Aircraft engine detection method and system based on multi-task time series analysis

Abstract

The application provides an aircraft engine detection method and system based on multi-task time series analysis, and relates to the technical field of aviation. First, an aircraft engine operation monitoring data set is acquired, which contains ECU operation data and fault data corresponding to a data acquisition time sequence. Then, cross-period comparison representation learning processing is performed on the aircraft engine operation monitoring data set to generate a target fault feature set. Next, a time series data set is constructed based on the target fault feature set and the data acquisition time sequence, a time series prediction model with fused multi-scale attention is called to generate a multi-scale time series prediction result. Then, a dynamic standard baseline curve of key parameters of an aircraft engine exhaust valve control system is established according to the prediction result. Finally, real-time ECU operation data is compared and analyzed with the dynamic standard baseline curve to generate an aircraft engine operation state detection result. The application can accurately detect the engine operation state and realize early fault warning.

Description

Technical Field

[0001] This invention relates to the field of aviation technology, and more specifically, to an aircraft engine detection method and system based on multi-task timing analysis. Background Technology

[0002] In the aviation field, aircraft engines are core components of aircraft, and their operational stability and reliability directly affect flight safety. Currently, the inspection of aircraft engines mainly relies on traditional monitoring methods and simple data analysis techniques.

[0003] Traditional monitoring methods typically involve installing various sensors on the engine to collect its operating parameters in real time, such as temperature, pressure, and speed. Then, by setting fixed thresholds, the collected parameters are compared to these thresholds; if a parameter exceeds the threshold range, an engine malfunction is determined. However, this method has significant limitations. Firstly, fixed thresholds are difficult to adapt to the complex changes in engine operating conditions. For example, the engine's operating status and parameter ranges vary greatly during different flight phases (takeoff, cruise, and landing), and fixed thresholds may lead to misjudgments or missed diagnoses. Secondly, traditional methods only focus on the anomaly of a single parameter, failing to comprehensively consider the correlation and temporal changes between multiple parameters, making it difficult to accurately identify potential and early-stage engine malfunctions.

[0004] Furthermore, most existing data analysis techniques are based on simple statistical methods or single models, which do not delve deeply enough into engine operating data. The aforementioned methods cannot effectively extract the differentiated features between normal operating conditions and fault operating conditions under different operating stages, nor can they accurately predict the dynamic change trends of key engine parameters, thus limiting the ability to accurately detect engine operating conditions and provide fault warnings. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an aircraft engine detection method based on multi-task timing analysis, the method comprising: Acquire a set of operational monitoring data generated by an aircraft engine during operation. The set of operational monitoring data includes ECU operational data and fault data, and the ECU operational data and fault data maintain a one-to-one correspondence with the data acquisition time series. The operation monitoring data set is subjected to cross-time period comparison representation learning processing. The difference features between normal operation status data and fault operation status data under different operating periods are extracted through the comparison representation learning model to generate a target fault feature set. The target fault feature set corresponds one-to-one with the information of each dimension of the ECU operation data. Based on the target fault feature set and the data acquisition time series, a time series dataset of the aircraft engine exhaust valve control system is constructed. A time series prediction model that integrates multi-scale attention is called to perform time series modeling processing on the time series dataset to generate multi-scale time series prediction results. Based on the multi-scale time-series prediction results, a dynamic standard reference curve for the key parameters of the aircraft engine exhaust valve control system is established. The dynamic standard reference curve corresponds to the key parameters of each dimension of the ECU operating data. The real-time ECU operating data collected from the aircraft engine is compared and analyzed in real time with the dynamic standard reference curve. By combining the correlation of target fault characteristics, abnormal information in which the real-time ECU operating data deviates from the dynamic standard reference curve is identified, and an aircraft engine operating status detection result is generated. The aircraft engine operating status detection result includes abnormal parameter dimension identifiers and abnormality degree information.

[0006] Furthermore, embodiments of the present invention also provide an aircraft engine testing system based on multi-task timing analysis, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned aircraft engine detection method based on multi-task timing analysis by executing the machine-executable instructions.

[0007] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described aircraft engine detection method based on multi-task timing analysis.

[0008] Based on the above, by acquiring a set of operational monitoring data containing ECU operating data and fault data that corresponds one-to-one with the data acquisition time series during aircraft engine operation, and performing cross-time period comparative representation learning processing on the operational monitoring data set, it is possible to deeply explore the differentiated characteristics of normal and faulty operating states under different operating stages. This generates a set of target fault features that corresponds one-to-one with the various dimensions of ECU operating data. A time-series dataset is constructed based on the target fault feature set and the data acquisition time series, and a time-series prediction model incorporating multi-scale attention is used for time-series modeling processing to generate multi-scale time-series prediction results. This fully considers the changing patterns of engine operating data at different time scales. Based on the multi-scale time-series prediction results, a dynamic standard benchmark curve for key parameters of the aircraft engine exhaust valve control system is established. This dynamic standard benchmark curve corresponds to each dimension of key parameters in the ECU operating data, and can dynamically adapt to changes in engine operating conditions. Finally, real-time comparison and analysis are performed between real-time ECU operating data and the dynamic standard benchmark curve. Combined with the correlation of target fault features, abnormal information is identified, generating aircraft engine operating status detection results containing abnormal parameter dimension identifiers and abnormality degree information. This achieves accurate detection of engine operating status and early warning of faults. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the execution flow of the aircraft engine detection method based on multi-task timing analysis provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of exemplary hardware and software components of an aircraft engine detection system based on multi-task timing analysis provided in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an aircraft engine detection method based on multi-task time series analysis according to an embodiment of the present invention. The following is a detailed description of the aircraft engine detection method based on multi-task time series analysis.

[0012] Step S110: Obtain the set of operation monitoring data generated by the aircraft engine during operation. The set of operation monitoring data includes ECU operation data and fault data. The ECU operation data and fault data maintain a one-to-one correspondence with the data acquisition time series.

[0013] In this embodiment, the object to be monitored is a turboprop aircraft engine used in short-haul regional flights, whose daily operation covers three main phases: takeoff, cruise, and landing. When acquiring the operational monitoring data set, data is first collected in real time through the engine's built-in sensor network (including intake pressure sensors, intake temperature sensors, propeller speed sensors, common rail pressure sensors, etc.), and then transmitted to the engine electronic control unit (ECU) storage module. The ECU operational data includes intake pressure data, intake temperature data, propeller speed data, and common rail pressure data, while fault data includes information such as exhaust valve jamming alarms and pressure sensor anomaly alarms. The data acquisition time series is generated at fixed intervals, with each time node corresponding to a complete set of ECU operational data and a record of whether fault data exists at that moment, ensuring a one-to-one correspondence. Finally, the data is integrated to form an operational monitoring data set covering multiple consecutive operating days (including different flight missions).

[0014] Step S111: Extract historical operating data and historical fault data from the ECU storage module of the aircraft engine in multiple consecutive operating periods, and combine the historical operating data and historical fault data of all periods to form an initial data set.

[0015] Historical operational and fault data from multiple consecutive operating days were extracted from the ECU storage module of the turboprop aircraft engine. Each operating day was divided into multiple operating segments (takeoff, cruise, and landing), forming multiple sets of consecutive operating segments. The historical operational data included real-time collected values ​​of intake pressure, intake temperature, propeller speed, and common rail pressure for each segment. The historical fault data included the fault alarm type, alarm trigger time, and alarm clearing time for each segment. All historical operational and fault data from all segments were combined chronologically to form an initial dataset. Each data entry in this dataset was labeled with its corresponding operating segment and data acquisition timestamp.

