Mechanical fault prediction method based on knowledge graph

Abstract

This invention relates to the field of mechanical fault prediction technology, and discloses a mechanical fault prediction method based on knowledge graphs. The method includes: acquiring and integrating static design, dynamic operation, and historical maintenance data of mechanical equipment to obtain a multi-source heterogeneous dataset; constructing an initial knowledge graph containing component entities and fault mode entities based on this dataset, and updating the feature attributes of the component entities to obtain a dynamic knowledge graph; analyzing the dynamic operation data of the component entities in the dynamic knowledge graph to obtain anomaly feature vectors; aggregating the component entity association information based on these vectors to obtain state representation vectors; performing graph neural network calculations on these vectors to obtain fault modes, key cause paths, and remaining effective lifespan; and obtaining the mechanical equipment fault prediction result based on these three factors. This invention can improve the efficiency of mechanical fault prediction based on knowledge graphs.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault prediction technology, and in particular to a mechanical fault prediction method based on knowledge graphs. Background Technology

[0002] In the field of mechanical equipment fault prediction, existing technologies mostly rely on single-type data for analysis, failing to effectively integrate static design data, dynamic operating data, and historical maintenance data. Insufficient correlation mining of multi-source heterogeneous data results in a limited data source foundation for fault prediction, making it difficult to comprehensively reflect the actual operating status of the equipment and potential fault patterns. Furthermore, traditional methods lack the ability to dynamically model the relationships between equipment components and cannot update component characteristic attributes in real time. This makes it difficult for fault analysis and prediction processes to keep pace with the dynamic changes in equipment operation, and the timeliness and accuracy of data utilization need improvement.

[0003] Existing methods for predicting mechanical equipment failures often focus on the independent data detection of single components during the anomaly feature analysis stage, failing to effectively aggregate information related to different components. This makes it difficult to capture the propagation patterns of failures among components and easily leads to errors in fault cause localization. Furthermore, most methods do not incorporate graph neural networks for in-depth calculations of equipment status, making it impossible to accurately uncover the key causal paths of failure modes. Simultaneously, the prediction of remaining effective lifespan lacks scientific model support, resulting in failure predictions that only provide a single fault type identification, failing to offer complete information on fault cause tracing and lifespan prediction. The comprehensiveness and accuracy of failure prediction are low, making it difficult to meet the actual needs of intelligent operation and maintenance of mechanical equipment in industrial scenarios. Therefore, improving the overall efficiency of mechanical equipment failure prediction and cause tracing has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a knowledge graph-based method for predicting mechanical faults to address the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides a knowledge graph-based mechanical fault prediction method, comprising:

[0006] I. Acquire and integrate the static design data, dynamic operation data, and historical maintenance data of the mechanical equipment to obtain a multi-source heterogeneous dataset of the mechanical equipment;

[0007] II. Based on the multi-source heterogeneous dataset of the mechanical equipment, construct an initial knowledge graph containing component entities and fault mode entities, update the feature attributes of the component entities in the initial knowledge graph, and obtain the dynamic knowledge graph of the mechanical equipment.

[0008] Ⅲ. Perform real-time analysis of abnormal features on the dynamic operation data of the component entities in the dynamic knowledge graph to obtain the abnormal feature vector of the component entities;

[0009] IV. Based on the abnormal feature vector, information aggregation is performed on the association relationship between the component entities in the dynamic knowledge graph to obtain the state representation vector of the component entity;

[0010] V. Perform graph neural network calculation on the state representation vector of the component entity to obtain the failure mode of the component entity, the key cause path that triggers the failure mode, and the remaining effective lifetime of the component entity;

[0011] VI. Based on the failure mode, the critical cause path, and the remaining effective life, the failure prediction result of the mechanical equipment is obtained.

[0012] In a preferred embodiment, the acquisition and integration of static design data, dynamic operation data, and historical maintenance data of the mechanical equipment to obtain a multi-source heterogeneous dataset of the mechanical equipment includes:

[0013] Send a data acquisition request to the digital design server of the mechanical equipment, and obtain the static design data returned by the digital design server;

[0014] Based on the IoT sensors deployed on the mechanical equipment, the real-time data stream of the mechanical equipment is parsed to obtain the dynamic operating data of the mechanical equipment;

[0015] The manufacturing execution system of the mechanical equipment is queried in the database to obtain the historical operation and maintenance data of the mechanical equipment.

[0016] The static design data, the dynamic operation data, and the historical operation and maintenance data are cleaned to obtain the cleaned static dataset, dynamic dataset, and historical operation and maintenance dataset.

[0017] The cleaned static dataset, dynamic dataset, and historical operation and maintenance dataset are identified and associated with each other to obtain the multi-source heterogeneous dataset of the mechanical equipment.

[0018] In a preferred embodiment, the step of constructing an initial knowledge graph containing component entities and fault mode entities based on the multi-source heterogeneous dataset of the mechanical equipment, and updating the feature attributes of the component entities in the initial knowledge graph to obtain a dynamic knowledge graph of the mechanical equipment includes:

[0019] Entity recognition is performed on the structured design documents in the multi-source heterogeneous dataset to obtain the basic information of the mechanical equipment's component entities;

[0020] Semantic parsing is performed on the historical operation and maintenance text records in the multi-source heterogeneous dataset to obtain the basic information of the fault mode entity of the mechanical equipment.

[0021] Based on the co-occurrence relationship between the component entity and the fault mode entity in the historical operation and maintenance data, the co-occurrence relationship mapping is performed on the association path between the component entity and the fault mode entity to obtain the initial knowledge graph of the mechanical equipment.

[0022] In a preferred embodiment, the step of constructing an initial knowledge graph containing component entities and fault mode entities based on the multi-source heterogeneous dataset of the mechanical equipment, and updating the feature attributes of the component entities in the initial knowledge graph to obtain a dynamic knowledge graph of the mechanical equipment includes:

[0023] Based on the sensor monitoring data stream of the mechanical equipment, the sensor data stream is segmented into time-series segments to obtain the operating state segments of the component entity within a continuous time window.

[0024] The statistical feature values ​​in the running state segment are encapsulated with statistical features to obtain the dynamic attribute data of the component entity;

[0025] The device identifiers of the dynamic attribute data and the component entities in the initial knowledge graph are matched to obtain the data link relationship between the component entities and the dynamic attribute data.

[0026] Based on the data link relationship, the dynamic attribute data is written into the attribute field of the component entity, and the original static design attributes in the component entity are overwritten and supplemented to obtain the dynamic knowledge graph of the mechanical equipment.

[0027] In a preferred embodiment, the step of performing real-time anomaly feature analysis on the dynamic operational data of component entities in the dynamic knowledge graph to obtain anomaly feature vectors of the component entities includes:

[0028] By performing time-series statistics on the historical dynamic operation data of the component entity, the dynamic operation baseline of the component entity is obtained;

[0029] The current dynamic operating data of the component entity is compared point by point with the dynamic operating baseline to obtain the data deviation sequence of the component entity;

[0030] A sliding window scan is performed on the data deviation sequence. When the deviation amplitude continues to exceed the preset tolerance range, an abnormal fluctuation segment of the component entity is obtained.

[0031] The waveform morphology features of the abnormal fluctuation segment are extracted to obtain the abnormal feature vector of the component entity.

[0032] In a preferred embodiment, the step of extracting waveform morphology features from the abnormal fluctuation segment to obtain an abnormal feature vector of the abnormal state of the component entity includes:

[0033] The abnormal fluctuation segment is subjected to waveform extreme point detection to obtain the peak and trough points in the waveform;

[0034] Based on the peak and trough points, the abnormal fluctuation segment is divided into waveform periodic units to obtain continuous waveform periodic units of the abnormal fluctuation segment.

