Data processing method and device, equipment and medium

Abstract

The invention provides a data processing method and device, equipment and a medium, and relates to the field of data processing, and the technical scheme comprises the following steps: obtaining maintenance business data of an aircraft and a knowledge graph configured for the aircraft, and obtaining recorded time sequence data from a fast access recorder in the aircraft; feature extraction is carried out on time sequence data, maintenance service data and data in a knowledge graph, the data dimension and redundancy are effectively reduced, and simpler data representation is provided for the subsequent feature fusion step. And for each target flight component, inputting the fusion feature and the reference case associated with the time sequence data into the component life prediction model, outputting the residual life prediction result of the target flight component, and introducing analogy reasoning of the reference case, so that the prediction precision of the component life prediction model can be improved. Based on the residual life prediction result of each target flight component, the maintenance strategy information of the aircraft can be generated more accurately.

Description

Technical Field

[0001] This application relates to the field of data processing, and more particularly to a data processing method, apparatus, device, and medium. Background Technology The reliability of each component in an aircraft directly determines flight safety and operating costs; therefore, it is necessary to maintain each component in the aircraft in a timely and accurate manner.

[0002] In traditional technologies, maintenance measures for aircraft are set based on human experience. However, due to the limitations of human experience, the maintenance effect of these measures is poor and the accuracy of the measures is low. Summary of the Invention

[0003] This application provides a data processing method, apparatus, device, medium, and product that can accurately generate maintenance strategies for aircraft, thereby accurately guiding the maintenance of aircraft.

[0004] Firstly, a data processing method is provided, the method comprising: The system acquires maintenance business data and knowledge graphs for aircraft configuration, and retrieves time-series data from the aircraft's fast access recorder. The time-series data retrieved from the fast access recorder is more accurate. By acquiring time-series data, maintenance business data, and knowledge graphs from three dimensions, it provides high-quality data references for subsequent data processing, which is conducive to improving the accuracy of subsequent remaining life prediction.

[0005] The maintenance business data includes aircraft maintenance work order information; entities in the knowledge graph include flight component models, component anomaly types, component maintenance actions, and component parameter thresholds; relationships in the knowledge graph include ternary relationships between flight component models, component parameter thresholds, and component anomaly types, as well as binary relationships between component anomaly types and component maintenance actions; and time-series data includes target time-series information of multiple target flight components contained in multiple subsystems of the aircraft, as well as sensor information of dedicated sensors connected to the target flight components in the aircraft. Target time-series information includes component time-series information of the target flight components and system time-series information of the subsystems. Component parameter thresholds are thresholds configured for the parameter values ​​corresponding to the target flight components. Feature extraction is performed on the target time series information and sensor information to obtain the basic features of the target flight component. Feature extraction is performed on the maintenance work order information to obtain the maintenance business features of the target flight component. Feature extraction is performed on the binary and ternary relations to obtain the graph embedding features of the target flight component. Feature extraction is performed on the data in the time series data, maintenance business data and knowledge graph, which effectively reduces data dimensionality and redundancy, and provides a more concise data representation for the subsequent feature fusion steps.

[0006] Feature fusion is performed based on basic features, maintenance business features, and graph embedding features to obtain fused features. By fusing multiple features, the fused features obtained by feature fusion are of higher quality than feature concatenation, which helps to improve the output accuracy of subsequent prediction models.

[0007] For each target flight component, the fusion features and reference cases associated with time-series data in the knowledge graph are input into the component life prediction model, and the remaining life prediction result of the target flight component is output. The analogy reasoning of the reference cases can improve the prediction accuracy of the component life prediction model.

[0008] This method, on the one hand, generates aircraft maintenance strategy information based on the remaining life prediction results of each target flight component, resulting in more accurate maintenance strategies. On the other hand, by acquiring time-series data, maintenance business data, and knowledge graphs, it considers the time-series dimension, historical maintenance dimension, and expert knowledge dimension of the data, resulting in higher-quality data that helps improve the accuracy of subsequent maintenance strategies. Furthermore, by using a component life prediction model to predict the remaining life of target flight components, it is not limited to human experience. Based on this, the maintenance strategy information generated from the remaining life prediction results is more accurate, which to some extent avoids the phenomenon of insufficient aircraft maintenance and improves maintenance effectiveness.

[0009] In one possible implementation of the first aspect, maintenance strategy information for the aircraft is generated based on the remaining life prediction results of each target flight component. Specifically, this includes: obtaining initial maintenance sub-schemes for each target flight component and a maintenance planning model configured for the aircraft. The initial maintenance sub-schemes and the remaining life prediction results of each target flight component are input into the maintenance planning model, which outputs component maintenance sub-schemes for each target flight component. The maintenance planning model is trained with the objectives of reducing maintenance costs, reducing downtime, and reducing safety risks. The maintenance sub-schemes for each component are then integrated to obtain the aircraft's maintenance strategy information.

[0010] In this embodiment, firstly, an initial maintenance sub-scheme is obtained, eliminating the need to generate a new scheme from scratch. This saves time compared to existing schemes, thus improving the efficiency of generating maintenance strategies. Secondly, the maintenance planning model is trained with the objectives of reducing maintenance costs, minimizing downtime, and mitigating safety risks. Simultaneously optimizing the three core objectives of cost, time, and safety improves the quality of remaining life prediction results. Thirdly, a corresponding component maintenance sub-scheme is configured for each target flight component, achieving component-level maintenance and refined maintenance.

[0011] In one possible implementation of the first aspect, the method further includes: periodically acquiring execution data of maintenance strategy information within a preset time period. The execution data includes the lifespan difference between the actual and predicted lifespan of the target flight component, and the recurrence rate of anomalies in the target flight component after maintenance. Based on the execution data, the maintenance planning model, the component lifespan prediction model, and the knowledge graph are updated.

[0012] In this embodiment, firstly, data is acquired periodically, and model parameters and knowledge graphs are updated based on the acquired execution data, forming a complete data loop and effectively improving the quality of prediction results. Secondly, based on the lifespan difference between the actual and predicted lifespan of the target flight component and the abnormal recurrence rate of the target flight component after maintenance, the model is continuously updated to avoid gradual performance degradation after model training and ensure that the model continuously outputs high-quality prediction results.

[0013] In one possible implementation of the first aspect, feature extraction is performed on the target time-series information and sensor information to obtain the basic features of the target flight components. Specifically, this includes: acquiring the anomaly curves for each target flight component; the anomaly curves characterize the relationship between the anomaly parameter threshold corresponding to the target flight component and the number of flight cycles. Based on the relationship between the target parameter values ​​in the target time-series information and the sensor parameter values ​​in the sensor information and the anomaly parameter thresholds corresponding to the number of flight cycles, outliers in the target parameter values ​​and sensor parameter values ​​are filtered out to obtain intermediate information of the target flight components; the intermediate information includes the target time-series information and sensor information after anomaly filtering. Data completion processing is performed on the intermediate information to obtain the completed information of the target flight components. Feature extraction is performed on the completed information to obtain the basic features of the target flight components.

[0014] In this embodiment, firstly, by acquiring anomaly parameter thresholds (anomaly curves) that change with the number of flight cycles, dynamic thresholds can be applied to the component's lifespan, avoiding misjudgments of anomalies in the target flight component and improving the accuracy of the parameter values ​​corresponding to the target flight component. Secondly, for data loss caused by sensor failure, transmission interruption, or recording errors, data completion processing is performed on intermediate information to obtain completed information, ensuring data integrity and providing comprehensive data reference for subsequent data processing. Thirdly, feature extraction is performed on the completed information to characterize the information, facilitating direct feature processing by the subsequent model.

[0015] In one possible implementation of the first aspect, acquiring recorded timing data from a fast access recorder in the aircraft specifically includes: determining the subsystem to which the target flight component of the aircraft belongs and the dedicated sensor connected to the target flight component. The subsystem includes an engine system, landing gear system, or avionics system, and the dedicated sensor includes an engine fluid sensor or a landing gear strain sensor. The recorded timing data is acquired from the subsystem and the dedicated sensor matching the preset reading frequency range in the fast access recorder of the aircraft, according to a preset reading frequency range in the fast access recorder.

[0016] In this embodiment, firstly, sensor information is acquired from dedicated sensors (oil sensors, strain sensors, etc.). Because these dedicated sensors are designed for the target flight components, the sensor information obtained is more accurate. Secondly, reading time-series data from a fast access recorder ensures precise time alignment of multi-source data, which is helpful for subsequent trend analysis and time-series correlation analysis, thereby improving the accuracy of subsequent maintenance strategy generation. Thirdly, by acquiring data from different data sources according to a preset reading frequency range, the data acquisition range is ensured, providing comprehensive data reference for subsequent data processing.

[0017] In another possible implementation of the first aspect, the method further includes: acquiring multiple preset cases from a knowledge graph and determining the parameter values ​​and component models corresponding to the target flight component; the parameter values ​​include target parameter values ​​or sensor parameter values. Based on the parameter values, component models, and ternary relationships corresponding to the target flight component, the target anomaly type of the target flight component is determined. A search is performed among the preset cases, and preset cases that match both the component model and the target anomaly type are used as reference cases associated with the recorded data.

[0018] In this embodiment, firstly, multiple preset cases are obtained from the knowledge graph, transforming tacit knowledge into an explicit knowledge base. Since the cases are preset and have a unified format, they facilitate subsequent retrieval and comparison. Secondly, retrieval based on the knowledge graph improves retrieval efficiency significantly compared to manually reviewing records. Thirdly, reference cases obtained through automatic reasoning using ternary relations in the knowledge graph enhance the efficiency and accuracy of obtaining reference cases.