[0016] Step S112: The initial data set is split into a subset of intake pressure data, a subset of intake temperature data, a subset of propeller speed data, a subset of common rail pressure data, and a subset of fault alarm data. Each data subset contains the correspondence between the corresponding dimension data and the data acquisition time series, and each data subset covers all continuous running segments.

[0017] The initial dataset is split along data dimensions: the intake pressure data subset retains only the intake pressure data collected for all time periods and their corresponding time series; the intake temperature data subset retains only the intake temperature data collected for all time periods and their corresponding time series; the propeller speed data subset retains only the propeller speed data collected for all time periods and their corresponding time series; the common rail pressure data subset retains only the common rail pressure data collected for all time periods and their corresponding time series; and the fault alarm data subset retains the fault alarm type, trigger time, release time, and corresponding time series for all time periods. Each data subset covers all continuous operating periods, and the timestamp of each data entry is fully aligned with other subsets. For example, the data collected at a certain timestamp in the intake pressure data subset can be directly associated with the alarm status at the same timestamp in the fault alarm data subset.

[0018] Step S113: Perform time-segmented data cleaning for each data subset. After standardizing each data subset after cleaning in each time segment, integrate the standardized intake pressure data subset, standardized intake temperature data subset, standardized propeller speed data subset, standardized common rail pressure data subset, and standardized fault alarm data subset of all time segments across time segments according to the data acquisition time sequence to form the operation monitoring data set.

[0019] During the time-segmented data cleaning process, for each runtime segment of each data subset, outliers (such as sudden increases or decreases in intake pressure data that exceed the normal operating range for that time segment) are first identified and replaced using linear interpolation of data from adjacent time nodes. Duplicate data (such as repeated collection values ​​at the same timestamp) is then deleted, retaining only the earliest recorded value. After cleaning, each data subset for each time segment is standardized: for continuous data such as intake pressure, intake temperature, propeller speed, and common rail pressure, a linear transformation is used to map the data to a fixed interval, eliminating dimensional differences between different dimensions; for fault alarm data (discrete), "no alarm" is marked with a specific symbol, and "alarm" is marked with different symbols according to the alarm type, achieving standardization. Finally, the standardized data subsets from all time segments are concatenated in chronological order of data collection time, ensuring that each time node contains standardized data across all five dimensions, forming the final operational monitoring data set.

[0020] Step S120: Perform cross-time period comparison representation learning processing on the operation monitoring data set. Extract the differential features between normal operation status data and fault operation status data under different operating periods through the comparison representation learning model, and generate a target fault feature set. The target fault feature set corresponds one-to-one with the various dimensions of the ECU operation data.

[0021] Step S121: Separate the normal operation status data subset and the fault operation status data subset for each time period from the operation monitoring data set according to the operation period. The normal operation status data subset includes ECU operation data of each dimension corresponding to the no-fault alarm data, and the fault operation status data subset includes ECU operation data of each dimension corresponding to the fault alarm data. Each time period corresponds to a set of normal operation status data subset and fault operation status data subset.

[0022] Data is separated for each operating period: For each period, firstly, "no alarm" time nodes are selected from the fault alarm data subset, and the corresponding intake pressure, intake temperature, propeller speed, and common rail pressure data are extracted to form the normal operating status data subset for that period; then, "alarm" time nodes (such as exhaust valve jamming alarm) are selected, and the corresponding four-dimensional ECU operating data are extracted to form the fault operating status data subset for that period. For example, during the cruise period of a certain operating day, the data corresponding to multiple time nodes without alarms form the normal operating status data subset for that period, and the data corresponding to multiple time nodes with exhaust valve jamming alarms form the fault operating status data subset for that period, ensuring that there are two sets of data subsets for each period.

[0023] Step S122: Input the subset of normal operation status data for all time periods into the feature encoding layer of the contrastive representation learning model for feature mapping processing, generate the normal status feature vector for each time period, and calculate the mean of the normal status feature vectors for all time periods to obtain a unified normal status feature vector across time periods.

[0024] The feature encoding layer of the contrastive representation learning model adopts a fully connected neural network structure. The number of input layer nodes is consistent with the number of dimensions of the ECU operating data. The hidden layer contains two layers, and the output layer outputs a fixed-length feature vector. Subsets of normal operating status data from all time periods are input into the feature encoding layer one by one. For the normal data of each time period, the four-dimensional raw data are mapped into a high-dimensional feature vector through linear transformation and activation function processing in the hidden layer, generating a normal state feature vector for that time period (each vector contains multiple feature dimensions, reflecting the characteristics of normal operation in that time period). After processing all time periods, the mean of all normal state feature vectors is taken according to their corresponding feature dimensions. That is, the mean of each feature dimension is equal to the sum of the feature values ​​of that dimension in all time periods divided by the number of time periods, resulting in a unified normal state feature vector across time periods. This vector reflects the common characteristics of normal operation in all time periods.

[0025] Step S123: Input a subset of fault operation status data from all time periods into the feature encoding layer of the contrastive representation learning model, perform feature mapping processing on the fault data of each time period, generate a fault status feature vector for each time period, calculate the mean of the fault status feature vectors of all time periods, and obtain a unified fault status feature vector across time periods.

[0026] Consistent with the normal data processing flow, subsets of fault operation status data from all time periods are sequentially input into the feature encoding layer of the contrastive representation learning model. After processing by the feature encoding layer, the fault data for each time period generates a fault status feature vector for that period (with the same length as the normal status feature vector, and corresponding feature dimensions, reflecting the characteristics of fault operation during that period). For example, the fault status feature vector for a cruise period on a certain operating day may contain features reflecting pressure and temperature characteristics when the exhaust valve is stuck. The average of all fault status feature vectors is taken according to their corresponding feature dimensions to obtain a unified fault status feature vector across time periods. This vector reflects the common characteristics of fault operation across all time periods.

[0027] Step S124: Call the Detector model to train the unified normal state feature vector and the unified fault state feature vector across time periods. Use the time-segmented cross-validation method to adjust the hyperparameters of the Detector model so that the Detector model can distinguish between the normal state feature vector and the fault state feature vector in different time periods.

[0028] Step S1241: Combine the normal state feature vector and the fault state feature vector of each time period to form a sample subset for that time period. Add a category label and a time period label to the samples in each sample subset. The category label includes a normal state label and a fault state label. The time period label identifies the runtime segment to which the sample belongs.

[0029] For each time period, the normal state feature vector (each vector as a sample) of that time period is labeled with the category label "normal" and the time period label (e.g., "a certain operating day - takeoff time period"). The fault state feature vector of that time period is labeled with the category label "fault" and the same time period label. These are combined to form a sample subset for that time period. Multiple sample subsets are formed for all time periods. Each sample subset contains two categories of samples, "normal" and "fault", and the corresponding time period identifier.

[0030] Step S1242: Combine the sample subsets of all time periods to form the training sample set of the Detector model. Divide the training sample set of the Detector model into multiple consecutive sub-sample sequences according to the time period order. Each sub-sample sequence contains sample subsets of at least two time periods.

[0031] All sample subsets are combined in time period order (from the takeoff time of the first operating day to the landing time of the last operating day) to form the training sample set for the Detector model. According to the principle of grouping by consecutive time periods, the training sample set is divided into multiple consecutive sub-sample sequences. Each sub-sample sequence contains sample subsets from multiple consecutive time periods, ensuring that each sub-sample sequence covers different operating phases (takeoff, cruise, and landing).

