maintenance, the entity features and attribute features are extracted. The entity features include the device parts, fault types, and operating conditions, and the attribute features include the specifications, status, and maintenance history of the equipment. By extracting the relationship between the components of the equipment and analyzing the fault propagation path between the components, the relationship features are extracted, and the features include the composition of the components and the influence between them; According to formulas (1), (2), and (3), the degree centrality, closeness centrality, and betweenness centrality of each device or component node in the fault map are calculated to identify key components: Among them, C D (v) represents the degree centrality of the node, deg(v) represents the degree of the node, and n-1 represents the total number of nodes in the graph except node v; C C (v) represents the closeness centrality of the node, d(v,u) represents the shortest path length from node v to any other node u, ∑ u≠v d(v,u) means summing the shortest path lengths from v to all nodes u different from v; C B (v) represents the betweenness centrality of the node, s and t represent any two different nodes in the graph, σ st represents the number of all shortest paths from node s to node t, σ st (v) represents the number of shortest paths from node s to node t through node v; Extract path and sequence features, including the order and frequency of failures, by extracting the sequence patterns of failures and analyzing the sequence of maintenance operations; Extract time series features by extracting the timestamps of fault occurrence and maintenance operations, including the period and time distribution of fault occurrence; Use correlation analysis and mutual information methods to select the most influential features, including the period and time distribution of fault occurrence; Apply PCA and t-SNE dimensionality reduction techniques to reduce the number of features and improve model training efficiency; The structured features and unstructured data are fused to obtain trainable data with extracted features. The trainable data is a structured data table, in which each column represents a feature and each row represents a sample. The structured features include entity attribute features, relationship features, centrality indicators, path and sequence features, and time series features. The unstructured data includes fault report text, sensor data, and log files.

7. The key equipment failure prediction method based on historical big data and failure map according to claim 1 is characterized in that: The method of using the isolation forest algorithm and the trainable data to train the equipment failure prediction model includes: Merge the trainable data into a unified dataset, and split the fused dataset into a training set and a test set; Initialize the isolation forest model and use the training set data to train the isolation forest model. The isolation forest model learns the "normal" pattern of normal data and identifies abnormal points that deviate from the "normal" pattern. Then use the trained model to calculate the anomaly score of the test set data, and use the predict method to detect anomalies on the test set data based on the calculated anomaly score. The isolation forest model marks each data point as normal (1) or abnormal (-1). Use precision, recall, and F1 score indicators to evaluate the performance of the model. Analyze the prediction results of the isolation forest model and identify the data points identified as abnormal by the model.

8. The key equipment failure prediction method based on historical big data and failure maps according to claim 1 is characterized in that: The real-time monitoring data obtained by the real-time monitoring module is input into the trained equipment failure prediction model to predict the failure probability of the equipment in real time, including: Perform data preprocessing on the real-time monitoring data obtained in the equipment real-time monitoring module to obtain input data suitable for model prediction; The processed data is input into the equipment fault prediction model for fault prediction. The specific process is as follows: The isolation forest model calculates an anomaly score S(x) for each data point x according to formula (4): Where d(x) is the distance from the data point x to its nearest outlier; Then, S(x) is normalized using formula (5) to obtain an abnormality score ranging from 0 to 1, which is the probability of equipment failure:

9. The key equipment failure prediction method based on historical big data and failure maps as claimed in claim 8 is characterized in that: When the failure probability exceeds the preset threshold, the system automatically sends out an early warning signal, and the inspection personnel conduct maintenance according to the alarm signal, specifically including: Set the alarm threshold T to determine whether the probability of equipment failure exceeds the alarm threshold T. If it exceeds, an alarm is issued: After the alarm is sounded, the staff will inspect and repair the equipment according to the alarm signal.

10. The key equipment failure prediction method based on historical big data and failure maps according to claim 1 is characterized in that: The successful warning records, error warning records and equipment failure records that have not been warned are synchronized to the historical failure data in real time to form new historical failure data, including: There are two situations when the staff conducts maintenance on the equipment. One is that after the early warning system issues an early warning, the staff conducts maintenance according to the early warning information; the other is that the early warning system does not issue an early warning, and the staff conducts maintenance on the equipment after the equipment has failed. Among them, the first situation is divided into successful early warning and error early warning; The staff will enter the successful warning records, error warning records and equipment failure records that have not been warned into the database in real time. The system will then automatically extract the sensor data of the current equipment failure and the recent log file data to form new historical failure data; The system updates the fault map and equipment failure prediction model in real time based on new historical fault data.

Images