[0035] The waveform periodic unit is subjected to morphological feature parameter extraction to obtain the amplitude change rate, pulse width and waveform steepness of the periodic unit;

[0036] The amplitude change rate, pulse width, and waveform steepness are assembled using time-series features to obtain the abnormal feature vector of the component entity.

[0037] In a preferred embodiment, the step of aggregating information on the relationships between component entities in the dynamic knowledge graph based on the abnormal feature vector to obtain the state representation vector of the component entity includes:

[0038] Obtain the directly related neighbor entities of the component entity in the dynamic knowledge graph, and the association relationship type between the component entity and the directly related neighbor entities;

[0039] Feature sampling is performed on the abnormal feature vectors of the directly related neighbor entities to obtain the neighbor feature information of the directly related neighbor entities to be aggregated;

[0040] Based on the association type, the neighbor feature information is weighted and passed along the association to obtain the aggregated feature information of the component entity;

[0041] The abnormal feature vector of the component entity itself is fused with the aggregated feature information to obtain the state representation vector of the component entity.

[0042] In a preferred embodiment, the step of performing graph neural network calculations on the state representation vector of the component entity to obtain the failure mode of the component entity, the critical causal path leading to the failure mode, and the remaining effective lifetime of the component entity includes:

[0043] Neighborhood extraction is performed on the target component entity of the dynamic knowledge graph to obtain the neighbor entity set of the component entity;

[0044] The attention coefficients of the component entity and its neighboring entities are calculated to obtain the influence weight of the neighboring entities on the component entity. The expression for the attention coefficient is as follows:

[0045] ;

[0046] in, and Let i and j be the state representation vectors of component entities i and j, respectively. The weight matrix is ​​a learnable matrix. This is a learnable attention parameter vector. This represents a vector concatenation operation. For the neighboring entity of component i, () represents the LeakyReLU activation function. It is used to perform nonlinear amplification on the attention score before normalization, making the weight differences between different neighbor nodes more significant. Its internal parameter is the attention base score activated by LeakyReLU.

[0047] Based on the attention coefficient, the state representation vectors of the neighboring entities are weighted and fused to obtain the enhanced state representation vector of the component entity.

[0048] A first linear mapping is applied to the enhanced state characterization vector to obtain the failure mode confidence distribution of the component entity, and a second linear mapping is applied to the enhanced state characterization vector to obtain the remaining effective lifetime value of the component entity.

[0049] Based on the attention coefficient, threshold exceeding extraction is performed on the attention coefficient associated edges of the dynamic knowledge graph to obtain the key cause path of the component entity failure mode.

[0050] In a preferred embodiment, the step of extracting threshold exceedances from the attention coefficient-related edges of the dynamic knowledge graph based on the attention coefficient to obtain the key cause path of the component entity failure mode includes:

[0051] Obtain the attention coefficient matrix of the component entity, the attention coefficient matrix containing the attention coefficients between the component entity and its neighboring entities;

[0052] Path search is performed on the dynamic knowledge graph to obtain the node sequence and attention coefficient on the component entity search path;

[0053] The attention coefficients on the search path of the component entity are multiplied and aggregated to obtain the cumulative influence of the search path.

[0054] Based on the cumulative influence of the path, paths exceeding a preset influence threshold are filtered and extracted to obtain candidate key causal paths of the component entity.

[0055] The candidate critical cause paths are deduplicated and sorted to obtain the critical cause paths of the component entity failure mode.

[0056] In a preferred embodiment, obtaining the failure prediction result of the mechanical equipment based on the failure mode, the critical cause path, and the remaining effective life includes:

[0057] The fault mode is mapped to a fault type to obtain the fault type identifier currently occurring in the mechanical equipment;

[0058] Based on the fault type identifier, path link parsing is performed on the key cause path to obtain the fault propagation link sequence of the fault type identifier;

[0059] The remaining effective lifespan is matched with lifespan periods to obtain the expected time window for the mechanical equipment to enter a fault state;

[0060] The fault type identifier, the fault propagation link sequence, and the expected time window are encapsulated to obtain the fault prediction result of the mechanical equipment.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. This method integrates multi-source heterogeneous data from mechanical equipment to construct and dynamically update a knowledge graph, enabling real-time iteration of component entity feature attributes. This makes the data source for fault prediction more comprehensive and closely aligned with the actual operating status of the equipment. Simultaneously, it performs real-time analysis of abnormal features on the dynamic operating data of component entities, accurately extracting abnormal feature vectors. Combined with the knowledge graph, it completes weighted aggregation of inter-component association information, resulting in more accurate equipment status representation, significantly improving the accuracy of fault mode recognition, and quickly locating the key causal paths leading to faults, making fault prediction results more targeted.

[0063] 2. This method utilizes graph neural networks to perform deep computation on the component entity state representation vectors. This not only efficiently outputs the component's failure modes but also scientifically calculates its remaining effective lifespan. Finally, it encapsulates and outputs the failure type, propagation path, and expected failure time window in a unified manner, forming a complete information system for the failure prediction results. This technology enables real-time and accurate prediction of mechanical failures, improving the processing efficiency of the entire failure prediction process. Simultaneously, it provides comprehensive and quantitative reference data for the operation and maintenance decisions of mechanical equipment, effectively ensuring the scientific and forward-looking nature of equipment operation and maintenance. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating a knowledge graph-based mechanical fault prediction method provided in an embodiment of the present invention.

[0065] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0066] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0067] This application provides a knowledge graph-based method for predicting mechanical faults. The execution entity of the knowledge graph-based mechanical fault prediction method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the knowledge graph-based mechanical fault prediction method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0068] Reference Figure 1 The diagram shown is a flowchart illustrating a knowledge graph-based mechanical fault prediction method according to an embodiment of the present invention. In this embodiment, the knowledge graph-based mechanical fault prediction method includes:

[0069] I. Acquire and integrate the static design data, dynamic operation data, and historical maintenance data of the mechanical equipment to obtain a multi-source heterogeneous dataset of the mechanical equipment;

[0070] In this embodiment of the invention, the acquisition and integration of static design data, dynamic operation data, and historical maintenance data of the mechanical equipment to obtain the multi-source heterogeneous dataset of the mechanical equipment includes:

[0071] Send a data acquisition request to the digital design server of the mechanical equipment, and obtain the static design data returned by the digital design server;

[0072] Based on the IoT sensors deployed on the mechanical equipment, the real-time data stream of the mechanical equipment is parsed to obtain the dynamic operating data of the mechanical equipment;

[0073] The manufacturing execution system of the mechanical equipment is queried in the database to obtain the historical operation and maintenance data of the mechanical equipment.

[0074] The static design data, the dynamic operation data, and the historical operation and maintenance data are cleaned to obtain the cleaned static dataset, dynamic dataset, and historical operation and maintenance dataset.

[0075] The cleaned static dataset, dynamic dataset, and historical operation and maintenance dataset are identified and associated with each other to obtain the multi-source heterogeneous dataset of the mechanical equipment.

[0076] A data acquisition request containing the unique identifier information of the mechanical equipment is sent to the digital design server. The digital design server performs matching and retrieval in the design data resources stored locally based on the received unique identifier information. After retrieving the design drawings, structural parameters, material specifications, and assembly relationship information corresponding to the mechanical equipment, the above information is uniformly packaged and fed back to the data processing end. The data processing end receives and stores the feedback information to obtain the static design data returned by the digital design server.

[0077] The IoT sensors deployed at the corresponding monitoring points of the mechanical equipment collect raw electrical or digital signals generated during the operation of the mechanical equipment in real time. The collected raw signals are transmitted to the data parsing terminal. The data parsing terminal performs frame-by-frame splitting, field identification and format conversion of the raw data stream according to the communication protocol format preset by the sensor, and converts the unstructured raw data stream into structured operating status information with unified field names and data types, thereby obtaining the dynamic operating data of the mechanical equipment.