[0019] In another possible implementation of the first aspect, feature fusion is performed based on basic features, maintenance business features, and graph embedding features to obtain fused features. Specifically, this includes: obtaining the feature weights of the maintenance business features and the graph embedding features, and obtaining the initial weights of the basic features. If the basic features are trend-changing features or collaborative features, the initial weights of the basic features are increased to obtain the feature weights of the basic features; if the basic features are not trend-changing features or collaborative features, the initial weights are used as the feature weights of the basic features. Based on each feature weight, the basic features, maintenance business features, and graph embedding features are mapped to a feature space. Through a three-dimensional encoding configured for the feature space, including time, decay degree, and semantics, feature fusion is performed on the basic features, maintenance business features, and graph embedding features within the feature space to obtain fused features.

[0020] In this embodiment, firstly, when the basic features are trend-changing features or collaborative features, the initial weight of the basic features is increased, thereby enhancing the weight of important features and preventing important weak signals from being overwhelmed by strong signals. Secondly, different features are mapped to a unified feature space, allowing for quantitative comparison between different features, which facilitates subsequent feature fusion processing. Thirdly, precise temporal context alignment is achieved. Based on 3D encoding, feature fusion is performed on basic features, maintenance business features, and graph embedding features within the feature space, realizing spatiotemporal semantic alignment of features and improving the accuracy of feature fusion.

[0021] Secondly, embodiments of this application provide a data processing apparatus, the apparatus comprising: The acquisition module is used to acquire maintenance business data of the aircraft and knowledge graph configured for the aircraft, and to acquire time-series data recorded from the fast access recorder in the aircraft. The maintenance business data includes maintenance work order information of the aircraft; entities in the knowledge graph include flight component models, component anomaly types, component maintenance actions, and component parameter thresholds; the relationships in the knowledge graph include ternary relationships between flight component models, component parameter thresholds, and component anomaly types, and binary relationships between component anomaly types and component maintenance actions; the time-series data includes target time-series information of multiple target flight components contained in multiple subsystems of the aircraft, and sensor information of dedicated sensors connected to the target flight components in the aircraft; the target time-series information includes component time-series information of the target flight components and system time-series information of the subsystems; and the component parameter thresholds are thresholds configured for parameter values ​​corresponding to the target flight components. The feature extraction module is used to extract features from the target time-series information and the sensor information to obtain the basic features of the target flight component; to extract features from the maintenance work order information to obtain the maintenance business features of the target flight component; and to extract features from the binary relation and the ternary relation to obtain the map embedding features of the target flight component. The feature fusion module is used to perform feature fusion based on the basic features, the maintenance business features, and the map embedding features to obtain fused features; The prediction module is used to input the fused features and the reference cases in the knowledge graph associated with the time series data into the component life prediction model for each target flight component, and output the remaining life prediction result of the target flight component. The information generation module is used to generate maintenance strategy information for the aircraft based on the remaining life prediction results of each of the target flight components. The maintenance strategy information is used to guide the target flight components in maintenance.

[0022] Thirdly, embodiments of this application provide a data processing apparatus, the method comprising: a memory and at least one processor. The memory is communicatively connected to the processor. The memory is used to store computer program code, the computer program code including computer instructions. When the processor executes the computer instructions, it causes the electronic device to perform the method as described in the first aspect and any possible implementation thereof.

[0023] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions. When these computer instructions are executed by a processor, they are used to implement the method as described in the first aspect and any possible implementation thereof.

[0024] Fifthly, embodiments of this application provide a computer program product that, when run on a computer or executed by the computer's processor, implements the method described in the first aspect and any possible design thereof. The computer may be the data processing device described in the third aspect and any possible implementation thereof.

[0025] Understandably, the beneficial effects achieved by the data processing apparatus of the second aspect, the data processing device of the third aspect, the computer-readable storage medium of the fourth aspect, and the computer program product of the fifth aspect provided above can be referred to as the beneficial effects of the first aspect and any possible implementation thereof, which will not be repeated here. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the structure of a data processing system provided in an embodiment of this application; Figure 2 A flowchart illustrating a data processing method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a condition-based maintenance system provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. Detailed Implementation

[0027] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this embodiment, unless otherwise stated, "a plurality of" means two or more.

[0028] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0029] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0030] The reliability of each component in an aircraft directly determines flight safety and operating costs; therefore, it is necessary to maintain each component in the aircraft in a timely and accurate manner.

[0031] In traditional technologies, the first method of aircraft maintenance relies on human experience to set up maintenance measures for the aircraft. However, due to the limitations of human experience, the maintenance effect of these measures is poor and the accuracy of the maintenance measures is low.

[0032] The second method of aircraft maintenance involves performing overall maintenance on a scheduled basis, which has the problem of over-maintenance.

[0033] The third method of aircraft maintenance involves acquiring relevant aircraft data, performing life analysis based on that data, and then performing maintenance based on the analysis results. However, this method only utilizes a small portion of the acquired aircraft data for component life analysis, which relies on empirical formulas, resulting in relatively poor accuracy.

[0034] This application provides a data processing method. Regarding the first approach, on one hand, the data processing method provided in this application generates aircraft maintenance strategy information based on the remaining life prediction results of each target flight component, resulting in a more accurate maintenance strategy. On the other hand, this application acquires time-series data, maintenance business data, and knowledge graphs, considering the time-series dimension, historical maintenance dimension, and expert knowledge dimension, resulting in higher quality data and contributing to improved accuracy of subsequent maintenance strategies. Furthermore, this application uses a component life prediction model to predict the remaining life of target flight components, not limited to human experience. Based on this, the maintenance strategy information generated from the remaining life prediction results is more accurate, to some extent avoiding insufficient aircraft maintenance and improving maintenance effectiveness.

[0035] Regarding the second approach, the data processing method provided in this application embodiment performs life analysis on each target flight component in the aircraft. Based on the remaining life prediction results of each target flight component, maintenance strategy information of the aircraft is generated. The strategy information naturally includes different maintenance strategies corresponding to multiple components, thereby achieving refined maintenance and avoiding the problem of over-maintenance.

[0036] Regarding the third approach, firstly, in the data processing method provided in this application embodiment, the acquired time-series data, maintenance business data, and knowledge graph data are all processed through feature extraction and feature fusion, and predictions are made based on the fused features, thus fully processing the data and eliminating any data omissions. Secondly, the lifespan analysis is based on the component lifespan prediction model and reference cases in the knowledge graph. Compared to empirical formulas, the results output by the component lifespan prediction model are more accurate. Further introducing analogical reasoning from reference cases can further improve the prediction accuracy of the component lifespan prediction model.

[0037] The data processing method provided in this application embodiment can be applied to, for example... Figure 1 The data processing system shown can be connected to the aircraft 101, which includes multiple target flight components 1011, a fast access recorder 1012, and dedicated sensors 1013.

[0038] The data processing system includes an edge processing node 102, a cloud storage cluster 103, and a processing server 104.

[0039] The processing server 104 can acquire maintenance business data and a knowledge graph configured for the aircraft, and retrieve recorded time-series data from the fast access recorder in the aircraft. The maintenance business data includes maintenance work order information for the aircraft; entities in the knowledge graph include flight component models, component anomaly types, component maintenance actions, and component parameter thresholds; the relationships in the knowledge graph include ternary relationships between flight component models, component parameter thresholds, and component anomaly types, and binary relationships between component anomaly types and component maintenance actions; the time-series data includes target time-series information of multiple target flight components contained in multiple subsystems within the aircraft, and sensor information of dedicated sensors connected to the target flight components in the aircraft. The target time-series information includes the component time-series information of the target flight component and the system time-series information of the subsystem; the component parameter thresholds are thresholds configured for the parameter values ​​corresponding to the target flight component. The processing server 104 can extract features from the target time series information and sensor information to obtain the basic features of the target flight component; extract features from the maintenance work order information to obtain the maintenance business features of the target flight component; and extract features from the binary and ternary relations to obtain the graph embedding features of the target flight component. The processing server 104 can perform feature fusion based on basic features, maintenance business features, and graph embedding features to obtain fused features; The processing server 104 can input the fused features and reference cases associated with time-series data in the knowledge graph into the component life prediction model for each target flight component, and output the remaining life prediction result of the target flight component. The processing server 104 can generate maintenance strategy information for the aircraft based on the remaining life prediction results of each target flight component. The maintenance strategy information is used to guide the target flight components to perform maintenance.

[0040] The edge processing node 102 can acquire time-series data, perform noise filtering and anomaly detection on the time-series data, and send the noise-filtered and anomaly-detected time-series data to the processing server 104 for further processing. Noise filtering can be based on Kalman filtering. Kalman filtering is an optimal recursive estimation algorithm for linear dynamic systems. It estimates the true state of the system in real time and continuously by combining the system's dynamic model with measurement data containing uncertainty (noise), filtering out noise in the process.

[0041] The cloud storage cluster 103 can be equipped with a knowledge graph database and model training platform. The cloud storage cluster 103 can be used to store and process data in the server 104, and can also be used to train component lifespan prediction models. Distributed storage can be based on Ceph. Ceph is an open-source, unified, software-defined distributed storage system designed to run on general-purpose hardware, providing three access interfaces: object storage, block storage, and file system storage. It features high scalability, high reliability, and no single point of failure.

