Improved saits algorithm-based health management system for navigation equipment

Abstract

The application discloses a navigation equipment health management system based on an improved SAITS algorithm and relates to the technical field of navigation equipment fault detection. The system comprises a multi-source data acquisition module, a data processing module, an interpolation and prediction module, an anomaly detection module, a root cause analysis module, a digital twin module, a warning module, an anomaly processing module and a model correction module. Multi-source data such as acquisition sensors, monitoring and scheduling logs are collected and cleaned and integrated into time series data. An improved SAITS algorithm that fuses STL decomposition, abnormal sample enhancement and a feedback mechanism is used to realize missing data interpolation and future operation data prediction. An adaptive anomaly detection algorithm is used to identify anomalies, a knowledge graph is used to locate fault roots, and a multi-dimensional digital twin model is used to simulate and verify anomaly confidence. An alarm and health assessment report is generated, and a disposal scheme and a maintenance plan are output. The application can improve the anomaly detection accuracy of navigation equipment, reduce the false alarm rate, realize active operation and maintenance and full-process closed-loop management, and is suitable for the health management and control of equipment in complex water transportation scenarios.

Description

Technical Field

[0001] This invention belongs to the field of general aviation equipment fault detection, specifically relating to a general aviation equipment health management system based on an improved SAITS algorithm. Background Technology

[0002] With the increasing number of navigation equipment in modern shipping, the detection and maintenance of these devices has become increasingly important. Navigation equipment, such as level gauges and locks, operates in complex and variable environments. Aging of the equipment itself, external environmental interference, and fluctuations in operating load can all lead to abnormal operating conditions. Existing methods are insufficient for effective management of equipment health, resulting in sudden equipment failures.

[0003] Traditional general aviation equipment health management methods typically rely on fixed thresholds or simple rule models. While simple, these methods have significant limitations when dealing with the vast number of different general aviation equipment and complex operating environments. Fixed thresholds struggle to adapt to dynamic changes in equipment status, leading to missed or false alarms during normal fluctuations. Furthermore, most existing general aviation equipment has built-in microcomputers that acquire various types of data, which traditional methods cannot fully utilize for accurate analysis. In addition, traditional methods only perform static monitoring of existing data and cannot provide reasonable early warnings about equipment health trends. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems by providing a health management system for navigation equipment based on an improved SAITS algorithm. This system utilizes edge computing nodes for real-time pre-screening of channel topology data and combines this with an environmental factor correction matrix to dynamically compensate sensor data, thereby improving the real-time performance, completeness, and resistance to environmental interference in data acquisition and ensuring the reliability of input data. It also improves the Self-Attention-based Imputation for TimeSeries (SAITS) model, using the improved SAITS algorithm to perform high-precision imputation of missing data and predict future equipment operating states, providing a high-quality data foundation for anomaly detection. Furthermore, it employs a dynamically adjusted anomaly detection algorithm combined with a fault mode knowledge graph to achieve accurate identification of current and future anomalies and rapid location of fault root causes. Finally, it constructs a multi-dimensional digital twin model of "ship-lock-channel," verifying the confidence level of anomaly results through simulation. By comparing physical entity data with the multi-dimensional digital twin model, it reduces the false alarm rate and improves the reliability of fault diagnosis results.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a general aviation equipment health management system based on an improved SAITS algorithm, comprising:

[0006] The multi-source data acquisition module collects multi-source data from the general aviation system, including sensor data, general aviation monitoring data, network monitoring data, and general aviation dispatching system log data. The data processing module cleans, integrates, and standardizes multi-source data, and stores it in time-series format; The interpolation and prediction module uses the improved SAITS algorithm to interpolate missing data in multi-source data, and at the same time predicts the operation data of general aviation equipment in future time periods. The anomaly detection module uses an adaptive anomaly detection algorithm to identify anomalies in the current and future operational data of general aviation equipment. The root cause analysis module is used to build a knowledge graph model and analyze the root causes of anomalies identified by the anomaly detection module. The digital twin module constructs a digital twin model of the general aviation system. It integrates the working data generated by the operation of the physical entity with the simulation data of the multi-dimensional digital twin model to form twin data. The anomaly types and parameters identified by the anomaly detection module are injected into the digital twin model for dynamic simulation to verify the confidence level of the equipment anomalies identified by the anomaly detection module. The early warning module generates real-time alarm information based on the anomaly detection results and generates a health assessment report for possible future anomalies. The health assessment report includes the current operating status of the equipment, fault type prediction, remaining life assessment, and potential faults that may be induced. The anomaly handling module generates anomaly response strategies for anomalies identified by the anomaly detection module; and generates fault handling plans and maintenance plans based on fault reasoning models for faults identified by the digital twin module.

[0007] The model correction module optimizes the adaptive anomaly detection algorithm model based on the feedback from the duty officer on the detection results. The optimization process includes adjusting the detection threshold, updating the anomaly sample library, and retraining the adaptive anomaly detection algorithm.

[0008] Preferably, the multi-source data acquisition module uses edge computing nodes to pre-screen the channel topology sensing data and perform spatial correlation verification on the cluster sensor data; it also establishes an environmental factor correction matrix to dynamically compensate for sensor data affected by environmental parameters.