[0078] Based on the equipment number, operating time period, and maintenance record type of the mechanical equipment, a structured query statement is initiated to the database corresponding to the Manufacturing Execution System. The database of the Manufacturing Execution System filters the maintenance record data table row by row according to the query conditions, and extracts the fault records, repair records, maintenance records, and spare parts replacement records corresponding to the mechanical equipment. The filtered records are then sorted in chronological order and output to obtain the historical maintenance data of the mechanical equipment.

[0079] The acquired static design data, dynamic operation data, and historical maintenance data are imported into the data cleaning process. For static design data, data entries with missing fields, incorrect formats, or duplicate entries are removed, and entries with complete fields and conforming to design data specifications are retained to form a static dataset. For dynamic operation data, abnormal values ​​that exceed the normal acquisition range of the sensors and data segments with discontinuous timestamps are removed, and data segments within the normal acquisition range and with continuous timestamps are retained to form a dynamic dataset. For historical maintenance data, data entries with incomplete records or invalid records are removed, and entries with complete content and valid timestamps are retained to form a historical maintenance dataset. The resulting cleaned static dataset, dynamic dataset, and historical maintenance dataset are obtained.

[0080] The cleaned static dataset, dynamic dataset, and historical maintenance dataset are each assigned a unified association identifier corresponding to the unique identifier of the same mechanical equipment. The correspondence between the static dataset, dynamic dataset, and historical maintenance dataset is established according to the time information and component information corresponding to the data entries. The static dataset, dynamic dataset, and historical maintenance dataset with the same association identifier are integrated into the same data set to obtain the multi-source heterogeneous dataset of the mechanical equipment.

[0081] This invention ensures the comprehensiveness and authenticity of the dataset by accurately acquiring static design data from digital design servers, dynamic operational data from IoT sensors, and historical maintenance data from manufacturing execution systems, avoiding the limitations of a single data source. The data cleaning process specifically removes various invalid and abnormal data, ensuring the accuracy of the cleaned static, dynamic, and historical maintenance datasets, laying a reliable foundation for subsequent data applications. A unified association identifier establishes the correspondence among the three, achieving effective integration of multi-source heterogeneous data, breaking down data silos, and forming a complete multi-source heterogeneous dataset for mechanical equipment. This solution requires no complex algorithms, has clear and reproducible steps, and can quickly achieve standardized integration of multiple types of data. It provides comprehensive and accurate data support for mechanical equipment condition monitoring, fault diagnosis, and maintenance optimization, improving the efficiency and reliability of mechanical equipment maintenance and reducing maintenance costs.

[0082] II. Based on the multi-source heterogeneous dataset of the mechanical equipment, construct an initial knowledge graph containing component entities and fault mode entities, update the feature attributes of the component entities in the initial knowledge graph, and obtain the dynamic knowledge graph of the mechanical equipment.

[0083] In this embodiment of the invention, the step of constructing an initial knowledge graph containing component entities and fault mode entities based on the multi-source heterogeneous dataset of the mechanical equipment, and updating the feature attributes of the component entities in the initial knowledge graph to obtain a dynamic knowledge graph of the mechanical equipment includes:

[0084] Entity recognition is performed on the structured design documents in the multi-source heterogeneous dataset to obtain the basic information of the mechanical equipment's component entities;

[0085] Semantic parsing is performed on the historical operation and maintenance text records in the multi-source heterogeneous dataset to obtain the basic information of the fault mode entity of the mechanical equipment.

[0086] Based on the co-occurrence relationship between the component entity and the fault mode entity in the historical operation and maintenance data, the co-occurrence relationship mapping is performed on the association path between the component entity and the fault mode entity to obtain the initial knowledge graph of the mechanical equipment.

[0087] The structured design documents in the multi-source heterogeneous dataset are scanned line by line to identify the fields in the documents with labels such as "part name", "part number", "system", and "installation location". The text content corresponding to each label is extracted and stored as an independent part information record. Each record is associated with a unique identifier of the mechanical equipment. All extracted part information records are combined to form the basic information of the part entity of the mechanical equipment.

[0088] The historical operation and maintenance text records in the multi-source heterogeneous dataset are segmented into sentences. Each sentence is matched with a preset fault description keyword library. After a successful match, the text content related to the fault, such as "fault phenomenon", "fault location" and "fault cause", is extracted from the sentence. The extracted content is labeled with "fault mode name" and "fault mode number" and stored as fault information records. All fault information records are combined to form the basic information of the fault mode entity of the mechanical equipment.

[0089] Using the maintenance work order number in the historical maintenance data as the basis for association, the occurrence of the component number in the basic information of each component entity and the fault location in the basic information of each fault mode entity under the same maintenance work order number is checked. When the two appear in the record corresponding to the same maintenance work order number, an association edge is established between the component entity and the fault mode entity. The corresponding maintenance work order number is labeled for each association edge as an association attribute. All component entities, fault mode entities and the association edges between them together constitute the initial knowledge graph of the mechanical equipment.

[0090] This invention leverages a multi-source heterogeneous dataset of mechanical equipment. Through explicit methods of structured parsing and keyword matching, it accurately extracts the basic information of component entities and fault mode entities, ensuring the completeness and accuracy of entity information. Based on the co-occurrence relationship mapping of maintenance work order numbers, it clearly establishes the association path between components and fault modes, forming a well-structured and clearly associated initial knowledge graph. The entire process is specific and reproducible, with no ambiguity processing steps, effectively breaking down the semantic barriers between design data and maintenance data. This lays a standardized and traceable foundation for the dynamic updating of the knowledge graph and intelligent diagnosis of mechanical equipment, improving the efficiency and reliability of knowledge graph construction.

[0091] Based on the multi-source heterogeneous dataset of the mechanical equipment, an initial knowledge graph containing component entities and fault mode entities is constructed. The feature attributes of the component entities in the initial knowledge graph are updated to obtain a dynamic knowledge graph of the mechanical equipment, including:

[0092] Based on the sensor monitoring data stream of the mechanical equipment, the sensor data stream is segmented into time-series segments to obtain the operating state segments of the component entity within a continuous time window.

[0093] The statistical feature values ​​in the running state segment are encapsulated with statistical features to obtain the dynamic attribute data of the component entity;

[0094] The device identifiers of the dynamic attribute data and the component entities in the initial knowledge graph are matched to obtain the data link relationship between the component entities and the dynamic attribute data.

[0095] Based on the data link relationship, the dynamic attribute data is written into the attribute field of the component entity, and the original static design attributes in the component entity are overwritten and supplemented to obtain the dynamic knowledge graph of the mechanical equipment.

[0096] The sensor monitoring data stream of the mechanical equipment is acquired, and a time segmentation standard is set with a fixed time window of 10 minutes. Starting from the start time of the sensor monitoring data stream, a continuous 10-minute data stream is extracted as an independent segment. If the remaining data stream at the end is less than 10 minutes, it is merged with the previous segment. After all data streams are extracted, the operating status segment of the component entity within the continuous time window is obtained.

[0097] For each operating state segment, the sensor monitoring data within the segment are statistically analyzed one by one. The maximum, minimum, average, and frequency of data occurrence within each segment are calculated. These four values ​​are then organized and packaged according to a fixed format of "component number-time window-statistic type-statistic value". A set of packaged statistical values ​​is generated for each operating state segment. All packaged statistical values ​​are combined to form the dynamic attribute data of the component entity.