[0042] In one embodiment, such as Figure 2 As shown, a data processing method is provided that can be applied to... Figure 1 The method can be implemented in a processing server 104, or in a cloud computing platform, edge computing device, chip, or computing device, etc. This application does not limit the specific form of the device executing the method; taking the application of the method to a processing server 104 as an example, the method specifically includes: S201, acquire maintenance business data of the aircraft and knowledge graph of the aircraft configuration, and retrieve time-series data recorded from the fast access recorder in the aircraft.

[0043] In some embodiments, an aircraft can refer to a device that flies within or outside the atmosphere. An aircraft can include flying devices such as aircraft, spacecraft, and rockets. Aircraft can include fixed-wing aircraft, helicopters, etc.

[0044] Maintenance operational data can be structured data that records the maintenance history of an aircraft. Maintenance operational data can include maintenance work order information, component replacement records, fault reports, oil analysis reports, maintenance personnel qualifications, etc.

[0045] Maintenance work order information may include work order number, maintenance date, personnel performing the work, fault description, replaced parts, test results, and labor costs. Parts replacement records are detailed documents documenting the removal and replacement of specific flight components. Fault reports are formal documents recording anomalies discovered during aircraft operation or inspection. Fluid analysis reports are condition assessment reports generated after laboratory analysis of aircraft fluids such as lubricating oil and hydraulic oil. Maintenance personnel qualifications are a set of certifications proving that maintenance personnel are legally qualified to perform specific flight maintenance work.

[0046] A knowledge graph is a structured semantic knowledge base that stores entities and relationships between entities in the field of flight maintenance in the form of a graph structure. A knowledge graph can be built on Neo4j. Neo4j is a native graph database management system specifically designed for efficiently storing, querying, and processing highly interconnected data. It uses an attribute graph model to store data as entities, relationships, and attributes, and is one of the mainstream technology platforms for building and manipulating knowledge graphs.

[0047] A Quick Access Recorder (QAR) is a device on an aircraft used to record flight parameter data. It collects hundreds to thousands of flight parameters (such as altitude, speed, engine speed, temperature, etc.) at a fixed frequency, and the data can be downloaded quickly via physical media or wirelessly.

[0048] Time-series data can be time-series data obtained from a fast access recorder, which records a set of real-time state parameters of the aircraft in different flight cycles.

[0049] The timing data includes target timing information of multiple target flight components contained in multiple subsystems of the aircraft, as well as sensor information of dedicated sensors connected to the target flight components in the aircraft.

[0050] A subsystem can be a collection of subsystems within an aircraft that perform specific functions. Subsystems may include engine systems, landing gear systems, and avionics systems, among others.

[0051] Target timing information can be the timing characteristic information corresponding to the target flight components in the subsystem. Target timing information can include component timing information of the target flight components and system timing information of the subsystem. Specifically, component timing information can include, in the engine system, high-pressure compressor vibration values, engine exhaust gas temperature (EGT), fuel flow, and the speed fluctuations of the low-pressure rotor (N1) and high-pressure rotor (N2) of the engine. System timing information can include, in the landing gear system, landing vertical load, buffer compression, wheel axle temperature, and brake wear. System timing information can also include, in the avionics system, bus voltage fluctuations, navigation signal signal-to-noise ratio, and Flight Management System (FMS) command response delays.

[0052] Specialized sensors can be testing devices installed on specific flight components to monitor critical parameters of those components. These specialized sensors could be fluid sensors on engines or strain sensors on landing gear.

[0053] Sensor information can be metadata and measurement data for a specific sensor, including sensor type, accuracy, installation location, measurement range, real-time / historical readings, etc. For example, the oil contamination level on an oil sensor, or the component strain value on a strain sensor.

[0054] Entities in a knowledge graph can include flight component models, component anomaly types, component maintenance actions, and component parameter thresholds.

[0055] The model number of a flight component can be the specific model number of a flight component in an aircraft.

[0056] Component anomaly types can be categorized into specific anomaly patterns that a component may exhibit. For example, component anomaly types could include fatigue cracks, stress corrosion cracking, wear, corrosion, and thermal fatigue.

[0057] Component maintenance actions can be standardized maintenance procedures prescribed for specific types of anomalies. Examples include borehole inspection, crack repair, replacement of seals, and sensor calibration.

[0058] The component parameter threshold can be a threshold configured for the parameter value corresponding to the target flight component, that is, the critical value for determining whether the flight component is in a normal or abnormal state.

[0059] Relations in a knowledge graph can include ternary relations and binary relations.

[0060] The ternary relation can be composed of three entities: flight component model, component parameter threshold, and component anomaly type. Specifically, the ternary relation indicates that if the parameter value corresponding to a flight component of a certain model meets the component parameter threshold, then the component is determined to have an anomaly of a certain anomaly type. For example, if the vibration value of a high-pressure compressor is greater than 15g (component parameter threshold), then fatigue cracks (component anomaly type) are determined in that high-pressure compressor.

[0061] A binary relation can consist of two entities: a component anomaly type and a component maintenance action. Specifically, a binary relation can represent the need for a component maintenance action when a component anomaly type occurs. For example, the appearance of a fatigue crack (component anomaly type) requires borehole inspection and crack repair (component maintenance action).

[0062] In one possible implementation, S201, the aircraft's maintenance operational data and a knowledge graph of the aircraft's configuration are acquired, and time-series data recorded is obtained from a fast in-flight storage recorder, specifically including: The processing server can directly retrieve recorded time-series data from the aircraft's storage recorder, and can also obtain aircraft maintenance data and knowledge graphs configured for the aircraft from the cloud storage cluster. By retrieving time-series data from the fast access recorder, the real-time nature of the data is ensured. Retrieving maintenance data and knowledge graphs from the cloud storage cluster allows for faster data retrieval, ensuring efficient data reading and thus improving the efficiency of subsequent data processing.

[0063] In one possible implementation, S201, the recorded timing data is acquired from the fast access recorder in the aircraft, specifically including: determining the subsystem to which the target flight component of the aircraft belongs, and the dedicated sensor connected to the target flight component. The subsystem includes the engine system, landing gear system, or avionics system, and the dedicated sensor includes an engine fluid sensor or a landing gear strain sensor. The recorded timing data is acquired from the subsystem and the dedicated sensor that match the preset reading frequency range in the fast access recorder of the aircraft, according to the preset reading frequency range.

[0064] A subsystem can be a collection of subsystems within an aircraft that perform specific functions. Subsystems may include engine systems, landing gear systems, and avionics systems, among others.

[0065] The target flight component can be a corresponding component within a subsystem. For example, in an engine system, the target flight component can be a high-pressure compressor, the engine, the engine's high-pressure rotor, or the engine's low-pressure rotor, etc.

[0066] Specialized sensors can be testing devices installed on specific flight components to monitor critical parameters of those components. These specialized sensors could be fluid sensors on engines or strain sensors on landing gear.

[0067] The preset read frequency range can be the frequency read range built into the fast access recorder. For example, from 1 Hz to 64 Hz.

[0068] Specifically, the timing data may include target timing information of multiple flight components contained in each of the multiple subsystems within the aircraft, as well as sensor information of dedicated sensors connected to the target flight components in the aircraft. Target timing information may include component timing information of the target flight components and system timing information of the subsystems. The processing server can obtain the system timing information of the subsystems, the component timing information of the target flight components, and the sensor information of the dedicated sensors connected to the target flight components from the fast access recorder in the aircraft.

[0069] In this embodiment, firstly, sensor information is acquired from dedicated sensors (oil sensors, strain sensors, etc.). Because these dedicated sensors are designed for the target flight components, the sensor information obtained is more accurate. Secondly, reading time-series data from a fast access recorder ensures precise time alignment of multi-source data, which is helpful for subsequent trend analysis and time-series correlation analysis, thereby improving the accuracy of subsequent maintenance strategy generation. Thirdly, by acquiring data from different data sources according to a preset reading frequency range, the data acquisition range is ensured, providing comprehensive data reference for subsequent data processing.

[0070] S202, feature extraction is performed on the target time series information and sensor information to obtain the basic features of the target flight component; feature extraction is performed on the maintenance work order information to obtain the maintenance business features of the target flight component; feature extraction is performed on the binary and ternary relations to obtain the map embedding features of the target flight component.

[0071] In some embodiments, the basic features can be mathematical representation vectors extracted from target time-series information and sensor information, reflecting the physical state and operational characteristics of the target flight components. These features are derived from the raw data through signal processing, time-series analysis, and statistical methods, and can quantitatively describe the real-time health status, performance, and degradation trend of the components.

[0072] Maintenance business characteristics can be structured feature vectors extracted from maintenance work order information, reflecting the historical maintenance status and operational background of the target flight component. These features transform unstructured maintenance records into computable mathematical representations, encoding the component's maintenance history, failure modes, and maintenance quality information.

[0073] Graph embedding features are low-dimensional dense vector representations learned from binary and ternary relations in a knowledge graph, reflecting the semantic location and associations of target flight components within a domain knowledge network. These features map discrete entities and relations to a continuous vector space using graph embedding techniques, preserving the topological structure and semantic information of the knowledge graph.

[0074] In one possible implementation, the processing server can extract features from the target's temporal information and sensor information to obtain the basic features of the target flight component. The processing server can also extract features from maintenance work order information to obtain the maintenance service features of the target flight component. Furthermore, the processing server can extract features from binary and ternary relations to obtain the graph embedding features of the target flight component.