[0009] In the interpolation prediction module, the improvements of the improved SAITS algorithm over the existing SAITS algorithm include: 1) STL decomposition is used for seasonal data. The operation of the navigation system is closely related to the water level, and the shipping volume is related to the flood season and the dry season. Therefore, the operation of some equipment shows obvious seasonal trends. By using STL decomposition technology, the original navigation system time series data is decomposed into trend, seasonality and residual, which are used as inputs to the SAITS model. This can more effectively impute missing values ​​in seasonal data and noisy data. 2) Abnormal sample augmentation: SAITS models lose accuracy when faced with outliers. By generating abnormal samples through data generation techniques and increasing the weight of abnormal samples when training the SAITS model, the learning ability of abnormal data can be effectively strengthened, and the accuracy of prediction of abnormal data can be enhanced. 3) Introducing a feedback mechanism: The SAITS model has a large error in predicting data, which can lead to anomaly detection errors. Introducing a feedback mechanism can improve the accuracy and generalization ability of the SAITS model. Based on the error type, such as false positives and false negatives, the strategy is adaptively adjusted and retrained, including adjusting the masking ratio, adding attention layers, and retraining using abnormal samples.

[0010] In the anomaly detection module, a deep Q-network (DQN) is used as one of the adaptive anomaly detection algorithms, specifically including: (a) Define the state space S and the action space A: The state space is the multi-dimensional historical sensor data corresponding to the device, and the action space set A is defined as: judging the current state as "normal", judging the current state as "abnormal", and adjusting the detection threshold. (ii) Setting the reward function: The reward function includes a reward for correctly detecting a device malfunction. Penalty for incorrectly detecting device malfunction Correctly judging the normal operation of the equipment is a reward And penalties for missing equipment malfunctions. The final reward function is a weighted sum of all rewards: ; This is the final reward result; All are weighting coefficients; The action value function, Q-value function, is composed of both current and future benefits, and is expressed as follows: The objective Q-value function represents the impact of the current action on the future. ; In the formula, Indicates the discount factor; These represent the rewards at time t, t+1, t+2, ..., t+n, respectively. These represent the states at time t and time t+1, respectively. These represent the actions at time t and time t+1, respectively. (iii) Iterative training; Execute actions at each time step Receive rewards and enter the next state. Then sample tuples Return the samples to the experience replay pool; during iterative training, randomly select a batch of samples from the experience replay pool for training; calculate the current state based on the main network. Values ​​and Targets loss function Current state Values ​​and Targets between the mean squared deviations ; The loss function is minimized and iterated until convergence is achieved. (iv) Anomaly detection; Using the future time series predicted by SAITS interpolation as input to DQN, the state of each time node in the time series is obtained, which is the detection and prediction of anomalies.

[0011] In the root cause analysis module, building a knowledge graph of equipment failure modes includes: Data collection and processing: Collect logs, call chain data and operation manuals generated by the general aviation system, filter the collected data, and process the filtered data into sentences before using it as input for the large language model; Knowledge extraction: Knowledge extraction is performed based on a large language model, and knowledge is divided into entity classes, relation classes, and attribute classes for extraction. Knowledge fusion and ontology construction: The extracted knowledge is fused, and the information of the same entity in multi-source data is integrated. For example, the same device from different data sources is merged into one entity, and the relationship is fused into repeated or contradictory relationship descriptions to form a consistent relationship expression. Graph storage and application: Use the neo4j database to store entity knowledge and connect entity relationships.

[0012] The digital twin module is configured to perform the following operations: Constructing a multidimensional digital twin model: Based on the "geometry-physics-behavior-rules" multidimensional model framework, a multidimensional joint model is established, including a geometric model of the ship-lock-channel, a hydrodynamic physical model, a lock operation behavior model, and a navigation rule model, in order to characterize the geometric features, physical properties, behavioral coupling relationships, and behavioral criteria of the physical entity; Generate twin data: Twin data is formed by fusing the working data generated by the operation of physical entities with the simulation data of virtual models. The twin data is used to drive the virtual model to control, optimize and predict the working state of physical entities. Fault simulation verification: The identified anomaly types and parameters are injected into the digital twin model, and dynamic simulation is performed using the digital twin simulation engine to verify the confidence level of the equipment anomalies identified by the adaptive anomaly detection algorithm.

[0013] Secondly, this invention provides a health management method for general aviation equipment based on an improved SAITS algorithm, comprising the following steps: S1: Acquire operational data of general aviation equipment; pre-screen channel topology sensing data using edge computing nodes and perform spatial correlation verification on cluster sensor data; establish an environmental factor correction matrix to dynamically compensate sensor data affected by environmental parameters; clean, integrate, and store operational data of general aviation equipment and synchronize it to the digital twin platform. S2: An improved SAITS algorithm is used to impute missing data and predict future operational data of general aviation equipment; S3: Based on the predicted data obtained in step S2, an adaptive anomaly detection algorithm is used to identify the current and future abnormal states of general aviation equipment; based on the identified equipment anomalies, an equipment failure mode knowledge graph is constructed. S4: Construct a digital twin model of the general aviation system. For the equipment anomalies identified in step S3, use the digital twin platform to simulate the operation of the general aviation system, verify the confidence level of the equipment anomalies identified in step S3, and generate a health assessment report. If the confidence level is not less than the confidence level threshold, then combine the equipment failure mode knowledge graph to generate anomaly response strategies. S5: Based on actual feedback from general aviation equipment testing, incremental learning or retraining of the adaptive anomaly detection algorithm is performed to improve detection accuracy.