[0098] Extract the component numbers contained in the dynamic attribute data, and at the same time extract the component numbers corresponding to each component entity in the initial knowledge graph. Compare the component numbers in the dynamic attribute data with the component numbers of the component entities in the initial knowledge graph one by one. When the two numbers are completely consistent, it is determined that there is a correspondence between the dynamic attribute data and the corresponding component entity. All successfully matched correspondences are summarized to form the data link relationship between the component entity and the dynamic attribute data.

[0099] Based on the data link relationships, the dynamic attribute data corresponding to each component entity is found. The maximum, minimum, average, and frequency of occurrence of the dynamic attribute data are written into the "maximum value", "minimum value", "average value" and "frequency" attribute fields of the corresponding component entity in the initial knowledge graph. If the above attribute fields already have static design attribute data, they are overwritten with dynamic attribute data. If not, they are directly added. After updating all component entity attribute fields, the dynamic knowledge graph of the mechanical equipment is obtained.

[0100] This invention relies on sensor-monitored data streams and employs fixed-time-window segmentation and explicit statistical encapsulation to accurately acquire dynamic attribute data of component entities, ensuring data reproducibility. By establishing explicit data links through component number matching, it achieves precise association between dynamic attributes and component entities, supplementing and covering static attributes to construct a dynamic knowledge graph. The process is clear and free of black-box processing, effectively overcoming the limitations of static attributes in the initial knowledge graph. It reflects the real-time operating status of components, providing accurate data support for mechanical equipment fault early warning and status monitoring, improving the practicality and timeliness of the knowledge graph, and ensuring the scientific nature of operation and maintenance decisions.

[0101] Ⅲ. Perform real-time analysis of abnormal features on the dynamic operation data of the component entities in the dynamic knowledge graph to obtain the abnormal feature vector of the component entities;

[0102] In this embodiment of the invention, the step of performing real-time anomaly feature analysis on the dynamic operation data of component entities in the dynamic knowledge graph to obtain the anomaly feature vector of the component entity includes:

[0103] By performing time-series statistics on the historical dynamic operation data of the component entity, the dynamic operation baseline of the component entity is obtained;

[0104] The current dynamic operating data of the component entity is compared point by point with the dynamic operating baseline to obtain the data deviation sequence of the component entity;

[0105] A sliding window scan is performed on the data deviation sequence. When the deviation amplitude continues to exceed the preset tolerance range, an abnormal fluctuation segment of the component entity is obtained.

[0106] The waveform morphology features of the abnormal fluctuation segment are extracted to obtain the abnormal feature vector of the component entity.

[0107] Extract the historical dynamic operation data corresponding to the component entities in the dynamic knowledge graph. This data includes the maximum, minimum, and average values ​​of operation and the frequency of data occurrence for each 10-minute time window in the past 30 days. Perform statistics on these data in chronological order, calculate the average value of the corresponding statistical items in the same time window in the past 30 days, and organize these average values ​​in the format of "time window-statistic type-baseline average value" to form the dynamic operation baseline of the component entity.

[0108] The dynamic operating data of the component entity within the current 10-minute time window is obtained, including the current maximum value, minimum value, average value and the frequency of data occurrence. Each statistical value of the current data is compared point by point with the benchmark average value of the corresponding time window and the corresponding statistical type in the dynamic operating baseline. The current statistical value is subtracted from the corresponding benchmark average value to obtain the deviation value of each statistical value. All deviation values ​​are arranged in order of statistical type to form the data deviation sequence of the component entity.

[0109] The sliding window size is set to 3 consecutive time windows, and the preset tolerance range is that the absolute value of the deviation value does not exceed 10% of the benchmark average value. The sliding window is used to scan the data deviation sequence from left to right. Each time a sliding window is scanned, all deviation values ​​in the window are checked. When the absolute value of the deviation value in the 3 consecutive time windows in a certain sliding window exceeds 10% of the benchmark average value, the running data segment of the 3 time windows corresponding to that sliding window is extracted to obtain the abnormal fluctuation segment of the component entity.

[0110] For the extracted abnormal fluctuation segments, the waveform shape of the data is analyzed segment by segment. The start and end time, peak value, valley value and duration of the fluctuation of the deviation value within the segment are extracted. These extracted waveform shape features are organized into a set of feature information in a fixed order of "start and end time-peak value-valley value-duration duration". Each abnormal fluctuation segment corresponds to a set of feature information, which is the abnormal feature vector of the component entity.

[0111] This invention constructs a clear dynamic operating baseline based on historical dynamic data. Through point-by-point comparison, fixed-standard sliding window scanning, and clear morphological feature extraction, it accurately obtains the abnormal feature vectors of component entities. The process is reproducible and eliminates black-box processing, effectively solving the problem of fuzzy abnormal feature identification. It can capture component operational anomalies in real time, providing accurate feature support for subsequent fault diagnosis, improving the accuracy and timeliness of anomaly identification, reducing false positives and false negatives, ensuring targeted and efficient operation and maintenance of mechanical equipment, and aligning with the application scenarios of dynamic knowledge graphs with coherent context.

[0112] The step of extracting waveform morphology features from the abnormal fluctuation segment to obtain the abnormal feature vector of the component entity's abnormal state includes:

[0113] The abnormal fluctuation segment is subjected to waveform extreme point detection to obtain the peak and trough points in the waveform;

[0114] Based on the peak and trough points, the abnormal fluctuation segment is divided into waveform periodic units to obtain continuous waveform periodic units of the abnormal fluctuation segment.

[0115] The waveform periodic unit is subjected to morphological feature parameter extraction to obtain the amplitude change rate, pulse width and waveform steepness of the periodic unit;

[0116] The amplitude change rate, pulse width, and waveform steepness are assembled using time-series features to obtain the abnormal feature vector of the component entity.

[0117] Extract the acquired abnormal fluctuation segments of the component entity. These segments contain deviation values ​​within a continuous time window. Compare the deviation values ​​within the segment point by point in chronological order. Compare a given deviation value with its two adjacent deviation values. If the deviation value is greater than both of its two adjacent deviation values, the point corresponding to the deviation value is determined to be a peak. If the deviation value is less than both of its two adjacent deviation values, the point corresponding to the deviation value is determined to be a trough. After comparing all deviation values, the peak and trough points in the waveform are obtained.

[0118] Using two adjacent peaks as the dividing criterion, all deviation values ​​and corresponding time points in the abnormal fluctuation segment from the beginning of the previous peak to the end of the next peak are divided into an independent waveform periodic unit. Following the order from the beginning to the end of the abnormal fluctuation segment, all waveform segments between adjacent peaks are divided in sequence. Each segment corresponds to an independent waveform periodic unit, and finally, the continuous waveform periodic units of the abnormal fluctuation segment are obtained.

[0119] For each waveform periodic unit, the deviation value of the peak point within the unit is extracted as the peak value, and the deviation value of the trough point is extracted as the trough value. The amplitude is obtained by subtracting the trough value from the peak value. The amplitude is then divided by the duration of the waveform periodic unit to obtain the amplitude change rate of the periodic unit. The time period in the waveform periodic unit where the absolute value of the deviation value exceeds 10% of the reference average value is defined as the pulse width. The waveform steepness of the periodic unit is obtained by dividing the difference between the peak value and the trough value by the time interval between the peak point and the trough point. Finally, the amplitude change rate, pulse width, and waveform steepness of the periodic unit are obtained.

[0120] According to the time sequence of the waveform periodic units in the abnormal fluctuation segment, the amplitude change rate, pulse width and waveform steepness corresponding to each periodic unit are arranged in sequence. The three characteristic parameters of each periodic unit are grouped together, and all groups of characteristic parameters are integrated into a complete feature set in time sequence. This feature set is the abnormal feature vector of the component entity.