[0075] In one possible implementation, the processing server can first perform anomaly detection on the target's timing information and sensor information, filtering out outliers to obtain regular data. The processing server can then extract features from the regular data to obtain the basic characteristics of the target's flight components. By pre-detecting anomalies in the target's timing information and sensor information and filtering out abnormal data, data contamination in subsequent data processing is avoided.

[0076] After data acquisition and before feature extraction, data anomalies may occur. Therefore, preprocessing is required for the data with anomalies before feature extraction. Specifically, in one possible implementation, S202, feature extraction is performed on the target time-series information and sensor information to obtain the basic features of the target flight components. This includes: acquiring the anomaly curves for each target flight component; the anomaly curves characterize the relationship between the anomaly parameter threshold corresponding to the target flight component and the number of flight cycles. Based on the relationship between the target parameter values ​​in the target time-series information and the sensor parameter values ​​in the sensor information and the anomaly parameter thresholds corresponding to the number of flight cycles, anomalies in the target parameter values ​​and sensor parameter values ​​are filtered out to obtain intermediate information of the target flight components; the intermediate information includes the target time-series information and sensor information after anomaly filtering. Data completion processing is performed on the intermediate information to obtain the completed information of the target flight components. Feature extraction is performed on the completed information to obtain the basic features of the target flight components.

[0077] The anomaly curve can be a dynamically changing function or curve used to characterize the dynamic threshold boundary for determining the abnormality of a target flight component's parameters under different flight cycles (or usage time). The anomaly curve is not a fixed value, but rather a threshold function that changes with the component's usage time. It can be understood that the anomaly curve can serve as a two-dimensional degradation benchmark of "flight cycle - parameter change," which allows for dynamic alignment of an individual component with a healthy baseline. For example, if the parameter value of a target flight component is outside the benchmark, it can be considered an anomaly.

[0078] An abnormal parameter threshold can be a specific numerical boundary for determining whether the parameters of a target flight component are abnormal after a certain number of flight cycles. The abnormal parameter threshold is the function value of the abnormality curve at a specific point, representing that a parameter value exceeding this threshold is considered abnormal during this usage phase.

[0079] Flight cycle count is the number of times an aircraft completes a full cycle from takeoff to landing. It is a key metric for measuring the usage intensity and lifespan of an aircraft and its components. One flight cycle typically corresponds to one takeoff and landing cycle, including the phases of takeoff, climb, cruise, descent, and landing.

[0080] Intermediate information may include the target time-series information for outlier filtering and the sensor information for outlier filtering.

[0081] Data completion refers to the process of repairing or reconstructing missing, discontinuous, or unreasonable data based on intermediate information after outlier removal, resulting in a complete, continuous, and high-quality dataset. Data completion addresses common issues in the original data, such as missing values, abrupt changes, and inconsistencies, providing reliable input for subsequent feature extraction.

[0082] Specifically, after obtaining the abnormal curves of each target flight component, the processing server can filter out abnormal values ​​in the target parameter values ​​and sensor parameter values ​​based on the relationship between the target parameter values ​​in the target time series information and the sensor parameter values ​​in the sensor information and the abnormal parameter thresholds corresponding to the number of flight cycles, thereby obtaining the intermediate information of the target flight component.

[0083] For example, the processing server can use the IQR algorithm and the Isolation Forest algorithm, combined with the abnormal parameter threshold corresponding to the number of flight cycles, to identify and remove abnormal values ​​(such as engine vibration jump greater than 20g) in the target parameter value and sensor parameter value, and obtain intermediate information of the target flight component (the proportion of abnormal values ​​in the intermediate information is less than 0.8%).

[0084] The IQR algorithm, or Interquartile Range Method, is a non-parametric outlier detection method based on the statistical distribution of data. It uses the first quartile (Q1, the 25th percentile), the third quartile (Q3, the 75th percentile), and the interquartile range (IQR = Q3 - Q1) to define the "normal" range of the data. Data points outside this range are considered outliers.

[0085] The Isolation Forest Algorithm (IFA) is an unsupervised anomaly detection algorithm based on ensemble learning. Its core idea is that anomalies are scarce and significantly different from normal points; therefore, by randomly partitioning the feature space, anomalies can be isolated with fewer random segmentations. Normal points require more segmentations to be isolated.

[0086] Specifically, after obtaining the intermediate data, the processing server can perform data completion processing on the intermediate data to obtain the completion information of the target flight component.

[0087] For example, after obtaining intermediate data, the processing server can perform data completion processing on the intermediate data based on the AR model to obtain the completed information of the target flight component. The AR model is used to complete the data in the intermediate information to obtain the completed information (parameters may be missing due to transmission interruption), and the data integrity of the completed information is greater than or equal to 99.9%.

[0088] Among them, the AR model, or Autoregressive Model, is a time series forecasting model that uses a linear combination of the same variable at several past time points to predict the current value. The AR(p) model means using the values ​​of the previous p time points to predict the current value.

[0089] Specifically, after obtaining the complete information, the processing server can extract features from the complete information to obtain the basic features of the target flight component.

[0090] For example, the processing server can extract features from the completed information using the feature extraction module of the life prediction model to obtain the basic features of the target flight component.

[0091] In this embodiment, firstly, by acquiring anomaly parameter thresholds (anomaly curves) that change with the number of flight cycles, dynamic thresholds can be applied to the component's lifespan, avoiding misjudgments of anomalies in the target flight component and improving the accuracy of the parameter values ​​corresponding to the target flight component. Secondly, for data loss caused by sensor failure, transmission interruption, or recording errors, data completion processing is performed on intermediate information to obtain completed information, ensuring data integrity and providing comprehensive data reference for subsequent data processing. Thirdly, feature extraction is performed on the completed information to characterize the information, facilitating direct feature processing by the subsequent model.

[0092] S203. Based on the basic features, maintenance business features, and graph embedding features, feature fusion is performed to obtain the fused features.

[0093] After obtaining the basic features, maintenance business features, and graph embedding features, feature fusion can be performed based on the basic features, maintenance business features, and graph embedding features to obtain fused features.

[0094] In some embodiments, the fusion feature can be a unified, complementary, and enhanced feature representation formed by organically combining the basic features (physical state), maintenance business features (historical operation and maintenance), and graph embedding features (domain knowledge) of the target flight component through multimodal fusion technology.

[0095] For example, basic features (1024 dimensions) can be combined with business features (512 dimensions) and graph embedding features (256 dimensions) for feature fusion processing, outputting a fused feature of 1792 dimensions. The fusion loss is less than or equal to 0.008, which improves feature utilization by 60% compared to traditional splicing methods. Feature fusion is not just simple splicing, but achieves information complementarity, conflict resolution, and representation enhancement through an intelligent fusion mechanism, forming a comprehensive, in-depth, and interpretable description of the component's health status.

[0096] In this embodiment, firstly, the triple verification of physical signals, service records, and expert knowledge improves information complementarity. Secondly, single-mode anomalies can be corrected through other modes, enhancing noise immunity and resulting in a more accurate fused feature vector, which helps improve the accuracy of subsequent maintenance schemes.

[0097] Before performing multi-feature fusion, it is necessary to consider the feature weights corresponding to each feature. Specifically, in a possible implementation, as described in S203 above, feature fusion is performed based on basic features, maintenance business features, and graph embedding features to obtain fused features. This includes obtaining the feature weights of the maintenance business features and graph embedding features, as well as obtaining the initial weights of the basic features. If the basic features are trend-changing features or collaborative features, the initial weights of the basic features are increased to obtain the feature weights of the basic features; if the basic features are not trend-changing features or collaborative features, the initial weights are used as the feature weights of the basic features. Based on the feature weights, the basic features, maintenance business features, and graph embedding features are mapped to the feature space. Through a three-dimensional encoding configured for the feature space, including time, decay degree, and semantics, feature fusion is performed on the basic features, maintenance business features, and graph embedding features within the feature space to obtain fused features.

[0098] The feature weight of maintenance business characteristics can be a numerical value that quantifies the relative importance of maintenance business characteristics in the current integration process. This weight can be automatically adjusted based on the quality, timeliness, and relevance of maintenance data, reflecting the contribution of historical maintenance information to the current health status assessment.

[0099] The feature weights of knowledge graph embedding features can be used to measure the credibility and applicability of knowledge graph features in current decision-making. These weights are dynamically determined based on the authority, timeliness, and matching accuracy of the knowledge source.

[0100] The initial weights are default weights before considering their specific characteristics (trend changes, synergies). These weights are based on the general importance preset of features, reflecting the fundamental role of data (target time-series information and sensor information) in health assessment.

[0101] Trend change characteristics characterize the fundamental features of how parameters change systematically and directionally over time or usage cycles. These characteristics reflect the long-term degradation trend of component performance and have significant predictive value.

[0102] Synergistic features characterize the fundamental relationships between multiple parameters. These features reflect the coupling effects within the system and can identify anomalous patterns that cannot be detected by a single parameter.

[0103] The initial weights of basic features are dynamically adjusted based on the feature type. Trend-changing features and collaborative features receive increased weights, reflecting their higher value in lifetime prediction.

[0104] The feature space can be a unified multidimensional vector space into which all features (base, maintenance, and graph) are projected and mapped. This feature space is designed to preserve the semantic information of the original features, facilitate the comparison and fusion of features from different modalities, and support subsequent 3D encoding and fusion operations.