[0014] S6: Generate a maintenance plan based on the aforementioned anomaly response strategy or fault handling scheme, and use the digital twin platform to simulate the impact of the maintenance plan on traffic volume and efficiency.

[0015] Preferably, step S4 further includes calculating the deviation between the predicted data and the simulation results of the general aviation equipment. If the deviation exceeds the deviation threshold, the domain expert will further diagnose and confirm the fault, and, in conjunction with the expert's opinion, adopt a fault reasoning model to form a fault handling plan.

[0016] Compared with the prior art, the beneficial effects of the present invention include: 1) This invention deeply integrates edge computing, the improved SAITS algorithm, digital twins, knowledge graphs, and incremental learning to form an integrated closed-loop system of "perception-analysis-simulation-optimization." Through environmental compensation, algorithm self-optimization, and expert collaboration mechanisms, it achieves strong adaptability to complex navigation operation conditions. Employing a dynamically adjusted anomaly detection algorithm combined with a fault mode knowledge graph, it achieves accurate identification of abnormal states (current and future) and rapid location of fault root causes in the navigation system, improving detection sensitivity and intelligence. By constructing a multi-dimensional digital twin model of "ship-lock-channel," it uses simulation to verify the confidence level of anomaly results, and by comparing physical entity data with virtual simulation, it reduces the false alarm rate and improves the reliability of diagnostic results. This invention has significant advantages in improving the management efficiency of navigation equipment, reducing failure risks, and optimizing maintenance costs, and is particularly suitable for water transport scenarios with stringent reliability and real-time requirements.

[0017] 2) This invention utilizes edge computing nodes to perform real-time pre-screening of waterway topology data, and combines environmental factor correction matrix to dynamically compensate sensor data, which significantly improves the real-time performance, integrity and anti-environmental interference capability of data acquisition, and ensures the reliability of input data.

[0018] 3) This invention uses the improved SAITS algorithm to perform high-precision interpolation of missing data and predict future equipment operating status, providing a high-quality data foundation for anomaly detection.

[0019] 4) This invention provides an expert collaborative decision-making mechanism: when the simulation deviation exceeds the threshold, domain experts and fault reasoning models are introduced to jointly diagnose the problem, forming a closed-loop decision of "machine intelligence + human experience" to ensure the scientific nature and accuracy of handling complex faults.

[0020] 5) Based on actual detection feedback, this invention performs incremental learning or retraining on the anomaly detection algorithm to achieve dynamic model optimization, adapt to long-term challenges such as equipment aging and environmental changes, and ensure the continuous high-precision operation of the system.

[0021] 6) This invention uses a digital twin model to simulate the impact of maintenance plans on air traffic efficiency, quantifies key indicators such as downtime and loss of air traffic, and assists in formulating the optimal maintenance strategy to minimize the risk of operational interruption.

[0022] 7) This invention provides a closed-loop link from data to decision-making: covering the entire process of "data acquisition → prediction → detection → simulation → optimization → execution", realizing real-time monitoring of equipment status, early warning, accurate diagnosis and proactive maintenance, which can extend equipment life and reduce operation and maintenance costs.

[0023] 8) This invention constructs a knowledge graph of equipment failure modes, which associates historical failure data, expert experience and real-time detection results, supports multi-dimensional failure reasoning and the generation of handling solutions, and can improve decision-making efficiency in complex scenarios. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Figure 1 This is a flowchart illustrating the health management method for general aviation equipment according to an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the prediction model of the improved SAITS algorithm according to an embodiment of the present invention.

[0027] Figure 3 This is a schematic diagram of the DQN according to an embodiment of the present invention.

[0028] Figure 4 This is a schematic diagram of the knowledge graph modeling in an embodiment of the present invention.

[0029] Figure 5 This is a schematic diagram of a digital twin platform according to an embodiment of the present invention. Detailed Implementation