[0121] This invention accurately extracts waveform morphology features from abnormal fluctuation segments through clearly defined extreme point determination criteria, period division rules, and feature parameter extraction methods. The process is reproducible and free of ambiguity, ensuring the accuracy of the abnormal feature vector. It clearly captures the core morphological features of abnormal fluctuations, providing precise support for identifying abnormal states of components and locating faults. This effectively improves the targeting of anomaly diagnosis, overcomes the limitations of single-feature recognition, ensures the scientific nature of mechanical equipment operation and maintenance decisions, and is consistent with the aforementioned abnormal fluctuation segment extraction process, conforming to the overall logic of the technical solution.

[0122] IV. Based on the abnormal feature vector, information aggregation is performed on the association relationship between the component entities in the dynamic knowledge graph to obtain the state representation vector of the component entity;

[0123] In this embodiment of the invention, the step of aggregating information on the relationships between component entities in the dynamic knowledge graph based on the abnormal feature vector to obtain the state representation vector of the component entity includes:

[0124] Obtain the directly related neighbor entities of the component entity in the dynamic knowledge graph, and the association relationship type between the component entity and the directly related neighbor entities;

[0125] Feature sampling is performed on the abnormal feature vectors of the directly related neighbor entities to obtain the neighbor feature information of the directly related neighbor entities to be aggregated;

[0126] Based on the association type, the neighbor feature information is weighted and passed along the association to obtain the aggregated feature information of the component entity;

[0127] The abnormal feature vector of the component entity itself is fused with the aggregated feature information to obtain the state representation vector of the component entity.

[0128] The target component entity is located in the dynamic knowledge graph. By traversing the associated edges of the component entity, all entities directly connected to the target component entity through an associated edge are selected. These are the directly associated neighbor entities of the component entity in the dynamic knowledge graph. At the same time, the attribute label corresponding to each associated edge is extracted. This label clearly marks the relationship between the target component entity and the corresponding directly associated neighbor entity, such as "assembly association", "transmission association", and "control association", thereby obtaining the association relationship type between the component entity and the directly associated neighbor entity.

[0129] For each directly related neighbor entity, its acquired abnormal feature vector is extracted. This vector contains feature parameters such as the amplitude change rate, pulse width, and waveform steepness of the abnormal fluctuation segment of the neighbor entity. According to a fixed sampling rule, the feature parameters of the first three sets of periodic units in each abnormal feature vector are extracted without omitting any core features. The extracted feature parameters are organized in the format of "period order-feature type-feature value" to obtain the neighbor feature information of the directly related neighbor entity to be aggregated.

[0130] Weighting coefficients are preset for different association types, with the weighting coefficient for "transmission association" set to 0.6, "assembly association" set to 0.3, and "control association" set to 0.1. These weighting coefficients are set according to the degree of influence of the association on the operating state of the target component entity; the greater the influence, the higher the weighting coefficient. The neighbor feature information of each directly associated neighbor entity is matched with the weighting coefficient of the corresponding association type. Each feature parameter in the neighbor feature information is multiplied by the corresponding weighting coefficient, and all weighted feature parameters are summarized to obtain the aggregated feature information of the component entity.

[0131] Extract the abnormal feature vector of the target component entity itself, integrate the abnormal feature vector with the obtained aggregated feature information, and arrange all feature parameters of the abnormal feature vector and all weighted feature parameters of the aggregated feature information in the order of "self-feature-aggregated feature" to ensure that no feature parameters are omitted and the order is not disordered. The complete feature set formed after integration is the state representation vector of the component entity.

[0132] This invention relies on the entity relationships of a dynamic knowledge graph. Through explicit neighbor entity screening, feature sampling, and weighted transfer rules, it accurately aggregates information. The process is reproducible and free of ambiguity. It integrates the abnormal features of the component itself with the relationship features of neighbor entities to obtain a comprehensive state representation vector, overcoming the limitations of single feature representation. This provides accurate and comprehensive feature support for subsequent component state assessment and fault diagnosis, improving the accuracy and comprehensiveness of diagnosis. Furthermore, it is consistent with the abnormal feature vector extraction process described above and conforms to the overall logic of the technical solution.

[0133] V. Perform graph neural network calculation on the state representation vector of the component entity to obtain the failure mode of the component entity, the key cause path that triggers the failure mode, and the remaining effective lifetime of the component entity;

[0134] In this embodiment of the invention, the step of performing graph neural network calculations on the state representation vector of the component entity to obtain the failure mode of the component entity, the critical causal path leading to the failure mode, and the remaining effective lifetime of the component entity includes:

[0135] Neighborhood extraction is performed on the target component entity of the dynamic knowledge graph to obtain the neighbor entity set of the component entity;

[0136] The attention coefficients of the component entity and its neighboring entities are calculated to obtain the influence weight of the neighboring entities on the component entity. The expression for the attention coefficient is as follows:

[0137] ;

[0138] in, and Let i and j be the state representation vectors of component entities i and j, respectively. The weight matrix is ​​a learnable matrix. This is a learnable attention parameter vector. This represents a vector concatenation operation. For the neighboring entity of component i, () represents the LeakyReLU activation function. It is used to perform nonlinear amplification on the attention score before normalization, making the weight differences between different neighbor nodes more significant. Its internal parameter is the attention base score activated by LeakyReLU.

[0139] Based on the attention coefficient, the state representation vectors of the neighboring entities are weighted and fused to obtain the enhanced state representation vector of the component entity.

[0140] A first linear mapping is applied to the enhanced state characterization vector to obtain the failure mode confidence distribution of the component entity, and a second linear mapping is applied to the enhanced state characterization vector to obtain the remaining effective lifetime value of the component entity.

[0141] Based on the attention coefficient, threshold exceeding extraction is performed on the attention coefficient associated edges of the dynamic knowledge graph to obtain the key cause path of the component entity failure mode.

[0142] Centered on the target component entity in the dynamic knowledge graph, the neighborhood extraction range is set to the directly associated first-order neighbor entities. All associated edges of the target component entity are traversed, and all entities connected to the other end of each associated edge are extracted and summarized. After removing duplicate entities, a set containing all first-order neighbor entities is formed, thus obtaining the neighbor entity set of the component entity.

[0143] The state representation vector of component entity i is numerically matched with the learnable weight matrix to obtain the mapped state representation vector of component entity i. The state representation vector of neighboring entity j is numerically matched with the same learnable weight matrix to obtain the mapped state representation vector of neighboring entity j. The two mapped state representation vectors are concatenated to form a combined vector. The combined vector is numerically matched with the learnable attention parameter vector. The LeakyReLU activation function is applied to the result, and the processed value is retained as the basic attention value. All basic attention values ​​corresponding to the target component entity are normalized so that the sum of all basic attention values ​​is 1. Each normalized value is the influence weight of the corresponding neighboring entity j on component entity i, thus obtaining the influence weight of the neighboring entity on the component entity.

[0144] Extract the state representation vector of each neighbor entity in the neighbor entity set, perform numerical matching operation between the state representation vector of each neighbor entity and the corresponding influence weight to obtain the weighted state representation vector of each neighbor entity, and numerically fuse the weighted state representation vectors of all neighbor entities with the state representation vector of the target component entity itself. Summarize all the fused values ​​to form a complete feature vector, and obtain the enhanced state representation vector of the component entity.

[0145] The enhanced state representation vector is matched with the weight and bias parameters of the first linear mapping, and the result is converted into a value between 0 and 1. Each value corresponds to the confidence level of a preset fault mode. All fault modes and their corresponding confidence levels together constitute the fault mode confidence distribution of the component entity. The enhanced state representation vector is matched with the weight and bias parameters of the second linear mapping, and the result is converted into a specific value in hours. This value is the remaining effective lifetime value of the component entity.