[0032] Step S1243: Using a time-segmented cross-validation strategy, select one sub-sample sequence as the validation sub-sequence and the remaining sub-sample sequences as the training sub-sequences. Input the training sub-sequences into the Detector model to start the training process of the Detector model.

[0033] In the first round of cross-validation, the first sub-sample sequence is selected as the validation sub-sequence, and the remaining sub-sample sequences are used as training sub-sequences. All samples from the training sub-sequences are input into the Detector model. The input layer of the Detector model receives the feature vector, the hidden layer is responsible for further feature extraction, and the classifier module outputs a prediction result indicating whether the sample belongs to "normal" or "faulty". During training, the cross-entropy loss function is used to calculate the error between the predicted result and the true class label, and the model parameters are updated using the backpropagation algorithm.

[0034] Step S1244: During the training of the Detector model, the classifier module of the Detector model learns the class distinction boundary and time period adaptation parameters at the same time, so that the Detector model can distinguish between normal and fault states and adapt to the feature differences of different time periods, and periodically calculates the training loss value of the training subsequence.

[0035] The classifier module of the Detector model includes a class discrimination branch and a time-period adaptation branch: the class discrimination branch establishes a "normal-fault" class distinction boundary by learning the feature differences between samples of different classes; the time-period adaptation branch learns time-period adaptation parameters (such as feature offsets for different time periods) by analyzing the feature fluctuations of samples in different time periods of the training subsequence to offset the impact of time-period differences on the classification results. During training, after a certain number of parameter updates (i.e., one training batch), the training loss value (output of the cross-entropy loss function) of the current training subsequence is calculated, and the curve of the loss value changing with the training batch is recorded.

[0036] Step S1245: When the training loss value is reduced to the set training threshold, the validation subsequence is input into the Detector model, the validation classification result output by the Detector model is obtained, the classification accuracy of the sample in the validation subsequence for each time period is calculated, and the average classification accuracy for all time periods is calculated.

[0037] Training of the current training subsequence is stopped when the training loss value remains consistently below the set training threshold for multiple training batches. Samples from the validation subsequence are input into the Detector model to obtain the classification prediction result for each sample. For each time period in the validation subsequence, the proportion of samples whose predicted results match the true labels within that time period is calculated out of the total number of samples in that time period, yielding the classification accuracy for each time period. The classification accuracy of all time periods is then summed and divided by the number of time periods to obtain the average classification accuracy across all time periods.

[0038] Step S1246: Adjust the hyperparameters of the Detector model based on the classification accuracy and average classification accuracy for each time period. The hyperparameters of the Detector model include the number of hidden layer neurons, activation function type, regularization coefficient, and time period adaptation coefficient.

[0039] If the classification accuracy for a certain period (such as a cruise period) is lower than the period accuracy threshold, it indicates that the model is not well adapted to the features of that period. The weight of the period adaptation coefficient corresponding to that period needs to be increased so that the model pays more attention to the feature differences of that period in subsequent training. If the average classification accuracy is lower than the overall accuracy threshold, it indicates that the model's overall classification ability is insufficient. The number of neurons in the hidden layer needs to be adjusted (such as increasing the number of neurons to improve feature extraction ability) or the activation function type needs to be changed (such as changing one type of activation function to another type of activation function to alleviate gradient vanishing). At the same time, the regularization coefficient should be adjusted to avoid model overfitting.

[0040] Step S1247: If the classification accuracy of any time period is lower than the time period accuracy threshold, increase the weight of the time period adaptation coefficient corresponding to that time period and retrain the Detector model; if the average classification accuracy is lower than the overall accuracy threshold, adjust the number of hidden layer neurons or the activation function type of the Detector model to reduce training loss.

[0041] For example, if the classification accuracy is lower than the threshold for a certain period, the weight of the time period adaptation coefficient is increased, the training subsequence is re-inputted to start training, and the training loss value is observed to further decrease and the classification accuracy for that period is improved. If the average classification accuracy is lower than the threshold, the number of neurons in the hidden layer is adjusted or the activation function type is changed. After retraining, the new training loss value and classification accuracy are calculated until the average classification accuracy reaches the target.

[0042] Step S1248: Take each sub-sample sequence as a validation sub-sequence and the rest as training sub-sequences, and repeat the training, validation and hyperparameter adjustment process of the Detector model to obtain multiple sets of hyperparameter combinations and the corresponding classification accuracy and average classification accuracy for each time period.

[0043] After completing the first round of cross-validation, other sub-sample sequences are selected sequentially as validation sub-sequences, and the rest are used as training sub-sequences. Steps S1243 to S1247 are repeated. After each round of validation, the current hyperparameter combination (number of hidden layer neurons, activation function type, regularization coefficient, and adaptation coefficient for each time period) and the corresponding classification accuracy and average classification accuracy for each time period are recorded. Finally, multiple sets of hyperparameter combinations and their corresponding performance indicators are obtained.

[0044] Step S1249: Select the hyperparameter combination with the highest average classification accuracy and the smallest difference in classification accuracy across different time periods as the final hyperparameters of the Detector model. Use the entire training sample set to perform the final training of the Detector model. At the same time, introduce a time-adaptive loss function to reduce the impact of feature differences across different time periods on the classification results, and obtain the trained Detector model.

[0045] From multiple hyperparameter combinations, the combination with the highest average classification accuracy is selected. If multiple combinations have the same average accuracy, the combination with the smallest standard deviation of classification accuracy across all time periods (i.e., the most stable performance across time periods) is chosen as the final hyperparameter. The entire training sample set (samples from all time periods) is input into the Detector model, and training is initiated using the final hyperparameter. A time-adaptive loss function is introduced—this function, based on the cross-entropy loss, adds a penalty term for feature differences between samples from different time periods, enabling the model to learn class distinctions while further reducing interference from time-period differences. After training, the model performance is verified using the classification accuracy of all samples, ensuring that the model can stably distinguish between normal and fault state feature vectors across different time periods, resulting in the trained Detector model.

[0046] Step S125: Calculate the feature difference between the cross-time period unified normal state feature vector and the cross-time period unified fault state feature vector using the trained Detector model. At the same time, calculate the deviation between the normal state feature vector of each time period and the cross-time period unified normal state feature vector, and the deviation between the fault state feature vector of each time period and the cross-time period unified fault state feature vector.

[0047] The trained Detector model includes a feature difference calculation module. This module inputs the unified normal state feature vector and the unified fault state feature vector across time periods. The feature difference degree is obtained by calculating the sum of the absolute differences in the corresponding feature dimensions of the two vectors (the larger the sum of the differences, the more significant the feature difference between normal and fault states). Simultaneously, for each time period's normal state feature vector, the sum of the absolute differences in the corresponding dimensions between it and the unified normal state feature vector across time periods is calculated to obtain the deviation degree of the normal state for that time period (reflecting the difference between the normal features of that time period and the common normal features). Similarly, for each time period's fault state feature vector, the sum of the absolute differences in the corresponding dimensions between it and the unified fault state feature vector across time periods is calculated to obtain the deviation degree of the fault state for that time period (reflecting the difference between the fault features of that time period and the common fault features).

[0048] Step S126: Extract the feature dimensions whose feature difference is greater than a set threshold and whose deviation is less than the deviation threshold in each time period, and use them as the target feature dimensions.