[0030] The general aviation equipment health management system based on the improved SAITS algorithm in this embodiment includes the following modules: The multi-source data acquisition module collects multi-source data from the general aviation system, including sensor data, general aviation monitoring data, network monitoring data, general aviation dispatch system log data, and manually entered data. The data processing module cleans, integrates, and standardizes the collected multi-source data, and stores it in a time-series format. Log data is parsed into event sequence data using the Drain algorithm, and database data is transformed into time series data using a Python library. The data processed by the data processing module is synchronized to the data base of the digital twin platform through the digital twin module. The interpolation and prediction module uses an improved SAITS algorithm to interpolate missing data and predict the operation data of general aviation equipment in future time periods. The digital twin platform integrates basic data, shared data, business data, 3D data, and monitoring data through a data base; based on a multi-dimensional model framework of "geometry-physical-behavior-rules," it constructs a multi-dimensional digital twin model, such as... Figure 5 As shown, the working data generated by the operation of the physical entity is fused with the simulation data of the multidimensional digital twin model to form twin data. The twin data is used to drive the multidimensional digital twin model to control, optimize and predict the working state of the physical entity. The anomaly detection module uses an adaptive anomaly detection algorithm to identify anomalies in the current and future operational data of general aviation equipment. The root cause analysis module is used to build a knowledge graph model and analyze the root causes of anomalies identified by the anomaly detection module. The digital twin module is used to upload the operational data of the general aviation system to the digital twin platform; upload the anomalies identified by the anomaly detection module to the digital twin platform for simulation to verify the confidence of the predicted anomalies; calculate the deviation between the predicted data and the simulation results of the general aviation equipment; if the deviation exceeds the deviation threshold, the fault is confirmed by combining the opinions of domain experts. The early warning module generates real-time alarm information based on the anomaly detection results and generates a health assessment report for possible future anomalies. The health assessment report includes the current operating status of the equipment, fault type prediction, remaining life assessment, and potential faults that may be induced. The anomaly handling module generates anomaly response strategies for anomalies identified by the anomaly detection module; and generates fault handling plans and maintenance plans based on fault reasoning models for faults identified by the digital twin module. The model correction module optimizes the adaptive anomaly detection model based on the feedback from the duty officer regarding the detection results. The optimization process includes adjusting the detection threshold, updating the anomaly sample library, and retraining the adaptive anomaly detection algorithm.

[0031] like Figure 1 As shown, the general aviation equipment health management method of this embodiment includes the following steps: Step S1: Acquire general aviation equipment operation data; use edge computing nodes to pre-screen the channel topology sensing data and perform spatial correlation verification on the cluster sensor data; establish an environmental factor correction matrix to dynamically compensate sensor data affected by environmental parameters; clean, integrate, and store the general aviation equipment operation data and synchronize it to the digital twin platform.

[0032] In this embodiment, multi-source data is acquired through the navigation business system, including physical indicators of mechanical equipment (such as stress, deformation, and damage), electrical parameters of electronic equipment (voltage, current, temperature, etc.), navigation equipment server logs, lock hydraulic characteristics, navigation monitoring, and other shipping equipment operation and maintenance data. The collected data is integrated, cleaned, missing values ​​are recorded, and then uploaded to the Three Gorges digital twin platform.

[0033] Taking motor parameter signals as an example, the sample format is shown in Table 1.

[0034] Table 1 Motor parameter signals

[0035] NaN represents the missing value.

[0036] An environmental factor correction matrix is ​​established to dynamically compensate sensor data under extreme environments.

[0037] Examples of correction matrices are shown in Table 2.

[0038] Table 2 Sensor Data Correction Algorithm

[0039] Wherein, C_sand represents the real-time monitored sand content value, P represents the theoretical hydraulic pressure under the design conditions, v represents the real-time water flow velocity, P_adj represents the corrected pressure, ΔT represents the temperature change, α represents the thermal expansion coefficient of the material, L represents the original length of the structure, ΔL represents the deformation of the metal structure, S represents the original output value of the sensor, k_fouling represents the bio-adhesion coefficient, and S_corrected represents the corrected sensor data.

[0040] In this embodiment, the drain algorithm is used to transform log data, and the Pandas library in Python is used to convert data from the database into time series data. The sample data is divided into a training set and a test set according to a certain ratio, with 70% as the training set and 30% as the test set, and parameters such as batch size are set.

[0041] The data is standardized using z-score.

[0042] Step S2: Employ the improved SAITS algorithm, such as... Figure 2 As shown, missing data is imputed, and future operational data of general aviation equipment is predicted. Step S201: Masking process; Predefine hyperparameters such as the number of attention heads and layers. For each batch of input... Random manual masking was performed on 15% of the observations, and the time series after manual masking is the masking feature vector. The missing mask vector is ; Step S202: Establish a joint optimization interpolation task MIT and a reconstruction task ORT. MIT is a prediction task for artificially masked values, and ORT is a reconstruction task for observed values. Masking feature vectors and its missing mask vector The connection, after being transformed by a linear layer, serves as the input feature data e of the first diagonally shielded self-attention block DMSA1; the input feature data e of DMSA1 is then position-encoded and processed... The system consists of stacked attention layers, each containing a diagonally masked multi-head attention layer and a feedforward network. A linear layer is then used for projection to obtain the first reconstructed feature vector. ; ; In the formula, We is the weight matrix, be and p are the biases; Z is the output after N attention layers. is the reconstructed matrix; FFN() represents the feedforward neural network function; DiagMaskedMHA() represents the diagonal masking multi-head attention function; These are the weights and biases of the linear layer, respectively. Will and The connection is used as the input feature data of the second diagonally masked self-attention block DMSA2. , ; In the second diagonally shielded self-attention block DMSA2, As input to the linear layer, after positional encoding, it is fed into N stacked attention layers, and then sequentially passes through a linear layer, an activation layer, and another linear layer to obtain the second reconstructed feature vector. , ; This is the input feature data for the second diagonally masked self-attention block DMSA2. This is the weight matrix of the first linear layer of DMSA2. p represents the bias of the linear layer; It is the output feature vector after passing through N attention layers of DMSA2. This is the second reconstructed feature vector; In the weighted combination block, dynamic adjustments are made based on time dependency and available information. and After weighting, a third reconstructed feature vector is obtained. ;use replace The missing values ​​in the data were used to obtain the imputed data. ; ; ; interpolation error This represents the average absolute error calculated between the manually inserted missing values ​​and their respective imputations. ; In the formula It is the mean absolute error function; Reconstruction losses This represents the mean absolute error calculated between the observed values ​​and their respective reconstructions. ; Step S203: Joint optimization training; The final loss function is a weighted sum of the MIT and ORT losses. , As a scaling factor, the network hyperparameters are adjusted by minimizing the loss function to improve the accuracy and generalization ability of the SAITS model; Step S204: Data prediction; The trained SAITS model is used for prediction; the future time series is treated as missing values ​​in the latest time series data and concatenated to the latest time series; the entire time series is input into the trained SAITS model, and after interpolation and reconstruction, prediction results for future time series data are generated for the evaluation of future equipment operation.