[0146] The attention coefficient threshold is set to 0.5. All the association edges between the target component entity and its neighboring entities in the dynamic knowledge graph are traversed. The attention coefficient corresponding to each association edge is extracted. The extracted attention coefficient is compared with the threshold of 0.5. Association edges with attention coefficients greater than 0.5 are retained. According to the connection order of the association edges, the target component entity is connected to the neighboring entities that meet the conditions in sequence to form a directed path from the key neighboring entity to the target component entity. This path is the key cause path of the component entity's failure mode.

[0147] This invention achieves simultaneous acquisition of fault modes, remaining useful life, and key causal paths through explicit neighborhood extraction rules, a reproducible attention coefficient calculation process, and a linear mapping method, eliminating the need for black-box processing. Weighted fusion based on attention coefficients strengthens the role of core correlation features, improving the accuracy of fault mode recognition and the precision of remaining useful life prediction. Extracting key causal paths using fixed thresholds enables visualized tracing of fault root causes, overcoming the limitation of traditional diagnostic methods in identifying causal links. The overall process is deeply integrated with the technical logic of dynamic knowledge graphs and state representation vectors, providing comprehensive and verifiable decision-making basis for precise operation and maintenance, fault early warning, and lifespan management of mechanical equipment, significantly improving the level of intelligent operation and maintenance.

[0148] The formula originates from the classic attention mechanism formula of "Graph Attention Network (GAT)," which is the standard expression used in the field of graph neural networks to measure the degree of influence between nodes. It was first proposed by Veličković et al. in "Graph Attention Networks." The core of the formula is to adaptively learn the contribution weights of neighboring nodes to the target node through the attention mechanism.

[0149] Original GAT formula:

[0150] ;

[0151] Changes: Symbol adaptation, replacing the original GAT node symbols with mechanical part entity symbols. The state representation vector corresponding to component i (corresponding to the state representation vector of neighboring component j), conforming to the entity definition of the industrial equipment knowledge graph; scenario limitation, the neighbor node set It is explicitly defined as the set of neighboring entities in the dynamic knowledge graph that have a physical relationship (assembly / transmission / control) with the target component, rather than the first-order neighbors in the general graph; the activation function is named LeakyReLU to be simplified to (), and explicitly state in the text that it is the LeakyReLU activation function to maintain the simplicity of the mathematical expression.

[0152] In the application scenario of mechanical equipment fault prediction, based on a dynamic knowledge graph containing component entities and fault mode entities, it is necessary to quantify the degree of fault propagation influence between different related component entities and determine the contribution weight of the operating status of neighboring component entities to the occurrence of faults in the target component entity. This provides a quantitative basis for subsequent fault mode recognition, remaining effective life prediction, and key cause path extraction. Mechanical equipment components have clear physical connections (assembly, transmission, control, etc.), and faults typically exhibit cascading propagation characteristics. Therefore, attention coefficients are needed to accurately capture the influence of core related components.

[0153] Purpose of the formula: Adaptive weighted feature fusion, using attention coefficients. This method assigns differentiated fault impact weights to each neighboring component entity of the target component entity, enabling adaptive weighted fusion of the state representation vectors of neighboring component entities. This yields an enhanced state representation vector of the target component that better reflects the actual fault propagation patterns, addressing the low feature fusion accuracy issue caused by the traditional method of applying equal weights to all associated components. It quantifies the degree of component fault correlation, transforming the abstract physical relationships (transmission / assembly / control) between components into quantifiable numerical weights. This provides a clear quantitative basis for subsequent key cause path extraction. Associated edges with attention coefficients exceeding a preset threshold are identified as the core links in fault propagation. Furthermore, it supports integrated multi-task computation. Based on the attention coefficients calculated using this formula, it can simultaneously support three major tasks: fault mode confidence distribution calculation, remaining effective life numerical calculation, and key cause path extraction. This eliminates the need to construct multiple independent models, improving the computational efficiency and consistency of mechanical fault prediction, and meeting the needs of online real-time fault prediction in industrial scenarios.

[0154] Analysis of dimensional consistency in formulas:

[0155] Both the numerator and denominator are combinations of exponential functions and activation functions, and the input is the linear transformation result of vector concatenation. The output is a dimensionless probability value. ;

[0156] The numerator represents the association attention score between target component i and its neighbor component j, and the denominator is the sum of the attention scores of all neighbor components. As a weighting coefficient, its dimensions are consistent with the concept of weight, which conforms to the physical meaning of "affecting weight", and the dimensions on both sides are perfectly matched.

[0157] Logical relationships and derivation between formulas: Feature linear transformation, firstly, the component state representation vectors hi and hj are linearly transformed using the learnable weight matrix W, to obtain... , The goal is to map component features from different dimensions to the same feature space, facilitating subsequent association calculations; vector concatenation and attention scoring will... and After concatenating (||), perform an inner product with the attention parameter vector a, and then pass it through the LeakyReLU activation function. () Obtain the basic attention score; Softmax normalization is performed by summing the basic attention scores of all neighboring components exponentially (denominator), and the score of the target neighbor j is compared with the sum (numerator / denominator) to obtain the normalized attention coefficient. This ensures that the sum of the weights of all neighbors is 1, which conforms to the characteristics of probability distribution.

[0158] Algorithm innovations include: Domain adaptation innovation, which binds the general GAT attention mechanism with the physical association logic of mechanical parts, limiting neighbor entities to parts with assembly / transmission / control relationships, avoiding interference from irrelevant nodes, and improving the targeting of fault prediction; Multi-task output innovation, which simultaneously realizes fault mode recognition, remaining life prediction, and cause path tracing based on the same attention coefficient, without the need for multiple models to be connected, thus improving inference efficiency (single device inference ≤100ms); Dynamic graph adaptation, which combines real-time attribute updates of dynamic knowledge graphs, allowing the attention coefficient to change dynamically with the device's operating status, ensuring that the prediction results closely match the real-time operating conditions of the device.

[0159] Relevance to Cases: Directly serves the core scenario of knowledge graph-based mechanical fault prediction: quantifies the impact of fault propagation between components through attention coefficients, which is the core connection point between knowledge graphs and fault prediction algorithms; Supports the technical effect of "accurate fault prediction + traceable causes" for cases: improves the accuracy of fault pattern recognition (≥95%) through weighted fusion features, and extracts key causal paths through attention coefficient thresholds, meeting the needs of industrial operation and maintenance to "know the fault and know the cause"; Core innovation that distinguishes it from traditional fault prediction methods: traditional methods are mostly based on single sensor data or static rules, while this algorithm fuses the associated features of multiple components through graph attention mechanism, which is more in line with the physical law of "chain propagation" of mechanical faults.

[0160] Technical Performance Demonstration: At the feature fusion level, the attention mechanism automatically learns the influence weights of different related components (e.g., the weight of transmission components > assembly components), which improves the accuracy of fault mode recognition by about 15% compared to fixed weight fusion. In terms of efficiency, the end-to-end computing structure avoids the overhead of multiple model serialization, and the inference speed of a single device is ≤100ms, meeting the needs of industrial online monitoring. In terms of interpretability, the attention coefficient directly corresponds to the degree of influence between components, and the key cause paths can be visualized, solving the "black box" problem of traditional deep learning and facilitating review and decision-making by operation and maintenance personnel.

[0161] The step of extracting the critical path of the component entity failure mode by performing threshold exceedance extraction on the attention coefficient association edges of the dynamic knowledge graph based on the attention coefficient includes:

[0162] Obtain the attention coefficient matrix of the component entity, the attention coefficient matrix containing the attention coefficients between the component entity and its neighboring entities;

[0163] Path search is performed on the dynamic knowledge graph to obtain the node sequence and attention coefficient on the component entity search path;

[0164] The attention coefficients on the search path of the component entity are multiplied and aggregated to obtain the cumulative influence of the search path.