[0105] Three-dimensional encoding, which incorporates temporal, decay-level, and semantic elements, can be a structured encoding scheme that assigns a three-dimensional contextual label to each point in the feature space, forming a triplet to achieve spatiotemporal semantic alignment of features. Three-dimensional encoding includes temporal encoding, decay-level encoding, and semantic encoding.

[0106] Since trend change features and co-occurrence features play a significant role in subsequent predictions, their weights can be increased. Specifically, after obtaining the initial weights of the basic features, if the basic features are trend change features or co-occurrence features, the processing server can increase the initial weights of the basic features to obtain their feature weights. If the basic features are neither trend change features (vibration values, exhaust temperature, for example, an increase of 5°C in exhaust temperature per 100 cycles indicates accelerated decay) nor co-occurrence features (landing load and buffer compression, load exceeding a threshold accompanied by a decrease in compression indicates buffer failure), the processing server can use the initial weights as the feature weights of the basic features.

[0107] To improve the prediction accuracy of subsequent models, feature fusion can be performed on multiple features beforehand. Specifically, after obtaining the weights of each feature and mapping the basic features, maintenance business features, and graph embedding features to the feature space, the processing server can perform feature fusion on the basic features, maintenance business features, and graph embedding features within the feature space using a three-dimensional encoding configured for the feature space that includes time, decay degree, and semantics, to obtain fused features.

[0108] For example, the processing server can perform feature fusion on basic features, maintenance business features and graph embedding features in the feature space by configuring three-dimensional encoding that includes time, decay degree and semantics. During the feature fusion process, the server can perceive semantic context and time features, and then perform cross-validation from time, decay and semantics to finally obtain the fused features.

[0109] In this embodiment, firstly, when the basic features are trend-changing features or collaborative features, the initial weight of the basic features is increased, thereby enhancing the weight of important features and preventing important weak signals from being overwhelmed by strong signals. Secondly, different features are mapped to a unified feature space, allowing for quantitative comparison between different features, which facilitates subsequent feature fusion processing. Thirdly, precise temporal context alignment is achieved. Based on 3D encoding, feature fusion is performed on basic features, maintenance business features, and graph embedding features within the feature space, realizing spatiotemporal semantic alignment of features and improving the accuracy of feature fusion.

[0110] S204: For each target flight component, the fusion features and reference cases associated with time-series data in the knowledge graph are input into the component life prediction model, and the remaining life prediction result of the target flight component is output.

[0111] In some embodiments, reference cases can be a complete record of historical maintenance instances similar to the current target flight component, retrieved from a knowledge graph. These cases include past actual component failures, repair processes, and final remaining life verification results, providing a basis for analogical reasoning for predicting the life of the current component.

[0112] The component life prediction model is an intelligent prediction system that integrates data-driven and physical mechanisms. The model takes the current component's integrated characteristics and similar reference cases as input, and outputs the remaining life prediction result (the probability distribution of the component's remaining service life from the current state to functional failure) through multi-model collaborative reasoning.

[0113] The remaining lifetime prediction result is a structured prediction information output by the component lifetime prediction model. It quantitatively describes the expected remaining service time of the target flight component from its current state to functional failure (or the need for preventive maintenance). The remaining lifetime prediction result can include multi-dimensional information such as probability distribution and confidence interval.

[0114] In one possible implementation, the remaining lifetime prediction results include the remaining lifetime value of the target flight component. The processing server can input the fused features and reference cases associated with time-series data in the knowledge graph into the component lifetime prediction model for each target flight component. Based on the PHM technology in the model, the remaining lifetime prediction results of the target flight component are output.

[0115] Among them, Prognostic and Health Management (PHM) is an integrated technology system and engineering methodology. It aims to assess the current health status of systems (such as engines and landing gear) by collecting and analyzing operational status data, predicting future performance degradation trends, and providing early warnings of failure risks. Its ultimate goal is to support the shift from planned / reactive maintenance to predictive / proactive maintenance, thereby improving safety, availability, and reducing total lifecycle costs.

[0116] PHM (Progressive Physical Modeling) technology is based on the theory of cumulative fatigue damage to predict Remaining Useful Life (RUL). Specifically, calculating RUL using cumulative fatigue damage records is a specific physical modeling method applied to mechanical structures subjected to cyclic loads (such as engine blades and landing gear). The core principle of this method is that fatigue failure of materials is the result of the gradual accumulation of micro-damage caused by alternating stress. When the accumulated damage reaches a critical threshold, cracks initiate or propagate to fracture.

[0117] In one possible implementation, the processing server can search the knowledge graph to identify three reference cases (Top-3 similar cases) associated with the time-series data. These three reference cases can include cases associated with the component model, cases associated with the degradation trend, and cases associated with the flight cycle. By searching the knowledge graph for the Top-3 similar cases associated with the time-series data, and taking into full account past actual component failures, maintenance processes, and final remaining life verification results, the server can provide a basis for analogical reasoning for the current component's life prediction.

[0118] In one possible implementation, the method further includes: acquiring multiple preset cases from a knowledge graph and determining the parameter values ​​and component model corresponding to the target flight component; the parameter values ​​include target parameter values ​​or sensor parameter values. Based on the parameter values, component model, and ternary relationships corresponding to the target flight component, the target anomaly type of the target flight component is determined. A search is performed among the preset cases, and preset cases that match both the component model and the target anomaly type are used as reference cases associated with the recorded data.

[0119] The preset cases can be structured fault-maintenance experience units that are pre-built and stored in a knowledge graph. Each case records a complete fault event that actually occurred in history, including the initial state, evolution process, handling measures, and final result. After standardization, these are formed into knowledge units that can be retrieved and reasoned about.

[0120] The parameter values ​​corresponding to the target flight component can be a set of real-time measured and calculated parameters directly related to the target flight component, extracted from the current aircraft's time-series data. These parameter values ​​reflect the component's current operating status and are key matching criteria for identifying anomalies and retrieving similar cases. For example, parameter values ​​could be engine exhaust temperature, engine rotor speed, landing gear buffer travel, landing load, etc.

[0121] The component model number is a unique, standardized identifier for a target flight component, precisely describing its design specifications, material properties, performance parameters, and maintenance requirements. The model number contains key information such as manufacturer, product line, and version number, and forms the basis for entity matching in the knowledge graph.

[0122] A ternary relationship is a relationship between three entities (flight component model, component parameter threshold, and component anomaly type).

[0123] Specifically, the processing server can determine the target anomaly type of the target flight component based on the parameter values, component model, and ternary relationship corresponding to the target flight component.

[0124] For example, after acquiring multiple preset cases from the knowledge graph and determining the parameter values ​​and component models corresponding to the target flight component, the processing server can select the ternary relation corresponding to the flight component model with the same model from among multiple ternary relations as the target ternary relation. The processing server can then determine the target anomaly type of the target flight component based on the component parameter thresholds and the magnitude relationships of the parameter values ​​in the target ternary relation.

[0125] Specifically, after obtaining the target anomaly type of the target flight component, the processing server can search for each preset case in the knowledge graph and use the preset case that matches the component model and the target anomaly type as a reference case associated with the recorded data.

[0126] In this embodiment, firstly, multiple preset cases are obtained from the knowledge graph, transforming tacit knowledge into an explicit knowledge base. Since the cases are preset and have a unified format, they facilitate subsequent retrieval and comparison. Secondly, retrieval based on the knowledge graph improves retrieval efficiency significantly compared to manually reviewing records. Thirdly, reference cases obtained through automatic reasoning using ternary relations in the knowledge graph enhance the efficiency and accuracy of obtaining reference cases.

[0127] S205, based on the remaining life prediction results of each target flight component, generates maintenance strategy information for the aircraft, which is used to guide the target flight components in maintenance.

[0128] In some embodiments, maintenance strategy information can be a structured, executable, and optimized maintenance action plan generated based on the remaining life prediction results of each target flight component. It is a multi-dimensional decision output that comprehensively considers multiple constraints such as safety, economy, and operational feasibility, providing a complete solution for each target flight component from its "current state" to its "optimal maintenance timing and method." For example, maintenance strategy information includes the maintenance window period, maintenance content, and required tool / spare part list for each target flight component.

[0129] In one possible implementation, the maintenance strategy information includes the maintenance window, maintenance content, and required tool / spare part list for each target flight component. The processing server can generate the maintenance window (e.g., completing borescope inspection within 10 flight cycles), maintenance content (e.g., replacing high-pressure compressor blades), and required tool / spare part list for each target flight component based on its remaining lifespan prediction. The generated maintenance strategy information, including the maintenance window, maintenance content, and required tool / spare part list, avoids over-maintenance of the aircraft, improves maintenance efficiency, and saves maintenance costs.

[0130] In one possible embodiment, S205, based on the remaining life prediction results of each target flight component, maintenance strategy information for the aircraft is generated. Specifically, this includes: obtaining initial maintenance sub-schemes for each target flight component and a maintenance planning model configured for the aircraft. The initial maintenance sub-schemes and the remaining life prediction results of each target flight component are input into the maintenance planning model, which outputs component maintenance sub-schemes for each target flight component. The maintenance planning model is trained with the goal of reducing maintenance costs, reducing downtime, and reducing safety risks. The maintenance sub-schemes for each component are integrated to obtain the aircraft's maintenance strategy information.

[0131] The initial maintenance sub-plan can be a basic maintenance action plan generated based on the remaining life prediction results of a single target flight component and standard maintenance procedures. It is a preliminary maintenance recommendation for a single component and has not yet considered coordination with other components, resource constraints, and overall optimization.