[0043] Taking the motor parameter signal as an example, the data sample after interpolation and prediction is shown in Table 3.

[0044] Table 3. Motor parameter signal interpolation data

[0045] Step S3: Based on the predicted data obtained in Step S2, an adaptive anomaly detection algorithm is used to identify the current and future abnormal states of the general aviation equipment; based on the identified equipment anomalies, a knowledge graph of equipment failure modes is constructed, such as... Figure 4 As shown.

[0046] In this embodiment, taking a motor as an example, a deep Q-network (DQN) is used as an adaptive anomaly detection algorithm for the motor prediction parameter signal, specifically including: (a) Define the state space S and the action space A: The state space is the multi-dimensional historical sensor data corresponding to the device, and the action space set A is defined as: judging the current state as "normal", judging the current state as "abnormal", and adjusting the detection threshold. (ii) Setting the reward function: The reward function includes a reward for correctly detecting a device malfunction. Penalty for incorrectly detecting device malfunction Correctly judging the normal operation of the equipment is a reward And penalties for missing equipment malfunctions. The final reward function is a weighted sum of all rewards: ; This is the final reward result; All are weighting coefficients.

[0047] The action value function, Q-value function, is composed of both current and future benefits, and is expressed as follows: The objective Q-value function represents the impact of the current action on the future. ; In the formula, Indicates the discount factor; These represent the rewards at time t, t+1, t+2, ..., t+n, respectively. These represent the states at time t and time t+1, respectively. These represent the actions at time t and time t+1, respectively.

[0048] (iii) Iterative training; Execute actions at each time step Receive rewards and enter the next state. Then sample tuples Return the samples to the experience replay pool; during iterative training, randomly select a batch of samples from the experience replay pool for training; calculate the current state based on the main network. Values ​​and Targets loss function Current state Values ​​and Targets between the mean squared deviations ; The loss function is minimized and iterated until convergence is achieved. (iv) Anomaly detection; Using the future time series predicted by SAITS interpolation as input to DQN, the state of each time node in the time series is obtained, which is the detection and prediction of anomalies.

[0049] The time series data interpolated and predicted by the improved SAITS algorithm are input into the trained DQN model to obtain anomaly predictions. Using motor signals as an example, the anomaly detection results are shown in Table 4.

[0050] Table 4 Motor anomaly detection results

[0051] In the embodiments, K-means clustering is used as an anomaly detection algorithm for data types that have almost no outliers, such as power supply equipment and lock gates (deformation damage).

[0052] Constructing a knowledge graph of equipment failure modes, specifically including: Data collection and processing: Collect logs, call chain data, papers and monographs, online articles and operation manuals generated by the general aviation system, filter the collected data, and process the filtered data into sentences before using it as input for the large language model; Knowledge extraction: Knowledge extraction is performed based on a large language model, and knowledge is divided into entity classes, relation classes, and attribute classes for extraction. Knowledge fusion and ontology construction: The extracted knowledge is fused, and the information of the same entity in multi-source data is integrated. For example, the same device from different data sources is merged into one entity, and the relationship is fused into repeated or contradictory relationship descriptions to form a consistent relationship expression. Graph storage and application: Entity knowledge is stored in the neo4j database and entity relationships are connected. When searching for relevant content in a large model, the neo4j database can be accessed directly.

[0053] Step S4: Construct a digital twin model of the general aviation system. For the equipment anomalies identified in Step S3, use the digital twin platform to simulate the operation of the general aviation system, verify the confidence level of the equipment anomalies identified in Step S3, and generate a health assessment report. If the confidence level is not less than the confidence level threshold, then combine the equipment failure mode knowledge graph to generate anomaly response strategies.

[0054] Step S401: Construct a multidimensional digital twin model and data base; Based on the "geometry-physics-behavior-rules" multidimensional model framework, a multidimensional joint model is established, including a geometric model of ships-locks-channels, a hydrodynamic physical model, a lock operation behavior model, and a navigation rule model, to comprehensively characterize the geometric features, physical properties, behavioral coupling relationships, and behavioral criteria of physical entities.

[0055] The data platform integrates basic data, shared data, business data, 3D data, and monitoring data to provide data support for multidimensional digital twin models.