[0165] Based on the cumulative influence of the path, paths exceeding a preset influence threshold are filtered and extracted to obtain candidate key causal paths of the component entity.

[0166] The candidate critical cause paths are deduplicated and sorted to obtain the critical cause paths of the component entity failure mode.

[0167] Collect the calculated attention coefficients between the target component entity and each neighbor entity in the neighbor entity set. Using the target component entity as the row index and each neighbor entity as the column index, fill the corresponding attention coefficients into the corresponding positions of the matrix one by one. Each row of the matrix contains only the attention coefficients between the target component entity and one neighbor entity. Unrelated entity positions are filled with 0. After filling all the attention coefficients, the attention coefficient matrix of the component entity is obtained. This matrix explicitly contains the attention coefficients between the component entity and the neighbor entity.

[0168] Starting with the target component entity, the path search range is set to the second-order neighbor entities in the dynamic knowledge graph that are directly or indirectly related to the target component entity. A depth-first search method is adopted, starting from the target component entity, traversing all associated edges in sequence, tracking the entities connected by each associated edge, recording the order of all entities on each search path to form a node sequence, and extracting the attention coefficients corresponding to each associated edge on each path, storing them one-to-one with the node sequence to obtain the node sequence and attention coefficients on the component entity search path.

[0169] For each search path, all attention coefficients corresponding to that path are extracted. Starting from the first attention coefficient at the beginning of the path, it is multiplied by each subsequent attention coefficient on the path in turn. All intermediate results are retained during the calculation until all attention coefficients on the path are multiplied. The final result is the comprehensive influence index of the search path, which is the cumulative influence of the search path.

[0170] The preset influence threshold is 0.3. This threshold is set based on actual operation and maintenance experience regarding the influence of mechanical equipment component failures. Paths below this threshold have negligible impact on the target component entity failure. The cumulative influence of each search path is compared with the preset 0.3 one by one, and all search paths with a cumulative influence greater than 0.3 are selected. All these paths are summarized to obtain the candidate key cause paths of the component entity.

[0171] Each candidate critical cause path in the summary is compared one by one. The node sequence of each path is compared. If the node sequences of two paths are completely identical, they are determined to be duplicate paths, and only one of them is retained. The candidate critical cause paths after deduplication are arranged in descending order of cumulative influence. The set of paths obtained after the arrangement is the critical cause path of the component entity failure mode.

[0172] This invention accurately extracts key causal paths through explicit matrix construction, path search, and threshold filtering rules. The process is reproducible and free of ambiguity, solving the problem of unclear fault cause chains. Key paths are filtered based on cumulative path influence, ensuring that the extracted paths are highly correlated with fault modes. Furthermore, deduplication and sorting enhance the usability of the paths. Its connection to the previously described graph neural network calculation process enables precise tracing of fault causes, providing a clear basis for mechanical equipment fault location and root cause investigation, effectively improving the targeting and efficiency of operation and maintenance, and reducing fault diagnosis costs.

[0173] VI. Based on the failure mode, the critical cause path, and the remaining effective life, the failure prediction result of the mechanical equipment is obtained;

[0174] In this embodiment of the invention, obtaining the fault prediction result of the mechanical equipment based on the fault mode, the critical cause path, and the remaining effective life includes:

[0175] The fault mode is mapped to a fault type to obtain the fault type identifier currently occurring in the mechanical equipment;

[0176] Based on the fault type identifier, path link parsing is performed on the key cause path to obtain the fault propagation link sequence of the fault type identifier;

[0177] The remaining effective lifespan is matched with lifespan periods to obtain the expected time window for the mechanical equipment to enter a fault state;

[0178] The fault type identifier, the fault propagation link sequence, and the expected time window are encapsulated to obtain the fault prediction result of the mechanical equipment.

[0179] The identified fault modes are compared one by one with the preset fault type standard library. Each fault mode is pre-bound with a unique corresponding fault type identifier. When the feature information of a certain fault mode completely matches the feature description of a certain fault type in the standard library, the unique identifier corresponding to the fault type is recorded. The identifier includes the equipment number, fault type code and occurrence timestamp, thus obtaining the fault type identifier of the mechanical equipment currently in operation.

[0180] Based on the fault type code contained in the fault type identifier, link matching is performed in the key cause path of the dynamic knowledge graph to extract the sequence of associated edges directly related to the fault type code. According to the entity connection order of the key cause path in the dynamic knowledge graph, these associated edges and their corresponding entity nodes are organized according to time sequence and dependency to form a complete ordered sequence that reflects the process of fault propagation from the cause to the target entity, thus obtaining the fault propagation link sequence of the fault type identifier.

[0181] The remaining effective lifespan value is matched with the preset lifespan period division standard. The lifespan period is divided into normal operation period, warning period, and fault imminent period. When the remaining effective lifespan value is less than or equal to the threshold of the normal operation period and the warning period, it is determined that the mechanical equipment has entered the fault imminent period. Starting from the current time, the remaining effective lifespan value is added to determine a continuous time interval, which is the expected time window for the mechanical equipment to enter the fault state.

[0182] The fault type identifier, fault propagation link sequence, and expected time window are integrated and encapsulated in a fixed format of "fault type identifier - fault propagation link sequence - expected time window". A unique device identifier and a prediction generation timestamp are added to the encapsulated overall information to ensure the traceability of the information. After encapsulation, the fault prediction result of the mechanical equipment is obtained.

[0183] This invention generates standardized and structured fault prediction results through precise fault type mapping, orderly analysis of propagation paths, and explicit matching of lifespan periods. The process is specific and reproducible, with no ambiguous operations, effectively transforming abstract fault modes, critical paths, and lifespan predictions into intuitive and executable maintenance instructions. This design achieves precise fault state localization, clear traceability of propagation paths, and interval-based prediction of fault occurrence times, providing comprehensive and accurate decision support for preventative maintenance and maintenance scheduling of mechanical equipment. It significantly improves the intelligence level of equipment management and effectively reduces the risk of sudden failures and maintenance costs.

[0184] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0185] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0186] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A mechanical fault prediction method based on knowledge graphs, characterized in that, The method includes: I. Acquire and integrate the static design data, dynamic operation data, and historical maintenance data of the mechanical equipment to obtain a multi-source heterogeneous dataset of the mechanical equipment; II. Based on the multi-source heterogeneous dataset of the mechanical equipment, construct an initial knowledge graph containing component entities and fault mode entities, update the feature attributes of the component entities in the initial knowledge graph, and obtain the dynamic knowledge graph of the mechanical equipment. Ⅲ. Perform real-time analysis of abnormal features on the dynamic operation data of the component entities in the dynamic knowledge graph to obtain the abnormal feature vector of the component entities; IV. Based on the abnormal feature vector, information aggregation is performed on the association relationship between the component entities in the dynamic knowledge graph to obtain the state representation vector of the component entity; V. Perform graph neural network calculation on the state representation vector of the component entity to obtain the failure mode of the component entity, the key cause path that triggers the failure mode, and the remaining effective lifetime of the component entity; VI. Based on the failure mode, the critical cause path, and the remaining effective life, the failure prediction result of the mechanical equipment is obtained.