[0132] The maintenance planning model is a multi-objective optimization decision system. It takes initial maintenance sub-schemes and remaining life predictions for each component as input, comprehensively considers multiple constraints such as safety, cost, and time, and outputs the globally optimal component maintenance sub-scheme. This model is trained using a maintenance planning model with the training objectives of reducing maintenance costs, reducing downtime, and lowering safety risks.

[0133] A component maintenance sub-scheme can be a feasible and optimized maintenance execution plan for a single target flight component, after optimization and adjustment based on the maintenance planning model. It incorporates global optimization considerations on the basis of the initial plan and is a detailed work plan that can be directly executed.

[0134] Specifically, after obtaining the remaining life prediction results for each target flight component, the processing server can obtain the initial maintenance sub-scheme for each target flight component, as well as the maintenance planning model for the aircraft configuration.

[0135] For example, the processing server can obtain flight schedules. Among the initial maintenance sub-schemes, the initial maintenance sub-scheme that does not conflict with the flight schedule is selected as the target sub-scheme. The processing server can input the target sub-scheme, as well as the remaining life prediction results of the target flight component corresponding to the target sub-scheme, into the maintenance planning model, and output the component maintenance sub-scheme for each target flight component.

[0136] Specifically, after obtaining the component maintenance sub-plans, the processing server can integrate the component maintenance sub-plans according to their execution order to obtain the aircraft's maintenance strategy information.

[0137] In this embodiment, firstly, an initial maintenance sub-scheme is obtained, eliminating the need to generate a new scheme from scratch. This saves time compared to existing schemes, thus improving the efficiency of generating maintenance strategies. Secondly, the maintenance planning model is trained with the objectives of reducing maintenance costs, minimizing downtime, and mitigating safety risks. Simultaneously optimizing the three core objectives of cost, time, and safety improves the quality of remaining life prediction results. Thirdly, a corresponding component maintenance sub-scheme is configured for each target flight component, achieving component-level maintenance and refined maintenance.

[0138] In one possible embodiment, the method further includes: periodically acquiring execution data of maintenance strategy information within a preset time period. The execution data includes the lifespan difference between the actual and predicted lifespan of the target flight component, and the recurrence rate of anomalies in the target flight component after maintenance. Based on the execution data, the maintenance planning model, the component lifespan prediction model, and the knowledge graph are updated.

[0139] Among them, the execution data can be a quantitative record of the actual execution results of the maintenance strategy collected within a preset period (such as monthly or quarterly).

[0140] The actual lifespan of the target flight component can be a real lifespan value verified through actual operational data.

[0141] The predicted lifespan of a target flight component can be the expected remaining service life of the component from its current state to functional failure, output by a component lifespan prediction model. The recurrence rate of anomalies in target flight components refers to the frequency with which the same or similar anomaly patterns reappear during operation after maintenance.

[0142] For example, the processing server can collect execution data of maintenance strategy information (such as the deviation between the actual and predicted lifespan of components, and the failure recurrence rate after maintenance) every 10 days based on the observation-adjustment-decision-action cycle. The processing server can update model parameters and knowledge graphs based on the execution data, improving the monthly average remaining lifespan prediction accuracy by ≥3.2%.

[0143] In this embodiment, firstly, data is acquired periodically, and model parameters and knowledge graphs are updated based on the acquired execution data, forming a complete data loop and effectively improving the quality of prediction results. Secondly, based on the lifespan difference between the actual and predicted lifespan of the target flight component and the abnormal recurrence rate of the target flight component after maintenance, the model is continuously updated to avoid gradual performance degradation after model training and ensure that the model continuously outputs high-quality prediction results.

[0144] In one embodiment, such as Figure 3 As shown, a condition-based maintenance system is provided, which is used to perform the above method steps. Specifically, the condition-based maintenance system includes: a hardware perception layer, a data fusion layer, a model inference layer, and an application service layer.

[0145] The hardware perception layer includes a multi-source health data acquisition unit, edge processing nodes, and a cloud storage cluster.

[0146] Specifically, the multi-source health data acquisition unit is used to acquire recorded time-series data from the fast access recorder in the aircraft.

[0147] For example, the multi-source health data acquisition unit supports the ARINC 429 / 717 dual-interface standard, is compatible with 30+ aircraft models, and collects 186 health parameters of core QAR components, covering: Engine system: high-pressure compressor vibration value (64Hz), EGT (32Hz), fuel flow (16Hz), N1 / N2 speed fluctuation (32Hz); Landing gear system: landing vertical load (64Hz), buffer compression (16Hz), wheel axle temperature (8Hz), brake wear (1Hz); Avionics system: bus voltage fluctuation (16Hz), navigation signal signal-to-noise ratio (8Hz), FMS command response delay (8Hz); It also integrates component-specific sensors (such as engine oil sensors and landing gear strain sensors) to acquire microscopic health data such as oil contamination and component strain values, with a sampling frequency of 1Hz-8Hz.

[0148] For example, the edge processing node is equipped with an edge GPU to achieve real-time filtering of QAR data (using Kalman filtering to suppress noise) and primary anomaly detection (such as triggering an early warning when the vibration value exceeds the threshold), and locally caches 72 hours of data to ensure data continuity when the network is interrupted.

[0149] Specifically, edge processing nodes can use the health parameters of core components, which are filtered in real time and detected for preliminary anomalies, as time-series data to be recorded in the fast access recorder of the aircraft.

[0150] For example, the cloud storage cluster consists of 10 GPU servers, configured with 15PB of Ceph distributed storage, and equipped with the Neo4j knowledge graph database and model training platform, providing data storage and model iteration services. The cloud storage cluster can be used to store time-series data.

[0151] In one possible implementation, the hardware deployment of the hardware awareness layer includes: 1. Multi-source health data acquisition unit: The QAR adapter module uses an acquisition card compliant with the DO-178C Level B standard, supports ARINC 429 / 717 dual interfaces, and is configured with 64 synchronous acquisition channels. The sampling frequency for key parameters such as engine vibration and landing gear load is set to 64Hz. The oil sensor uses spectral analysis technology, and the oil contamination data is updated once per flight cycle.

[0152] 2. Edge processing node: Deploy edge GPUs, configure 32GB of memory and 2TB of storage to achieve basic anomaly detection (such as triggering an alert when the engine vibration value is >18g).

[0153] 3. Cloud storage cluster: 10 GPU servers are managed using a Kubernetes cluster with a network bandwidth of ≥100Gbps. The Neo4j knowledge graph database supports 100,000+ queries per second.

[0154] The data fusion layer includes a data access module, a preprocessing module, a knowledge structuring module, and a fusion module.

[0155] The data access module is used to acquire recorded time-series data, maintenance business data, component attribute data, and knowledge rule data. It then constructs a knowledge graph based on the component attribute data and knowledge rule data.

[0156] Maintenance operational data includes repair work orders, component replacement records, fault reports, fluid analysis reports, and maintenance personnel qualifications. Component attribute data includes component model, design life (flight cycles / hour), manufacturing date, historical degradation curves, and material fatigue parameters. Knowledge rule data includes PHM theoretical formulas (such as remaining service life prediction models), maintenance standards (such as engine borescope inspection thresholds), and aircraft maintenance manuals (AMM / CMM).

[0157] The preprocessing module is used to remove outliers and fill in missing values ​​in the target time series information and sensor information to obtain the completed information.

[0158] The knowledge structuring module is used to build a knowledge graph of component decay, which contains 3200+ entities (flight component models, component parameter thresholds, component anomaly types, component maintenance actions, etc.) and 2800+ relationships (binary and ternary relationships between flight component models, component parameter thresholds, component anomaly types, and component maintenance actions, etc.).

[0159] For example, physical entities include engine high-pressure compressor A (flight component model), fatigue cracks (component anomaly type), borehole inspection (maintenance action), and vibration value of 15g (component parameter threshold).

[0160] The relationships include a ternary relationship between flight component model, component parameter threshold, and component anomaly type (high pressure compressor vibration value > 15g → fatigue crack), and a binary relationship between component anomaly type and component maintenance action (fatigue crack depth 5mm → borehole inspection + crack repair).

[0161] In one possible implementation, the data processing flow of the data fusion layer includes: 1. Data Access: The data access module receives real-time QAR data via the MQTT protocol, obtains work order data from the maintenance management system via the HTTP protocol, receives spare parts inventory information via the WebSocket protocol, and achieves data alignment based on flight cycle numbers.

[0162] 2. Preprocessing execution: The preprocessing module is used to remove extreme outliers (usually a small percentage, such as 0.7% or 0.5%) from the QAR data based on the IQR algorithm, and then completes the missing points through an adaptive AR model to obtain the complete information of the target flight component.

[0163] 3. Knowledge Graph Construction: The knowledge structuring module uses the TransE algorithm to train 256-dimensional entity embedding vectors, and a knowledge graph can be constructed based on these entity embedding vectors. After review by aircraft maintenance experts, the accuracy of the constructed knowledge graph exceeds a preset percentage (such as 99.4% or 98%).

[0164] 4. Cross-modal fusion: The fusion module trains a 12-layer cross-modal attention network using the AdamW optimizer. After 400 iterations, the fusion loss converges to a preset range (for example, the preset range can be a loss less than or equal to 0.0075).

[0165] The model inference layer includes a health feature extraction module, a remaining life prediction submodule, a condition-based maintenance planning submodule, and a closed-loop optimization submodule.