[0056] Step S402: Generate twin data; Twin data is formed by fusing the working data generated by the operation of physical entities with the simulation data of virtual models. The twin data is used to drive the virtual model to control, optimize and predict the working state of physical entities. Step S403: Fault simulation verification; The anomaly types and parameters identified in step S3 are injected into the digital twin model. Dynamic simulation is performed using the digital twin simulation engine. The confidence level of the fault prediction is calculated based on the degree of agreement between the predicted data of the general aviation equipment and the simulation results. A health assessment report is generated for possible future anomalies. The health assessment report includes the current operating status of the equipment, the predicted fault type, and the potential faults that may be induced.

[0057] Step S404: Exception Handling; If the confidence level is not less than the confidence level threshold, an anomaly response strategy is generated by combining the equipment failure mode knowledge graph; the deviation between the predicted data and simulation results of the general aviation equipment is calculated; if the deviation exceeds the deviation threshold, the domain experts further diagnose and confirm the failure, and a failure handling plan is formed by combining the expert opinions and adopting a failure reasoning model.

[0058] In this embodiment, the Deepseek-R1 model from Deepseek Inc. is used as the fault reasoning model.

[0059] The Deepseek-R1 model features small parameters and convenient local deployment. After deployment on a local server, it can be combined with the aforementioned knowledge graph to serve as an auxiliary tool for handling faults and generating fault resolution solutions.

[0060] The following is an example of a patterned input for a fault reasoning model: [Clearly describe the abnormal phenomenon occurring in the device / system] "The hydraulic gate hoist of the third-stage lock of the Three Gorges South Route Ship Lock operates under the condition of high water level difference (ΔH>28m) during the flood season:" - Pressure pulsation amplitude exceeds threshold (+32%) - Door closing time extended (designed 120s → actual measured 158s) - Accompanied by high-frequency vibration (250Hz component increases by 15dB) Step S5: Based on the actual feedback from general aviation equipment detection, perform incremental learning or retraining on the adaptive anomaly detection algorithm to improve detection accuracy.

[0061] After handling alarms, the duty officer will provide feedback on the accuracy of the detection results. If an anomaly is correctly detected, the sample is marked as an anomaly and saved. If a false positive occurs, the sensitivity of the adaptive anomaly detection algorithm is increased, and normal data is added for incremental learning to prevent model overfitting. If a false negative occurs, the threshold of the adaptive anomaly detection algorithm is adjusted, the reward function weights in the DQN model are adjusted, and the anomaly event data is labeled with an anomaly event before incremental learning.

[0062] Step S6: Generate a maintenance plan based on the anomaly response strategy or fault handling scheme, and simultaneously synchronize the maintenance plan to the station dispatch system to recreate the dispatch plan; simultaneously upload it to the digital twin platform, and based on the virtual debugging function, simulate the impact of the maintenance scheme on traffic volume and efficiency.

[0063] Based on the implementation results, this invention, by using equipment sensors, monitoring systems, and operation and maintenance logs as sample data, can accurately and comprehensively assess and predict the operating status of equipment. Combined with the improved SAITS algorithm, it can effectively complete missing values ​​in multivariate time series data and efficiently predict future operating data of shipping equipment, providing accurate data support for fault detection and prediction of Three Gorges shipping equipment. Employing an adaptive anomaly detection algorithm, it can dynamically learn and adaptively adjust, accurately identifying anomalies in SAITS prediction data and effectively reducing the risk of false positives and false negatives. Combined with the Three Gorges equipment knowledge graph, it can quickly and accurately locate fault types and potential related faults. The feedback mechanism enhances the fit to the data, further improving the model's accuracy and generalization ability. The integration of the digital twin platform provides comprehensive support for the health management of equipment in the Three Gorges navigation system, effectively improving prediction accuracy and collaborative operation and maintenance within the basin.

[0064] Taking the motors of the Three Gorges Dam ship locks as an example, the accuracy of motor fault detection has increased by approximately 23%, the false alarm rate has decreased from 18% to 4.5%, the average warning time has increased from 4 days to 21 days, the cost per maintenance has been significantly reduced, the frequency of unplanned downtime has decreased by approximately 65%, and losses due to navigation losses have been reduced by approximately 75 million yuan per year. Furthermore, through simulation and prediction using a digital twin platform, the staggered three-day maintenance between the Three Gorges and Gezhouba Dams can reduce shipping losses by approximately 120 million yuan, and the shared testing platform saves over 5 million yuan in costs; through spare parts sharing and expert collaboration, fault recovery time has been improved by 300%.

[0065] Based on the implementation results, this invention significantly improves the operational efficiency of navigation equipment at the Three Gorges Dam through the deep integration of a digital twin platform and intelligent health management technology. At the same time, based on basin-level data sharing and multi-hub collaborative operation and maintenance, it promotes resource complementarity and joint scheduling of hubs such as Gezhouba Dam, thereby improving the power generation efficiency of the entire basin and forming a comprehensive benefit integrating safety, economy and ecology. This provides a technical benchmark for the intelligent management of the Yangtze River's golden waterway.