[0049] Set a feature difference threshold (to filter feature dimensions with high discriminative power) and a deviation threshold (to filter feature dimensions with strong commonality across time periods). Iterate through all feature dimensions, first determining if the feature difference of that dimension is greater than the set threshold, then determining if the deviation of the normal state and the deviation of the fault state for that dimension across all time periods are both less than the deviation threshold. For example, if the feature difference of a certain feature dimension is greater than the set threshold, and the normal deviation and the fault deviation for that dimension across all time periods are both less than the deviation threshold, it indicates that this dimension can effectively distinguish between normal and fault states and also has commonalities across time periods; therefore, this dimension is determined as the target feature dimension. Finally, multiple target feature dimensions are selected, covering feature dimensions related to intake pressure, intake temperature, propeller speed, and common rail pressure.

[0050] Step S127: For each target feature dimension, extract the first feature value distribution range of the cross-time period unified normal state feature vector and the second feature value distribution range of the cross-time period unified fault state feature vector under that target feature dimension, and calculate the difference region and the proportion of the difference region between the first feature value distribution range and the second feature value distribution range.

[0051] For each target feature dimension, firstly, the fluctuation range of the characteristic values ​​of the unified normal state feature vector across time periods in that dimension is statistically analyzed, i.e., the first feature value distribution range (from minimum to maximum value); then, the fluctuation range of the characteristic values ​​of the unified fault state feature vector across time periods in that dimension is statistically analyzed, i.e., the second feature value distribution range. The difference region between the two ranges is calculated: if the two ranges do not overlap, the difference region is the sum of the two ranges; if the two ranges overlap, the difference region is the sum of the remaining regions after subtracting the overlapping portion from each range. When calculating the proportion of the difference region, firstly, the interval lengths of the first and second feature value distribution ranges are calculated separately, and the average of the two is taken as the baseline interval length; then, the interval length of the difference region is calculated, and the difference region interval length is divided by the baseline interval length to obtain the proportion of the difference region. This proportion reflects the degree of difference between the distribution of normal and fault feature values ​​under that target feature dimension; the higher the proportion, the stronger the effectiveness of distinguishing between normal and fault states through that dimension.

[0052] Step S128: Integrate the difference region, feature difference degree, difference region ratio and the identification information of the target feature dimension in the ECU operating data corresponding to each target feature dimension to generate a target fault feature set.

[0053] For each target feature dimension, first clarify its identification information in the ECU operating data. This identification information must clearly correspond to the specific dimension and feature type of the ECU operating data. For example, the identification information for the "intake pressure fluctuation feature dimension" is "ECU intake pressure data - fluctuation frequency feature", the identification information for the "common rail pressure peak deviation feature dimension" is "ECU common rail pressure data - peak deviation feature", and the identification information for the "propeller speed response delay feature dimension" is "ECU propeller speed data - response delay feature". This ensures that each target feature dimension can be directly associated with the specific dimension of the ECU operating data.

[0054] Next, the difference region, feature difference degree, and difference region proportion corresponding to each target feature dimension are bound to the above-mentioned identification information: taking the "intake pressure fluctuation feature dimension" as an example, its difference region is "the sum of the non-overlapping regions of the normal feature value distribution range and the fault feature value distribution range", the feature difference degree is "the sum of the absolute differences between the unified normal and fault feature vectors across time periods in this dimension", and the difference region proportion is "the ratio of the difference region interval length to the baseline interval length". These three parameters are integrated with the "ECU intake pressure data - fluctuation frequency feature" identification information into a complete feature record.

[0055] All target feature dimensions are processed in the same way to form multiple feature records. These records are then categorized according to the dimensions of the ECU operating data (e.g., "intake pressure related feature records," "common rail pressure related feature records," and "propeller speed related feature records"), and finally integrated to form a target fault feature set. Each record in this set contains "ECU data dimension identifier + difference region + feature difference degree + difference region percentage," and corresponds one-to-one with each dimension of the ECU operating data, clearly reflecting the key feature parameters that distinguish normal and fault states under each dimension.

[0056] Step S130: Based on the target fault feature set and the data acquisition time series, construct a time series dataset of the aircraft engine exhaust valve control system, call the time series prediction model that integrates multi-scale attention to perform time series modeling processing on the time series dataset, and generate multi-scale time series prediction results.

[0057] Step S131: Extract feature dimensions related to the aircraft engine exhaust valve control system from the target fault feature set. The feature dimensions include intake pressure feature dimension, intake temperature feature dimension, propeller speed feature dimension, and common rail pressure feature dimension. At the same time, calculate the association weight between each feature dimension related to the aircraft engine exhaust valve control system and the exhaust valve control logic.

[0058] The core control logic of the aircraft engine exhaust valve control system is to "change the engine exhaust flow rate by adjusting the exhaust valve opening, thereby affecting the intake pressure, common rail pressure, and propeller speed. At the same time, the intake temperature indirectly affects the exhaust valve regulation effect by changing the gas density." When calculating the correlation weight based on this logic, two factors need to be considered: first, the direct causal relationship between the feature dimension and the exhaust valve control logic; and second, the correlation frequency between the feature dimension and the abnormal alarm of the exhaust valve in historical fault data.

[0059] For example, the intake pressure characteristic dimension directly reflects the change in exhaust flow after the exhaust valve opening is adjusted (exhaust valve opening increases → exhaust flow increases → intake pressure decreases), and in historical exhaust valve jamming alarm data, intake pressure abnormalities occur most frequently, therefore its correlation weight is the highest; the common rail pressure characteristic dimension is indirectly affected by exhaust flow (exhaust flow changes → fuel injection pressure adjustment → common rail pressure changes), with the second highest correlation frequency and correlation weight; the propeller speed characteristic dimension is linked to exhaust flow (exhaust flow changes → engine power adjustment → propeller speed changes), with a moderate correlation frequency and correlation weight; the intake temperature characteristic dimension only indirectly affects the exhaust valve adjustment effect through gas density (increased intake temperature → decreased gas density → decreased exhaust valve adjustment sensitivity), with the lowest correlation frequency and correlation weight.

[0060] Step S132: Based on the correlation weight, filter the feature dimensions related to the aircraft engine exhaust valve control system, retain the feature dimensions with a correlation weight greater than the weight threshold as candidate feature dimensions, and arrange the ECU operating data corresponding to the candidate feature dimensions in the order of the data acquisition time series to form a single-dimensional time series data sequence for each candidate feature dimension.

[0061] When setting weight thresholds, the feature importance requirements of the exhaust valve control system should be considered—ensuring that all retained candidate feature dimensions can effectively characterize the abnormal state of the exhaust valve. During the screening process, if the correlation weight of a feature dimension is lower than the weight threshold (e.g., the intake temperature feature dimension may be lower than the threshold due to its lowest correlation weight), it should be removed; feature dimensions with correlation weights greater than the threshold (e.g., intake pressure feature dimension, common rail pressure feature dimension, propeller speed feature dimension) should be retained as candidate feature dimensions.

[0062] For each candidate feature dimension, the corresponding ECU operating data is extracted (e.g., the fluctuation frequency calculation result of the ECU intake pressure data corresponding to the intake pressure feature dimension, and the peak deviation calculation result of the ECU common rail pressure data corresponding to the common rail pressure feature dimension). These data are then arranged according to the chronological order of the data acquisition time series (from the first time node of the earliest operating day to the last time node of the latest operating day) to form a single-dimensional time-series data sequence. In each sequence, each time node corresponds to a feature value, and the operating period to which that time node belongs is labeled (e.g., "a certain operating day - takeoff period").