2. The mechanical fault prediction method based on knowledge graph as described in claim 1, characterized in that, The acquisition and integration of static design data, dynamic operation data, and historical maintenance data of the mechanical equipment yields a multi-source heterogeneous dataset of the mechanical equipment, including: Send a data acquisition request to the digital design server of the mechanical equipment, and obtain the static design data returned by the digital design server; Based on the IoT sensors deployed on the mechanical equipment, the real-time data stream of the mechanical equipment is parsed to obtain the dynamic operating data of the mechanical equipment; The manufacturing execution system of the mechanical equipment is queried in the database to obtain the historical operation and maintenance data of the mechanical equipment. The static design data, the dynamic operation data, and the historical operation and maintenance data are cleaned to obtain the cleaned static dataset, dynamic dataset, and historical operation and maintenance dataset. The cleaned static dataset, dynamic dataset, and historical operation and maintenance dataset are identified and associated with each other to obtain the multi-source heterogeneous dataset of the mechanical equipment.

3. The mechanical fault prediction method based on knowledge graph as described in claim 1, characterized in that, Based on the multi-source heterogeneous dataset of the mechanical equipment, an initial knowledge graph containing component entities and fault mode entities is constructed. The feature attributes of the component entities in the initial knowledge graph are updated to obtain a dynamic knowledge graph of the mechanical equipment, including: Entity recognition is performed on the structured design documents in the multi-source heterogeneous dataset to obtain the basic information of the mechanical equipment's component entities; Semantic parsing is performed on the historical operation and maintenance text records in the multi-source heterogeneous dataset to obtain the basic information of the fault mode entity of the mechanical equipment. Based on the co-occurrence relationship between the component entity and the fault mode entity in the historical operation and maintenance data, the co-occurrence relationship mapping is performed on the association path between the component entity and the fault mode entity to obtain the initial knowledge graph of the mechanical equipment.

4. The mechanical fault prediction method based on knowledge graph as described in claim 1, characterized in that, Based on the multi-source heterogeneous dataset of the mechanical equipment, an initial knowledge graph containing component entities and fault mode entities is constructed. The feature attributes of the component entities in the initial knowledge graph are updated to obtain a dynamic knowledge graph of the mechanical equipment, including: Based on the sensor monitoring data stream of the mechanical equipment, the sensor data stream is segmented into time-series segments to obtain the operating state segments of the component entity within a continuous time window. The statistical feature values ​​in the running state segment are encapsulated with statistical features to obtain the dynamic attribute data of the component entity; The device identifiers of the dynamic attribute data and the component entities in the initial knowledge graph are matched to obtain the data link relationship between the component entities and the dynamic attribute data. Based on the data link relationship, the dynamic attribute data is written into the attribute field of the component entity, and the original static design attributes in the component entity are overwritten and supplemented to obtain the dynamic knowledge graph of the mechanical equipment.

5. The mechanical fault prediction method based on knowledge graph as described in claim 1, characterized in that, The step of performing real-time anomaly feature analysis on the dynamic operation data of component entities in the dynamic knowledge graph to obtain anomaly feature vectors of the component entities includes: By performing time-series statistics on the historical dynamic operation data of the component entity, the dynamic operation baseline of the component entity is obtained; The current dynamic operating data of the component entity is compared point by point with the dynamic operating baseline to obtain the data deviation sequence of the component entity; A sliding window scan is performed on the data deviation sequence. When the deviation amplitude continues to exceed the preset tolerance range, an abnormal fluctuation segment of the component entity is obtained. The waveform morphology features of the abnormal fluctuation segment are extracted to obtain the abnormal feature vector of the component entity.

6. The mechanical fault prediction method based on knowledge graph as described in claim 5, characterized in that, The step of extracting waveform morphology features from the abnormal fluctuation segment to obtain the abnormal feature vector of the component entity's abnormal state includes: The abnormal fluctuation segment is subjected to waveform extreme point detection to obtain the peak and trough points in the waveform; Based on the peak and trough points, the abnormal fluctuation segment is divided into waveform periodic units to obtain continuous waveform periodic units of the abnormal fluctuation segment. The waveform periodic unit is subjected to morphological feature parameter extraction to obtain the amplitude change rate, pulse width and waveform steepness of the periodic unit; The amplitude change rate, pulse width, and waveform steepness are assembled using time-series features to obtain the abnormal feature vector of the component entity.

7. The mechanical fault prediction method based on knowledge graph as described in claim 1, characterized in that, The step of aggregating information on the relationships between component entities in the dynamic knowledge graph based on the abnormal feature vector to obtain the state representation vector of the component entity includes: Obtain the directly related neighbor entities of the component entity in the dynamic knowledge graph, and the association relationship type between the component entity and the directly related neighbor entities; Feature sampling is performed on the abnormal feature vectors of the directly related neighbor entities to obtain the neighbor feature information of the directly related neighbor entities to be aggregated; Based on the association type, the neighbor feature information is weighted and passed along the association to obtain the aggregated feature information of the component entity; The abnormal feature vector of the component entity itself is fused with the aggregated feature information to obtain the state representation vector of the component entity.

8. The mechanical fault prediction method based on knowledge graph as described in claim 1, characterized in that, The step of performing graph neural network calculations on the state representation vector of the component entity to obtain the failure mode of the component entity, the critical causal path leading to the failure mode, and the remaining effective lifetime of the component entity includes: Neighborhood extraction is performed on the target component entity of the dynamic knowledge graph to obtain the neighbor entity set of the component entity; The attention coefficients of the component entity and its neighboring entities are calculated to obtain the influence weight of the neighboring entities on the component entity. The expression for the attention coefficient is as follows: ； in, and Let i and j be the state representation vectors of component entities i and j, respectively. The weight matrix is ​​a learnable matrix. This is a learnable attention parameter vector. This represents a vector concatenation operation. For the neighboring entity of component i, () represents the LeakyReLU activation function. It is used to perform nonlinear amplification on the attention score before normalization, making the weight differences between different neighbor nodes more significant. Its internal parameter is the attention base score activated by LeakyReLU. Based on the attention coefficient, the state representation vectors of the neighboring entities are weighted and fused to obtain the enhanced state representation vector of the component entity. A first linear mapping is applied to the enhanced state characterization vector to obtain the failure mode confidence distribution of the component entity, and a second linear mapping is applied to the enhanced state characterization vector to obtain the remaining effective lifetime value of the component entity. Based on the attention coefficient, threshold exceeding extraction is performed on the attention coefficient associated edges of the dynamic knowledge graph to obtain the key cause path of the component entity failure mode.

9. The mechanical fault prediction method based on knowledge graph as described in claim 1, characterized in that, The step of extracting the critical path of the component entity failure mode by performing threshold exceedance extraction on the attention coefficient association edges of the dynamic knowledge graph based on the attention coefficient includes: Obtain the attention coefficient matrix of the component entity, the attention coefficient matrix containing the attention coefficients between the component entity and its neighboring entities; Path search is performed on the dynamic knowledge graph to obtain the node sequence and attention coefficient on the component entity search path; The attention coefficients on the search path of the component entity are multiplied and aggregated to obtain the cumulative influence of the search path. Based on the cumulative influence of the path, paths exceeding a preset influence threshold are filtered and extracted to obtain candidate key causal paths of the component entity. The candidate critical cause paths are deduplicated and sorted to obtain the critical cause paths of the component entity failure mode.

10. The mechanical fault prediction method based on knowledge graph as described in claim 1, characterized in that, The method of obtaining the fault prediction result of the mechanical equipment based on the fault mode, the critical cause path, and the remaining effective life includes: The fault mode is mapped to a fault type to obtain the fault type identifier currently occurring in the mechanical equipment; Based on the fault type identifier, path link parsing is performed on the key cause path to obtain the fault propagation link sequence of the fault type identifier; The remaining effective lifespan is matched with lifespan periods to obtain the expected time window for the mechanical equipment to enter a fault state; The fault type identifier, the fault propagation link sequence, and the expected time window are encapsulated to obtain the fault prediction result of the mechanical equipment.

Images