[0166] The health feature extraction module is used to extract features from the target time-series information and sensor information to obtain the basic features of the target flight components; to extract features from the maintenance work order information to obtain the maintenance business features of the target flight components; and to extract features from the binary and ternary relations to obtain the map embedding features of the target flight components.

[0167] The health feature extraction module is also used to increase the initial weight of the basic feature when the basic feature is a trend change feature or a collaborative feature, so as to obtain the feature weight of the basic feature; when the basic feature is not a trend change feature or a collaborative feature, the initial weight is used as the feature weight of the basic feature.

[0168] The Remaining Life Prediction submodule outputs the remaining life prediction results for each target flight component based on the Retrieval Enhanced Generation (RAG) strategy. The RAG strategy aims to improve the accuracy, relevance, and reliability of the responses generated by the large language model by optimizing the retrieved context.

[0169] The condition-based maintenance planning submodule generates maintenance strategy information by combining flight scheduling plans and remaining life prediction results with the objective function of "lowest maintenance cost, shortest downtime, and lowest safety risk". The maintenance strategy information includes maintenance windows (e.g., "it is recommended to complete the borescope inspection within 10 flight cycles"), maintenance content (e.g., "replace the high-pressure compressor blades"), a list of required tools / spare parts, and a compliance rate greater than or equal to a preset percentage, such as 98% or 96%.

[0170] The closed-loop optimization submodule is used to periodically acquire the execution data of maintenance strategy information within a preset time period. The execution data includes the life difference between the actual life and the predicted life of the target flight component, and the recurrence rate of abnormalities of the target flight component after maintenance. Based on the execution data, the maintenance planning model, the component life prediction model, and the knowledge graph are updated.

[0171] In one possible implementation, the model training and inference process in the model inference layer includes: 1. Health Feature Extraction Sub-model: Training data: Data from 50,000 flights, including 12 types of decline pattern labels.

[0172] Training parameters: 500 iterations, learning rate 1e-4, reconstruction loss converges to a preset value (e.g., 0.012, 0.01, etc.).

[0173] 2. Remaining lifetime prediction sub-model: Base model: Based on model fine-tuning, LoRA technology is used to reduce training costs.

[0174] For example, for the high-pressure compressor of an engine (1200 flight cycles, vibration value 14.5g), the top-3 similar cases are searched, and the remaining life is output as 115±4 cycles. The error (prediction error) between the output remaining life and the actual remaining life is 3.5%.

[0175] 3. Closed-loop optimization sub-model: For example, the closed-loop optimization sub-model: based on 8 months of maintenance feedback data from a certain airline (execution data corresponding to the maintenance strategy information output by the condition-based maintenance planning sub-module, including 3,000 actual component life records), it is iterated every 10 days. Tests show that after 8 months, the prediction error has decreased from the initial 5.8% to 3.7%.

[0176] The application service layer includes a health monitoring terminal, a maintenance planning platform, a spare parts management module, and an effect evaluation module.

[0177] The health monitoring terminal is deployed in the aircraft maintenance monitoring center to display the health status of core components (health level 1-10), remaining life prediction curve, and decline trend warning in real time.

[0178] The maintenance planning platform is used to automatically generate maintenance work orders, supports manual fine-tuning (such as adjusting maintenance windows), and pushes them synchronously to the maintenance personnel's terminals.

[0179] The spare parts management module is used to generate spare parts demand plans in advance based on the tool / spare parts list and maintenance window period in the maintenance strategy information (e.g., "2 high-pressure compressor blades need to be stockpiled after 10 cycles"). Some test results show that generating spare parts demand plans 15 days in advance can improve spare parts turnover rate by 40%.

[0180] The effectiveness evaluation module is used to calculate the component life extension rate and cost savings after maintenance, generate maintenance effectiveness reports, and support model iteration.

[0181] In this application embodiment, firstly, the prediction accuracy is significantly improved: test results show that using the solution provided in this application embodiment, the prediction error of the remaining life of core components is ≤4%, which is a significant improvement in accuracy compared to the traditional empirical formula (error 18%), and can identify latent degradation 30-50 flight cycles in advance. Secondly, maintenance efficiency is optimized: the efficiency of maintenance plan execution is improved, aircraft downtime is shortened, and the rate of unplanned downtime is reduced. Thirdly, costs are significantly reduced: the over-maintenance rate and the annual maintenance cost per aircraft are reduced. Fourthly, safety risks are reduced: the failure recurrence rate of core components is greatly reduced, the early identification rate of latent degradation is improved, and the level of flight safety is significantly improved.

[0182] Figure 4 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. Figure 4 As shown, the data processing device includes: an acquisition module 401, a feature extraction module 402, a feature fusion module 403, a prediction module 404, and an information generation module 405.

[0183] The acquisition module 401 is used to acquire maintenance business data of the aircraft and a knowledge graph configured for the aircraft, and to acquire time-series data recorded from the fast access recorder in the aircraft. The maintenance business data includes maintenance work order information of the aircraft; entities in the knowledge graph include flight component models, component anomaly types, component maintenance actions, and component parameter thresholds; the relationships in the knowledge graph include ternary relationships between flight component models, component parameter thresholds, and component anomaly types, and binary relationships between component anomaly types and component maintenance actions; the time-series data includes target time-series information of multiple target flight components contained in multiple subsystems of the aircraft, and sensor information of dedicated sensors connected to the target flight components in the aircraft; the target time-series information includes component time-series information of the target flight components and system time-series information of the subsystems; and the component parameter thresholds are thresholds configured for parameter values ​​corresponding to the target flight components. The feature extraction module 402 is used to extract features from the target time-series information and the sensor information to obtain the basic features of the target flight component; to extract features from the maintenance work order information to obtain the maintenance business features of the target flight component; and to extract features from the binary relation and the ternary relation to obtain the map embedding features of the target flight component. Feature fusion module 403 is used to perform feature fusion based on the basic features, the maintenance business features, and the map embedding features to obtain fused features; The prediction module 404 is used to input the fused features and the reference cases in the knowledge graph associated with the time series data into the component life prediction model for each target flight component, and output the remaining life prediction result of the target flight component. The information generation module 405 is used to generate maintenance strategy information for the aircraft based on the remaining life prediction results of each of the target flight components. The maintenance strategy information is used to guide the target flight components to perform maintenance.

[0184] In other embodiments, the information generation module 405 is further configured to acquire the initial maintenance sub-schemes for each of the target flight components and the maintenance planning model configured for the aircraft; input the initial maintenance sub-schemes and the remaining life prediction results for each of the target flight components into the maintenance planning model, and output the component maintenance sub-schemes for each of the target flight components; wherein the maintenance planning model is trained with the training objectives of reducing maintenance costs, reducing downtime and reducing safety risks; and integrate the component maintenance sub-schemes to obtain the maintenance strategy information of the aircraft.

[0185] In other embodiments, the data processing device may further include an update module for periodically acquiring execution data of the maintenance strategy information within a preset time period; wherein the execution data includes the lifespan difference between the actual lifespan and the predicted lifespan of the target flight component, and the abnormality recurrence rate of the target flight component after maintenance; based on the execution data, the maintenance planning model, the component lifespan prediction model, and the knowledge graph are updated.

[0186] In other embodiments, the feature extraction module 402 is further configured to acquire anomaly curves for each of the target flight components; the anomaly curves characterize the relationship between the anomaly parameter threshold corresponding to the target flight component and the number of flight cycles; based on the target parameter values ​​in the target time-series information and the sensor parameter values ​​in the sensor information, and their respective magnitudes relative to the anomaly parameter thresholds corresponding to the number of flight cycles, outliers in the target parameter values ​​and sensor parameter values ​​are filtered out to obtain intermediate information of the target flight component; the intermediate information includes target time-series information and sensor information from which outliers have been filtered out; data completion processing is performed on the intermediate information to obtain completed information of the target flight component; feature extraction is performed on the completed information to obtain the basic features of the target flight component.

[0187] In other embodiments, the acquisition module 401 is further configured to determine the subsystem to which the target flight component of the aircraft belongs and the dedicated sensor connected to the target flight component; wherein the subsystem includes an engine system, a landing gear system, or an avionics system, and the dedicated sensor includes an engine oil sensor or a landing gear strain sensor; and to acquire recorded time-series data from the subsystem matching the preset reading frequency range and the dedicated sensor matching the preset reading frequency range according to the preset reading frequency range in the fast access recorder of the aircraft.

[0188] In other embodiments, the data processing apparatus may further include a case generation module, used to acquire multiple preset cases in the knowledge graph and determine the parameter values ​​and component models corresponding to the target flight component; the parameter values ​​include target parameter values ​​or sensor parameter values; based on the parameter values ​​corresponding to the target flight component, the component model, and the ternary relationship, determine the target anomaly type of the target flight component; search among the preset cases, and use the preset cases that match the component model and the target anomaly type as the reference cases associated with the recorded data.

[0189] In other embodiments, the feature fusion module 403 is further configured to obtain the feature weights of the maintenance business feature and the graph embedding feature, and to obtain the initial weight of the basic feature; if the basic feature is a trend change feature or a collaborative feature, increase the initial weight of the basic feature to obtain the feature weight of the basic feature; if the basic feature is not a trend change feature or a collaborative feature, use the initial weight as the feature weight of the basic feature; based on each feature weight, map the basic feature, the maintenance business feature, and the graph embedding feature to a feature space; and perform feature fusion on the basic feature, the maintenance business feature, and the graph embedding feature in the feature space by configuring a three-dimensional encoding that includes time, decay degree, and semantics for the feature space to obtain a fused feature.