[0063] Step S133: Align and integrate all single-dimensional time series data sequences according to the data acquisition time nodes, so that each data acquisition time node corresponds to a set of time series data for each candidate feature dimension. At the same time, add the identifier of the runtime segment to which each set of time series data belongs to form a time series dataset of the aircraft engine exhaust valve control system.

[0064] Based on the timestamp of the data acquisition time series, the single-dimensional time series data sequences of all candidate feature dimensions are aligned: for example, if the timestamp "at a certain time T1" corresponds to the feature value "F1" in the intake pressure time series, the feature value "P1" in the common rail pressure time series, and the feature value "R1" in the propeller speed time series, then "F1, P1, R1" are integrated into a set of time series data corresponding to timestamp T1.

[0065] Simultaneously, each data set is labeled with its corresponding operating period identifier—based on the operating phase (takeoff, cruise, landing) corresponding to the timestamp, information such as "Operating Day - Takeoff Period" or "Operating Day - Cruise Period" is added. The integrated time-series dataset contains "timestamp + operating period identifier + intake pressure characteristic value + common rail pressure characteristic value + propeller speed characteristic value" for each data entry, fully reflecting the changing patterns of relevant characteristics of the exhaust valve control system over time, and distinguishing the characteristic differences between different operating periods.

[0066] Step S134: Divide the time series dataset of the aircraft engine exhaust valve control system into a training dataset, a validation dataset, and a test dataset. The training dataset is used for parameter training of the time series prediction model that integrates multi-scale attention. The validation dataset is used for hyperparameter adjustment of the time series prediction model that integrates multi-scale attention. The test dataset is used for final prediction accuracy verification of the time series prediction model that integrates multi-scale attention.

[0067] The partitioning process must follow the "time-order partitioning" principle to avoid data leakage (i.e., the training data does not contain future data, ensuring the model's generalization ability). Specifically, the partitioning method is as follows: select data from the first majority of time points in the time-series dataset as the training dataset (covering complete periods across multiple consecutive operating days, including takeoff, cruise, and landing phases) for iterative updates of model parameters; select data from the middle portion of time points as the validation dataset (covering all periods of a complete operating day) for performance verification during hyperparameter tuning; and select data from the latter portion of time points as the test dataset (covering all periods of a complete operating day) for final accuracy verification after model training.

[0068] When dividing the datasets, it is necessary to ensure that the runtime distribution of the three datasets is consistent. For example, the proportion of data during takeoff, cruise, and landing in the training dataset should be the same as the proportion in the validation and test datasets to avoid bias in the model's prediction performance for a certain time period due to uneven distribution of time periods.

[0069] Step S135: Input the training dataset into the time series prediction model that integrates multi-scale attention. The encoder module of the time series prediction model that integrates multi-scale attention contains multiple attention layers at different time scales, which perform attention modeling on time series data of various scale periods and extract time series dependencies at different scales.

[0070] The encoder module of the time-series prediction model that integrates multi-scale attention contains three attention layers at different time scales, corresponding to "short-scale period", "medium-scale period" and "long-scale period", respectively. The functions and processing logic of each layer are as follows: Short-scale attention layer: corresponding to the scale of "continuous time nodes within a single time period" (such as multiple consecutive time nodes within a certain takeoff period). By calculating the attention weight of the time series data of adjacent time nodes at this scale, it focuses on the temporal dependencies within a short period of time - for example, capturing the fluctuation pattern of intake pressure in a short period of time after the exhaust valve is adjusted (such as the downward trend of intake pressure in multiple consecutive time nodes after the exhaust valve opening is adjusted).

[0071] Mesoscale attention layer: corresponding to the "single-day cross-time period" scale (such as the takeoff-cruise-landing period of a certain operating day). By calculating the attention weight of time series data in different time periods under this scale, the temporal dependencies of different operating stages of a single day are explored. For example, the correlation between the change of intake pressure during takeoff and the change of common rail pressure during cruise is captured (higher intake pressure during takeoff → higher common rail pressure during cruise).

[0072] Long-scale attention layer: corresponding to the "multi-day spanning time period" scale (such as the same time period of multiple operating days, such as the cruise period of multiple consecutive operating days). By calculating the attention weight of the time series data of the same time period of multiple days under this scale, long-term common time series patterns are extracted - for example, capturing the stable variation range of propeller speed during the cruise period of multiple consecutive operating days, and the long-term correlation with the exhaust valve adjustment.

[0073] After the training dataset is input into the encoder module, the temporal data of each candidate feature dimension is first converted into a high-dimensional embedding vector through the embedding layer (eliminating the difference in the scale of data of different dimensions and enhancing the feature representation ability); then the embedding vector is input into three attention layers respectively, and each layer outputs a temporal feature map of the corresponding scale (containing temporal dependency information at that scale).

[0074] Step S136: The features output by the attention layers at each time scale are weighted and fused through the fusion layer of the time series prediction model that integrates multi-scale attention. The learning rate, number of iterations, batch size and scale parameters of each attention layer of the time series prediction model that integrates multi-scale attention are adjusted by using a hyperparameter optimization method, so that the prediction error of the time series prediction model that integrates multi-scale attention on the training dataset is gradually reduced.

[0075] The fusion layer adopts an "attention-weighted fusion" strategy: first, the importance weights of the output feature maps of the three attention layers are adaptively learned through model training (e.g., in the task of predicting anomalies in exhaust valves, short-scale feature maps are more important for representing short-term anomalies and have higher weights); then, each feature map is multiplied by its corresponding weight and concatenated to form a global temporal feature map that integrates multi-scale features—this feature map contains temporal dependency information at short, medium, and long scales, which can comprehensively reflect the temporal change law of the exhaust valve control system.

[0076] The hyperparameter optimization process uses the Bayesian optimization method. The hyperparameters to be optimized include: learning rate (controlling the step size of model parameter updates, affecting the training convergence speed), number of iterations (controlling the total number of training rounds to avoid overfitting or underfitting), batch size (controlling the number of samples in each training session, affecting training efficiency and stability), and scale parameters of each attention layer (such as the number of time nodes in a single time period for short-scale attention layers, the number of time periods in a single day for medium-scale attention layers, and the number of time periods in multiple days for long-scale attention layers).

[0077] During optimization, a probabilistic model is first established based on the tried hyperparameter combinations and their corresponding prediction errors (such as the mean squared error between the model's prediction of time series data and the actual data in the training dataset). Then, the expected error of the untried hyperparameter combinations is predicted using this probabilistic model, and the combination with the smallest expected error is selected for verification. This combination is then substituted into the model for training, and the actual prediction error after training is recorded and fed back to the probabilistic model to update the parameters. The above process is repeated until the model's prediction error on the training dataset is reduced to a set range.

[0078] Step S137: Input the validation dataset into the adjusted time-series prediction model that integrates multi-scale attention, obtain the validation prediction results output by the adjusted time-series prediction model that integrates multi-scale attention, calculate the error value between the validation prediction results and the actual data in the validation dataset, and if the error value is less than the set error threshold, then use the test dataset to perform the final validation of the time-series prediction model that integrates multi-scale attention.