Claims

1. A general aviation equipment health management system based on an improved SAITS algorithm, characterized in that, Includes the following modules: The multi-source data acquisition module collects multi-source data from the general aviation system, including sensor data, general aviation monitoring data, network monitoring data, and general aviation dispatching system log data. The data processing module cleans, integrates, and standardizes multi-source data, and stores it in time-series format; The interpolation and prediction module uses the improved SAITS algorithm to interpolate missing data in multi-source data, and at the same time predicts the operation data of general aviation equipment in future time periods. The anomaly detection module uses an adaptive anomaly detection algorithm to identify anomalies in the current and future operational data of general aviation equipment. The root cause analysis module is used to build a knowledge graph model and analyze the root causes of anomalies identified by the anomaly detection module. The digital twin module constructs a multi-dimensional digital twin model of the general aviation system. It integrates the working data generated by the operation of the physical entity with the simulation data of the multi-dimensional digital twin model to form twin data. The anomaly types and parameters identified by the anomaly detection module are injected into the multi-dimensional digital twin model for dynamic simulation to verify the confidence level of the equipment anomalies identified by the anomaly detection module. The early warning module generates real-time alarm information based on the anomaly detection results and generates a health assessment report for possible future anomalies. The health assessment report includes the current operating status of the equipment, fault type prediction, remaining life assessment, and potential faults that may be induced.

2. The general aviation equipment health management system according to claim 1, characterized in that, The multi-source data acquisition module uses edge computing nodes to pre-screen channel topology sensing data and perform spatial correlation verification on cluster sensor data.

3. The general aviation equipment health management system according to claim 1, characterized in that, In the interpolation prediction module, the improvements of the improved SAITS algorithm over the existing SAITS algorithm include: 1) STL decomposition is used for seasonal data. The operation of the navigation system is closely related to the water level, and the shipping volume is related to the flood season and the dry season. Therefore, the operation of some equipment shows obvious seasonal trends. By using STL decomposition technology, the original navigation system time series data is decomposed into trend, seasonality and residual, which are used as inputs to the SAITS model. This can more effectively impute missing values ​​in seasonal data and noisy data. 2) Abnormal sample augmentation: SAITS models lose accuracy when faced with outliers. By generating abnormal samples through data generation techniques and increasing the weight of abnormal samples when training the SAITS model, the learning ability of abnormal data can be effectively strengthened, and the accuracy of prediction of abnormal data can be enhanced. 3) Introducing a feedback mechanism: The SAITS model has a large error in predicting data, which can lead to anomaly detection errors. Introducing a feedback mechanism can improve the accuracy and generalization ability of the SAITS model. Based on the error type, such as false positives and false negatives, the strategy is adaptively adjusted and retrained, including adjusting the masking ratio, adding attention layers, and retraining using abnormal samples.

4. The general aviation equipment health management system according to claim 3, characterized in that, The interpolation prediction module is configured to perform the following steps: 1) Concealment measures; Assume the input to the SAITS model is Introduce an indicator mask vector for the original missing values. For each batch of input Observations are randomly selected and artificially masked according to a set ratio; the time series after artificial masking is the masking feature vector. The missing mask vector is ; 2) Establish a joint optimization interpolation task MIT and a reconstruction task ORT. MIT is the prediction task for artificially masked values, and ORT is the reconstruction task for observed values. Masking feature vectors and its missing mask vector The connection, after being transformed by a linear layer, serves as the input feature data e of the first diagonally shielded self-attention block DMSA1; the input feature data e of DMSA1 is then position-encoded and processed... The system consists of stacked attention layers, each containing a diagonally masked multi-head attention layer and a feedforward network. A linear layer is then used for projection to obtain the first reconstructed feature vector. ; ； In the formula, We is the weight matrix, be and p are the biases; Z is the output after N attention layers. is the reconstructed matrix; FFN() represents the feedforward neural network function; DiagMaskedMHA() represents the diagonal masking multi-head attention function; These are the weights and biases of the linear layer, respectively. Will and The connection is used as the input feature data of the second diagonally masked self-attention block DMSA2. , ; In the second diagonally shielded self-attention block DMSA2, As input to the linear layer, after positional encoding, it is fed into N stacked attention layers, and then sequentially passes through a linear layer, an activation layer, and another linear layer to obtain the second reconstructed feature vector. , ； This is the input feature data for the second diagonally masked self-attention block DMSA2. This is the weight matrix of the first linear layer of DMSA2. p represents the bias of the linear layer; It is the output feature vector after passing through N attention layers of DMSA2. This is the second reconstructed feature vector; In the weighted combination block, dynamic adjustments are made based on time dependency and available information. and After weighting, a third reconstructed feature vector is obtained. ;use replace The missing values ​​in the data were used to obtain the imputed data. ; ； ； interpolation error This represents the average absolute error calculated between the manually inserted missing values ​​and their respective imputations. ; In the formula It is the mean absolute error function; Reconstruction losses This represents the mean absolute error calculated between the observed values ​​and their respective reconstructions. ; 3) Joint optimization training; The final loss function is a weighted sum of the MIT and ORT losses. , As a scaling factor, the network hyperparameters are adjusted by minimizing the loss function to improve the accuracy and generalization ability of the SAITS model; 4) Make predictions using the trained SAITS model; In the latest time series data, future time series are treated as missing values ​​and concatenated after the latest time series; The entire time series is input into the trained SAITS model, and after interpolation and reconstruction, prediction results for future time series data are generated for the evaluation of future equipment operation.