[0190] The data processing apparatus provided in this application embodiment can execute the methods shown in the above method embodiments. Its implementation principle and beneficial effects can be referred to the relevant descriptions in the method embodiments, and will not be repeated here.

[0191] Figure 5 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. Figure 5 As shown, the data processing device includes: a memory 501, a transceiver 502, and at least one processor 503.

[0192] The transceiver 502 is used to interact with other devices to send and receive data. For example, in this embodiment, the transceiver 502 can specifically be used to send and receive timing data, etc.

[0193] The memory 501 is used to store computer program code, which includes computer instructions. These computer instructions run in the data processing device described above to implement the method shown in the above method embodiments. For example, the memory may include high-speed random access memory (RAM), and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be a USB flash drive, portable hard drive, read-only memory, disk, or optical disc, etc.

[0194] Processor 503 can be a general-purpose processor, including a Central Processing Unit (CPU), a network processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. Processor 503 can also be other general-purpose processors. The general-purpose processor can be a microprocessor or any conventional processor.

[0195] The memory 501, transceiver 502, and processor 503 are communicatively connected. For example, the memory 501 and transceiver 502 can be connected to the processor 503 via a system bus and communicate with each other. The system bus can be a peripheral component interconnect (PCI) bus, an extended industry standard architecture (EISA) bus, an industry standard architecture (ISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the figure, but this does not mean that there is only one bus or one type of bus.

[0196] Optionally, the memory 501 can be either standalone or integrated with the processor 503. When the memory 501 is set up independently, it is connected to the processor 503 via a system bus.

[0197] This application also provides a chip for executing instructions, which is used to execute the data processing method described in the above embodiments.

[0198] This application also provides a computer-readable storage medium storing computer instructions. When these computer instructions are executed by a processor, they are used to implement the technical solution of the data processing method described in the above embodiments. Specifically, when the computer instructions are executed by a processor, the data processing device can execute the technical solution of the data processing method described in the above embodiments.

[0199] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and when the at least one processor executes the computer program, it can implement the technical solution of the data processing method in the above embodiments.

[0200] The aforementioned computer-readable storage media can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The computer-readable storage media can be any available medium accessible to a general-purpose or special-purpose computer.

[0201] An exemplary computer-readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the computer-readable storage medium can also be a component of the processor. The processor and the computer-readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the computer-readable storage medium can exist as discrete components in an electronic control unit or main control device; this application does not limit this.

[0202] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0203] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0204] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0205] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0206] It should be understood that the steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.

[0207] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A data processing method, characterized by, The method includes: Acquire maintenance business data of the aircraft and knowledge graph configured for the aircraft, and obtain time-series data recorded from the fast access recorder in the aircraft; The maintenance business data includes maintenance work order information of the aircraft; entities in the knowledge graph include flight component models, component anomaly types, component maintenance actions, and component parameter thresholds; the relationships in the knowledge graph include ternary relationships between flight component models, component parameter thresholds, and component anomaly types, and binary relationships between component anomaly types and component maintenance actions; the time-series data includes target time-series information of multiple target flight components contained in multiple subsystems of the aircraft, and sensor information of dedicated sensors connected to the target flight components in the aircraft; the target time-series information includes component time-series information of the target flight components and system time-series information of the subsystems; and the component parameter thresholds are thresholds configured for parameter values ​​corresponding to the target flight components. Feature extraction is performed on the target time-series information and the sensor information to obtain the basic features of the target flight component; feature extraction is performed on the maintenance work order information to obtain the maintenance business features of the target flight component; feature extraction is performed on the binary relation and the ternary relation to obtain the map embedding features of the target flight component. Based on the basic features, the maintenance business features, and the map embedding features, feature fusion is performed to obtain fused features; For each target flight component, the fused features and the reference cases associated with the time-series data in the knowledge graph are input into the component life prediction model, and the remaining life prediction result of the target flight component is output. Based on the remaining life prediction results of each of the target flight components, maintenance strategy information for the aircraft is generated, which is used to guide the maintenance of the target flight components.

2. The method according to claim 1, characterized in that, The process of generating maintenance strategy information for the aircraft based on the remaining life prediction results of each of the target flight components includes: Obtain the initial maintenance sub-scheme for each of the target flight components, and the maintenance planning model configured for the aircraft; The initial maintenance sub-scheme and the remaining life prediction results of each of the target flight components are input into the maintenance planning model, and the component maintenance sub-scheme for each of the target flight components is output; wherein, the maintenance planning model is trained with the training objectives of reducing maintenance costs, reducing downtime and reducing safety risks; The maintenance sub-schemes of each component are integrated to obtain the maintenance strategy information of the aircraft.

3. The method of claim 2, wherein, The method further includes: The execution data of the maintenance strategy information within a preset time period is periodically acquired; wherein, the execution data includes the life difference between the actual life and the predicted life of the target flight component, and the abnormality recurrence rate of the target flight component after maintenance; Based on the execution data, the maintenance planning model, the component life prediction model, and the knowledge graph are updated.

4. The method of claim 1, wherein, The step of extracting features from the target's temporal information and sensor information to obtain the basic features of the target's flight components includes: Obtain the anomaly curve for each of the target flight components; the anomaly curve is used to characterize the relationship between the anomaly parameter threshold corresponding to the target flight component and the number of flight cycles. Based on the target parameter values ​​in the target time series information and the sensor parameter values ​​in the sensor information, and their respective magnitudes relative to the abnormal parameter thresholds corresponding to the number of flight cycles, abnormal values ​​in the target parameter values ​​and sensor parameter values ​​are filtered out to obtain intermediate information of the target flight component; the intermediate information includes target time series information and sensor information for which abnormal values ​​have been filtered out. The intermediate information is processed to complete the data, thereby obtaining the complete information of the target flight component; Feature extraction is performed on the completed information to obtain the basic features of the target flight component.

5. The method of claim 1, wherein, The process of retrieving recorded time-series data from the fast access recorder in the aircraft includes: Identify the subsystem to which the target flight component of the aircraft belongs, and the dedicated sensors connected to the target flight component; The subsystems include engine systems, landing gear systems or avionics systems, and the dedicated sensors include engine oil sensors or landing gear strain sensors. According to the preset read frequency range in the fast access recorder in the aircraft, the recorded time-series data is acquired from the subsystem matching the preset read frequency range and the dedicated sensor matching the preset read frequency range.

6. The method of claim 1, wherein, The method further includes: Multiple preset cases are obtained from the knowledge graph, and the parameter values ​​and component models corresponding to the target flight component are determined; the parameter values ​​include target parameter values ​​or sensor parameter values; Based on the parameter values ​​corresponding to the target flight component, the component model, and the ternary relationship, the target anomaly type of the target flight component is determined; A search is conducted among the preset cases, and the preset cases that match the component model and the target anomaly type are used as the reference cases associated with the recorded data.

7. The method of claim 1, wherein, The step of fusing features based on the basic features, the maintenance business features, and the graph embedding features to obtain fused features includes: Obtain the feature weights of the maintenance business features and the graph embedding features, and obtain the initial weights of the basic features; If the basic feature is a trend change feature or a collaborative feature, increase the initial weight of the basic feature to obtain the feature weight of the basic feature; if the basic feature is not a trend change feature or a collaborative feature, use the initial weight as the feature weight of the basic feature. Based on the aforementioned feature weights, the basic features, the maintenance business features, and the graph embedding features are mapped to the feature space; By configuring a three-dimensional encoding that includes time, decay degree, and semantics for the feature space, feature fusion is performed on the basic features, maintenance business features, and graph embedding features within the feature space to obtain fused features.

8. A data processing apparatus, characterized by, The device includes: The acquisition module is used to acquire maintenance business data of the aircraft and knowledge graph configured for the aircraft, and to acquire time-series data recorded from the fast access recorder in the aircraft. The maintenance business data includes maintenance work order information of the aircraft; entities in the knowledge graph include flight component models, component anomaly types, component maintenance actions, and component parameter thresholds; the relationships in the knowledge graph include ternary relationships between flight component models, component parameter thresholds, and component anomaly types, and binary relationships between component anomaly types and component maintenance actions; the time-series data includes target time-series information of multiple target flight components contained in multiple subsystems of the aircraft, and sensor information of dedicated sensors connected to the target flight components in the aircraft; the target time-series information includes component time-series information of the target flight components and system time-series information of the subsystems; and the component parameter thresholds are thresholds configured for parameter values ​​corresponding to the target flight components. The feature extraction module is used to extract features from the target time-series information and the sensor information to obtain the basic features of the target flight component; to extract features from the maintenance work order information to obtain the maintenance business features of the target flight component; and to extract features from the binary relation and the ternary relation to obtain the map embedding features of the target flight component. The feature fusion module is used to perform feature fusion based on the basic features, the maintenance business features, and the map embedding features to obtain fused features; The prediction module is used to input the fused features and the reference cases in the knowledge graph associated with the time series data into the component life prediction model for each target flight component, and output the remaining life prediction result of the target flight component. The information generation module is used to generate maintenance strategy information for the aircraft based on the remaining life prediction results of each of the target flight components. The maintenance strategy information is used to guide the target flight components in maintenance.

9. A data processing device, characterized by include: A memory and at least one processor; the memory is communicatively connected to the processor; the memory is used to store computer program code, the computer program code including computer instructions; when the processor executes the computer instructions, the data processing device causes the data processing device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

Images