[0079] The validation dataset is input into the model after the hyperparameters are adjusted. The model extracts multi-scale temporal features of the validation data through the encoder module, and then the global temporal feature map is converted into validation prediction results for each candidate feature dimension through the decoder module (such as the intake pressure feature value and common rail pressure feature value of each time node in the validation period).

[0080] When calculating the error value, the mean squared error (MSE) metric is used—the predicted value at each time point is substituted with the actual value of the validation dataset to obtain the overall validation error value. If this error value is less than a set error threshold (set based on the accuracy requirements of the time-series prediction of the exhaust valve control system to ensure that the model still has good predictive performance on unseen validation data), then the final validation stage begins: the test dataset is input into the model, the test prediction results are obtained, the MSE value is calculated again, and the model's generalization ability on new test data is evaluated.

[0081] Step S138: If the prediction error of the test dataset is also less than the set error threshold, the trained time series prediction model is output, and the trained time series prediction model is called to perform time series modeling processing on the time series dataset of the aircraft engine exhaust valve control system, and output multi-scale time series prediction results containing the data change trends of each candidate feature dimension of various time periods.

[0082] If the test error is less than the set error threshold, it indicates that the model training is complete and the performance meets the standard, and the completed time series prediction model is output. When using this model to perform time series modeling on the complete time series dataset (including training, validation, and test datasets) of an aircraft engine exhaust valve control system, the model processing flow is as follows: The encoder module has three attention layers that perform attention modeling on short, medium, and long-scale temporal data of the complete dataset, respectively, and extract temporal dependencies at different scales. The fusion layer performs weighted fusion of the output features of each layer to form a global temporal feature map. The decoder module converts the global feature map into a prediction sequence for each candidate feature dimension, and finally outputs multi-scale temporal prediction results.

[0083] The results include three types of trends: short-scale trends (such as the intake pressure fluctuation trend at multiple consecutive time points in the future, which can be used to predict whether the exhaust valve may jam in the short term), medium-scale trends (such as the common rail pressure change trend across time periods of a future operating day, which can be used to predict the regulation effect of the exhaust valve at different operating stages of a single day), and long-scale trends (such as the propeller speed response trend at the same time period of multiple future operating days, which can be used to predict the stability of the exhaust valve control system in the long term), and each trend corresponds to specific prediction data for each candidate feature dimension.

[0084] Step S140: Based on the multi-scale time-series prediction results, establish dynamic standard reference curves for key parameters of the aircraft engine exhaust valve control system. The dynamic standard reference curves correspond to the key parameters of each dimension of the ECU operating data.

[0085] Step S141: Extract the prediction data sequences of each candidate feature dimension for each time period of various scales from the multi-scale time series prediction results. Each candidate feature dimension corresponds to multiple sets of prediction data sequences for different time periods.

[0086] For each candidate feature dimension (such as inlet pressure feature dimension, common rail pressure feature dimension, and propeller speed feature dimension), predictive data sequences for three scale periods are extracted from the multi-scale time series prediction results: Short-scale forecast data sequences: corresponding to the scale of "continuous time nodes within a single time period", such as "predicted intake pressure values ​​for multiple consecutive time nodes within a future takeoff period" or "predicted common rail pressure values ​​for multiple consecutive time nodes within a future cruise period". Each sequence contains the predicted values ​​for each time node within that time period and the corresponding timestamp.

[0087] Mesoscale forecast data sequence: corresponding to the "single-day spanning time period" scale, such as "the propeller speed forecast sequence during the takeoff-cruise-landing period of a future operating day", which includes the forecast values ​​for each time period of that operating day and the corresponding time period identifier.

[0088] Long-scale forecast data series: corresponding to the "multi-day spanning time period" scale, such as "intake pressure forecast series for cruise periods of multiple future operating days" and "common rail pressure forecast series for takeoff periods of multiple future operating days", including forecast values ​​for the same time period of each operating day and the corresponding operating day identifier.

[0089] Each candidate feature dimension corresponds to multiple sets of prediction data sequences at different scales and time periods, fully covering the prediction of feature changes in different time ranges in the future.

[0090] Step S142: Perform trend fusion processing on multiple sets of predicted data sequences for each candidate feature dimension, calculate the weighted average of multiple sets of predicted data at the same data collection time node, and obtain the fused predicted data sequence for each candidate feature dimension. The weighting weight is determined according to the error of the prediction results in each time period. The smaller the error, the larger the weighting weight.

[0091] Taking the intake pressure characteristic dimension as an example, its multiple sets of predicted data sequences include "short-scale - predicted sequence for a future takeoff time", "medium-scale - predicted sequence for a future operating day across time periods", and "long-scale - predicted sequence for cruise time periods across multiple future operating days". For the same data collection time node (such as "time node T2 for takeoff time on a future operating day"), this node may simultaneously contain both short-scale and meso-scale predicted data (the short-scale sequence contains node T2 for this takeoff time, and the meso-scale sequence also contains node T2 for this takeoff time).

[0092] When calculating the weighted average, first obtain the error value corresponding to each prediction data (e.g., the error value for short-scale prediction of this node is "error A", and the error value for mesoscale prediction of this node is "error B"); then calculate the weight based on the error value—the weight is inversely proportional to the error value (the smaller the error, the larger the weight), that is, "weight A = 1 / error A, weight B = 1 / error B", and normalize the weights (so that weight A + weight B = 1); finally, calculate the weighted average of this node by "(short-scale prediction value × weight A) + (mesoscale prediction value × weight B)".

[0093] All time points are processed in the same way, and all predicted data sequences for each candidate feature dimension are merged into a fused predicted data sequence. This sequence can more accurately reflect the true changing trend of the feature dimension (it combines the advantages of prediction at different scales and reduces the error of prediction at a single scale).

[0094] Step S143: Perform trend analysis processing on the fusion prediction data sequence for each candidate feature dimension to identify stable trend segments, fluctuating trend segments, and trend transition nodes in the fusion prediction data sequence that change with the data acquisition time. The trend transition node is the boundary point between the stable trend segment and the fluctuating trend segment.

[0095] Trend analysis uses the "sliding window method," and the specific process is as follows: Set a fixed-length sliding window (the window size is determined based on the data acquisition interval and the frequency of feature changes to ensure effective capture of trend changes), slide the window along the fused prediction data sequence, and calculate the variance of the data within each window (reflecting the degree of data fluctuation). If the variance of the data within the window is less than the stable variance threshold, and multiple consecutive windows meet this condition, then the time range covered by the above window is determined to be a stable trend segment—the data in this segment changes smoothly over time without significant fluctuations (e.g., the fused prediction value of intake pressure during the cruise period on a certain operating day, where the variance of multiple consecutive windows is less than the threshold, is determined to be a stable trend segment).

[0096] If the variance of the data within the window is greater than the fluctuation variance threshold, and the data exhibits periodic or random fluctuation characteristics (such as the common rail pressure fusion prediction value during the landing period on a certain operating day, which shows periodic fluctuations with the landing process, and the variance of multiple windows is greater than the threshold), then the time node range covered by the above window is determined as the fluctuation trend segment—the data in this segment changes significantly over time, and the fluctuation pattern needs to be closely monitored to match the dynamic needs of exhaust valve regulation.