5. The general aviation equipment health management system according to claim 1, characterized in that, In the anomaly detection module, a deep Q-network (DQN) is used as one of the adaptive anomaly detection algorithms, specifically including: (a) Define the state space S and the action space A: The state space is the multi-dimensional historical sensor data corresponding to the device, and the action space set A is defined as: judging the current state as "normal", judging the current state as "abnormal", and adjusting the detection threshold. (ii) Setting the reward function: The reward function includes a reward for correctly detecting a device malfunction. Penalty for incorrectly detecting device malfunction Correctly judging the normal operation of the equipment is a reward And penalties for missing equipment malfunctions. The final reward function is a weighted sum of all rewards: ； This is the final reward result; All are weighting coefficients; The action value function, Q-value function, is composed of both current and future benefits, and is expressed as follows: The objective Q-value function represents the impact of the current action on the future. ； In the formula, Indicates the discount factor; These represent the rewards at time t, t+1, t+2, ..., t+n, respectively. These represent the states at time t and time t+1, respectively. These represent the actions at time t and time t+1, respectively. (iii) Iterative training; Execute actions at each time step Receive rewards and enter the next state. Then sample tuples Return the samples to the experience replay pool; during iterative training, randomly select a batch of samples from the experience replay pool for training; calculate the current state based on the main network. Values ​​and Targets loss function Current state Values ​​and Targets between the mean squared deviations ； The loss function is minimized and iterated until convergence is achieved. (iv) Anomaly detection; Using the future time series predicted by SAITS interpolation as input to DQN, the state of each time node in the time series is obtained, which is the detection and prediction of anomalies.

6. The general aviation equipment health management system according to claim 1, characterized in that, In the root cause analysis module, building a knowledge graph of equipment failure modes includes: Data collection and processing: Collect logs, call chain data and operation manuals generated by the general aviation system, filter the collected data, and process the filtered data into sentences before using it as input for the large language model; Knowledge extraction: Knowledge extraction is performed based on a large language model, and knowledge is divided into entity classes, relation classes, and attribute classes for extraction. Knowledge fusion and ontology construction: The extracted knowledge is fused, and the information of the same entity in multi-source data is integrated. For example, the same device from different data sources is merged into one entity, and the relationship is fused into repeated or contradictory relationship descriptions to form a consistent relationship expression. Graph storage and application: Use the neo4j database to store entity knowledge and connect entity relationships.

7. The general aviation equipment health management system according to claim 1, characterized in that, The digital twin module is configured to perform the following steps: Constructing a multidimensional digital twin model: Based on the "geometry-physics-behavior-rules" multidimensional model framework, a multidimensional joint model is established, including a geometric model of the ship-lock-channel, a hydrodynamic physical model, a lock operation behavior model, and a navigation rule model, in order to characterize the geometric features, physical properties, behavioral coupling relationships, and behavioral criteria of the physical entity; Generate twin data: Twin data is formed by fusing the working data generated by the operation of physical entities with the simulation data of virtual models. The twin data is used to drive the virtual model to control, optimize and predict the working state of physical entities. Fault simulation verification: The identified anomaly types and parameters are injected into the digital twin model, and dynamic simulation is performed using the digital twin simulation engine to verify the confidence level of the equipment anomalies identified by the adaptive anomaly detection algorithm.

8. The general aviation equipment health management system according to claim 1, characterized in that, The system also includes an anomaly handling module: for anomalies identified by the anomaly detection module, it generates anomaly response strategies; for faults judged by the digital twin module, it generates fault handling plans and maintenance plans by combining fault reasoning models.

9. The general aviation equipment health management system according to claim 1, characterized in that, The system also includes a model correction module: optimizing the adaptive anomaly detection algorithm model based on the feedback from the duty officer on the detection results. The optimization process includes adjusting the detection threshold, updating the anomaly sample library, and retraining the adaptive anomaly detection algorithm.

10. The method for a general aviation equipment health management system based on the improved SAITS algorithm as described in any one of claims 1-9, characterized in that, Includes the following steps: S1: Obtain operational data of general aviation equipment; Edge computing nodes are used to pre-screen channel topology sensing data and perform spatial correlation verification on cluster sensor data. An environmental factor correction matrix is ​​established to dynamically compensate sensor data affected by environmental parameters; operational data of general aviation equipment is cleaned, integrated, and stored. S2: An improved SAITS algorithm is used to impute missing data and predict future operational data of general aviation equipment; S3: Based on the predicted data obtained in step S2, an adaptive anomaly detection algorithm is used to identify the current and future abnormal states of general aviation equipment; based on the identified equipment anomalies, an equipment failure mode knowledge graph is constructed. S4: Construct a digital twin model of the general aviation system. For the equipment anomalies identified in step S3, use the digital twin platform to simulate the operation of the general aviation system, verify the confidence level of the equipment anomalies identified in step S3, and generate a health assessment report. If the confidence level is not less than the confidence level threshold, then combine the equipment failure mode knowledge graph to generate anomaly response strategies. S5: Based on actual feedback from general aviation equipment testing, incremental learning or retraining of the adaptive anomaly detection algorithm is performed to improve detection accuracy.

Images