[0097] Trend transition nodes are the dividing points between stable and fluctuating trend segments. The determination logic is as follows: when the variance calculated by the sliding window abruptly changes from being less than the stable variance threshold to being greater than the fluctuating variance threshold, the starting time node corresponding to that window is the trend transition node; conversely, when the variance changes from being greater than the fluctuating variance threshold to being less than the stable variance threshold, the starting time node corresponding to that window is also a trend transition node. For example, in the intake pressure fusion prediction sequence during the cruise period of a certain operating day, if the variance of multiple windows in the first half is less than the stable threshold (stable trend segment), and the variance of a certain window in the second half abruptly changes to be greater than the fluctuating threshold (fluctuating trend segment), then the starting time node of that window is the trend transition node. It is necessary to record the feature value and timestamp at this node to ensure that the subsequent baseline curve is smoothly connected at this node.

[0098] Step S144: For a stable trend segment, calculate the mean and standard deviation of the fused prediction data within the stable trend segment. Use the mean as the benchmark value of the stable trend segment and the standard deviation as the benchmark fluctuation range of the stable trend segment. At the same time, record the start and end times of the stable trend segment.

[0099] Taking a stable trend segment of the intake pressure characteristic dimension (such as "the first half of the cruise period on a future operating day") as an example, first iterate through all the fused prediction data values ​​in the trend segment, calculate the sum of these data values, and then divide by the number of data values ​​to obtain the average value of the stable trend segment. This average value is used as the benchmark value—reflecting the normal reference level of intake pressure in this period, and matching the expected value of intake pressure under the stable adjustment state of the exhaust valve.

[0100] Next, calculate the standard deviation of this stable trend segment: first calculate the difference between each data value and the average value, then square each difference, calculate the average of all squared values, and finally take the square root of the average value to obtain the standard deviation. Use this standard deviation as the benchmark fluctuation range—reflecting the allowable fluctuation range of intake pressure under normal conditions during this period. Exceeding this range may indicate abnormal exhaust valve regulation.

[0101] At the same time, the start time (corresponding to a certain timestamp of the data collection time series) and end time (corresponding to another timestamp of the data collection time series) of the stable trend segment are recorded to clarify the effective time interval of the benchmark value and the benchmark fluctuation range, so as to ensure that the corresponding benchmark parameters can be located when comparing real-time data in the future.

[0102] Step S145: For the fluctuation trend segment, calculate the moving average and moving standard deviation of the fused prediction data within the fluctuation trend segment. The size of the moving window is dynamically adjusted according to the fluctuation frequency. A dynamic benchmark value and dynamic fluctuation range that change over time are generated according to the data collection time interval.

[0103] For fluctuation trend segments (such as "intake pressure fusion prediction sequence during the landing period of a future operating day"), since the data fluctuates significantly over time, the sliding window method is required to calculate the dynamic baseline value and dynamic fluctuation range: First, determine the initial sliding window size based on the fluctuation frequency (a smaller window for high fluctuation frequency to ensure high-frequency fluctuations are captured; a larger window for low fluctuation frequency to avoid frequent changes in the baseline value). Then, slide the sliding window along the fusion prediction data sequence of the fluctuation trend segment. Each time the window is slid, calculate the average value (i.e., the sliding average, which serves as the dynamic baseline value for the corresponding time node of the window) and the standard deviation (i.e., the sliding standard deviation, which serves as the dynamic fluctuation range for the corresponding time node of the window).

[0104] For example, if the data in the fluctuation trend segment exhibits periodic fluctuations, the sliding window size is set to the number of time points contained in one fluctuation cycle, ensuring that each window can cover a complete fluctuation cycle. The calculated moving average reflects the average level within that cycle, and the moving standard deviation reflects the degree of fluctuation within that cycle. As the window slides, each time point corresponds to a set of dynamic benchmark values ​​and dynamic fluctuation ranges, forming a time-varying sequence of benchmark parameters that matches the dynamic adjustment characteristics of the exhaust valve within the fluctuation trend segment.

[0105] Step S146: For trend transition nodes, calculate the rate of change of data before and after the trend transition node, determine the benchmark transition value at the trend transition node, and smoothly connect the benchmark value of the stable trend segment and the dynamic benchmark value of the fluctuating trend segment at the trend transition node.

[0106] Taking the trend transition node from a stable trend segment to a fluctuating trend segment as an example, first extract the last data value of the stable trend segment before the node (denoted as "previous value") and the first data value of the fluctuating trend segment after the node (denoted as "post value"), and calculate the rate of change: Rate of change = (post value - previous value) / (timestamp corresponding to the post value - timestamp corresponding to the previous value). This rate of change reflects the speed of data change during the trend transition.

[0107] The baseline transition value is determined based on the rate of change: if the rate of change is small (data transition is smooth), the baseline transition value is the average of the previous and subsequent values; if the rate of change is large (data transition is abrupt), multiple transition time nodes are added before and after the node, and the baseline value is gradually adjusted according to the rate of change, so that the baseline value of the stable trend segment gradually transitions to the first dynamic baseline value of the fluctuating trend segment through the transition value, avoiding the baseline value from abruptly changing at the node, which would lead to misjudgment in subsequent real-time data comparison.

[0108] For example, the benchmark value of the stable trend segment is "a certain value A", and the first dynamic benchmark value of the fluctuating trend segment is "a certain value B". When the rate of change is large, two transition nodes are added before and after the trend transition node, with transition values ​​of "(A+B)×0.75" and "(A+B)×0.5" respectively, so that the benchmark value smoothly decreases from A to B through the two transition values, ensuring that the benchmark curve has no obvious discontinuity at the node.

[0109] Step S147: Associate the stable trend segment baseline value, stable trend segment baseline fluctuation range, fluctuation trend segment dynamic baseline value, fluctuation trend segment dynamic fluctuation range, trend transition node and baseline transition value of each candidate feature dimension with the corresponding data collection time series to form the dynamic baseline data sequence of that candidate feature dimension.

[0110] For each candidate feature dimension (such as intake pressure feature dimension and common rail pressure feature dimension), the parameters obtained in steps S144 to S146 are arranged in the order of timestamps of the data acquisition time series: the timestamp of the stable trend segment is associated with the corresponding benchmark value and benchmark fluctuation range; the timestamp of the fluctuating trend segment is associated with the corresponding moving average (dynamic benchmark value) and moving standard deviation (dynamic fluctuation range); the timestamp of the trend transition node is associated with the corresponding benchmark transition value.

[0111] For example, in the dynamic baseline data sequence of the intake pressure characteristic dimension, the timestamp "T3-T4" (stable trend segment) is associated with "baseline value A, baseline fluctuation range R1"; the timestamp "T4-T5" (trend transition node interval) is associated with "transition value A1, transition value A2"; the timestamp "T5-T6" (fluctuation trend segment) is associated with "dynamic baseline value B1 (T5), dynamic fluctuation range R2 (T5)" and "dynamic baseline value B2 (T6), dynamic fluctuation range R3 (T6)", etc., forming a complete dynamic baseline data sequence, with each timestamp corresponding to a specific baseline parameter.

[0112] Step S148: Connect the dynamic benchmark data sequence according to the chronological order of the data acquisition time sequence to generate the dynamic standard benchmark curve corresponding to the candidate feature dimension. The horizontal axis of the dynamic standard benchmark curve is the data acquisition time, and the vertical axis is the benchmark data value, benchmark fluctuation range and trend transformation node identifier of the candidate feature dimension.

[0113] Using the data collection time as the horizontal axis (arranged in timestamp order) and the baseline parameters of the candidate feature dimensions as the vertical axis, the parameters in the dynamic baseline data sequence are plotted as curves: the stable trend segment plots the "baseline value straight line" and the "baseline fluctuation range upper and lower limit straight line" (forming a band-shaped area representing the normal fluctuation range); the fluctuation trend segment plots the "dynamic baseline value curve" and the "dynamic fluctuation range upper and lower limit curve" (forming a band-shaped area that changes over time); the trend transition nodes are marked with "transition node identifier" (such as a special symbol) and the corresponding baseline transition value.

[0114] For example, in the dynamic standard benchmark curve of the common rail pressure characteristic dimension, the horizontal axis is the timestamp of "08:00-12:00 on a future operating day", and the vertical axis is the common rail pressure value; from 08:00 to 09:30 (stable trend segment), the "benchmark value line C" and the "upper and lower limits of the fluctuation range C1, C2" are drawn; from 09:30 to 09:40 (trend transition node interval), the "transition node" and transition values ​​C3, C4 are marked; from 09:40 to 12:00 (fluctuation trend segment), the "dynamic benchmark value curve D" and the "upper and lower limits of the dynamic fluctuation range D1, D2" are drawn, forming a complete dynamic standard benchmark curve, which intuitively reflects the normal benchmark range of common rail pressure in different time periods.

[0115] Step S149: Set dynamic update trigger conditions. When the amount of newly added aircraft engine operation data reaches the set threshold or when an unincluded trend type appears in the newly added data, the dynamic standard baseline curve update process is triggered. Repeat the above steps to recalculate the baseline value and fluctuation range to achieve dynamic adjustment of the dynamic standard baseline curve.

[0116] Step S1491: Define the data volume triggering condition and trend type triggering condition for dynamic updates.

[0117] Data volume trigger condition: Set a threshold for the amount of new data. This threshold is determined based on the historical data collection frequency and the number of candidate feature dimensions—ensuring that the amount of new data is sufficient to reflect changes in operational status and avoiding frequent updates. For example, if the data collection interval is a fixed duration, the threshold for the amount of new data is set to cover the total amount of data for a complete operational day. That is, when the amount of new data reaches the amount for one operational day, an update is triggered.

[0118] Trend type triggering conditions: Establish a trend type library, which includes common trend types such as stable trend, linear fluctuation trend, periodic fluctuation trend, and random fluctuation trend. Each trend type corresponds to specific characteristic parameters (such as the variance range of a stable trend and the period length of a periodic fluctuation trend). When the trend characteristics of newly added data differ from the characteristic parameters of all types in the trend type library by more than the trend difference threshold, it is determined that an unincluded trend type has appeared, and an update is triggered.

[0119] Step S1492: Based on the data volume triggering condition, count the amount of aircraft engine ECU operating data added since the last dynamic standard baseline curve update. When the data volume reaches the threshold of the newly added data volume, trigger the data volume update signal.

[0120] The data acquisition system counts the number of newly added ECU operating data in real time (counted by timestamp, with each timestamp corresponding to a complete data record), and compares it with the threshold for the amount of newly added data. If the count reaches the threshold, a data update signal is immediately triggered, and the trigger time and the time range of the newly added data are recorded (e.g., "operating data from the last update to the current operating day").

[0121] Step S1493: For the trend type triggering condition, identify the time series trend of the newly added data, extract the trend feature parameters of the newly added data and compare them with the feature parameters in the trend type library. If the difference is greater than the trend difference threshold, then trigger the trend type update signal.

[0122] The sliding window method is used to extract trend feature parameters (such as variance, fluctuation period, rate of change, etc.) of the newly added data, and compare them with the feature parameter ranges of each type in the trend type library. For example, if the variance of the newly added data is greater than the upper limit of the variance of the stable trend, and the fluctuation period does not match the period range of the periodic fluctuation trend, and the rate of change does not meet the rate of change range of the linear fluctuation trend, then it is determined that the difference with all existing trend types is greater than the trend difference threshold, triggering the trend type update signal, and recording the trend feature parameters and differences of the newly added data.

[0123] Step S1494: Monitor data volume update signals and trend type update signals in real time. When any update signal is detected, start the update process of the dynamic standard baseline curve, recalculate the baseline parameters and adjust the curve.

[0124] After the update process is initiated, the new data is integrated with the original time-series dataset to form an updated time-series dataset. Steps S130 to S148 are repeated to regenerate the multi-scale time-series prediction results based on the updated dataset, and new stable trend segment baseline values, fluctuating trend segment dynamic baseline values, and trend transition node parameters are calculated. The new parameters are compared with the original parameters, the differences are replaced, and an updated dynamic standard baseline curve is generated to ensure that the curve can adapt to changes in engine operating conditions (such as baseline value shifts caused by decreased exhaust valve adjustment sensitivity after long-term use).

[0125] Step S1410: Perform the above-mentioned dynamic benchmark data sequence construction and dynamic standard benchmark curve generation operations on all candidate feature dimensions to obtain dynamic standard benchmark curves corresponding to the key parameters of each dimension of ECU operating data.

[0126] For the three candidate feature dimensions—intake pressure, common rail pressure, and propeller speed—steps S141 to S149 are repeated to generate dynamic standard reference curves for each dimension: the intake pressure curve reflects the normal reference range of intake pressure at different times, the common rail pressure curve reflects the normal reference range of common rail pressure at different times, and the propeller speed curve reflects the normal reference range of propeller speed at different times. These three curves correspond to three key parameters of the ECU operating data.

[0127] Step S150: The real-time ECU operating data collected by the aircraft engine is compared and analyzed in real time with the dynamic standard reference curve. The abnormal information of the real-time ECU operating data deviating from the dynamic standard reference curve is identified by combining the correlation of the target fault characteristics. The aircraft engine operating status detection result is generated, which includes the abnormal parameter dimension identifier and the abnormality degree information.

[0128] Step S151: Acquire real-time ECU operating data collected by the aircraft engine. The real-time ECU operating data includes real-time intake pressure data, real-time intake temperature data, real-time propeller speed data, real-time common rail pressure data, and real-time fault alarm data. The real-time ECU operating data collected by the aircraft engine has a real-time data acquisition timestamp and operating condition identifier.

[0129] The ECU operating data is collected in real time through the aircraft engine's sensor network: the intake pressure sensor collects real-time intake pressure data, the intake temperature sensor collects real-time intake temperature data, the propeller speed sensor collects real-time propeller speed data, the common rail pressure sensor collects real-time common rail pressure data, and the ECU system records real-time fault alarm data (such as exhaust valve jamming alarms and sensor malfunction alarms). Each real-time data point is accompanied by a real-time data acquisition timestamp (accurate to the data acquisition interval) and an operating condition identifier (such as "takeoff condition," "cruise condition," and "landing condition") to ensure matching with the corresponding time period and operating condition of the dynamic standard reference curve.

[0130] Step S152: Retrieve the latest version of the dynamic standard reference curve from the storage location, and filter out the dynamic standard reference curve segments that match the current operating conditions based on the operating condition identifier of the ECU operating data collected in real time from the aircraft engine.

[0131] The dynamic standard reference curve is divided into different segments according to the operating conditions (takeoff, cruise, and landing). For example, the intake pressure dynamic standard reference curve includes "takeoff operating condition segment", "cruise operating condition segment", and "landing operating condition segment". When the operating condition of the real-time ECU operating data is identified as "cruise operating condition", the latest version of the dynamic standard reference curve for each candidate feature dimension is retrieved from the storage location, and the "cruise operating condition segment" of each curve is selected as the reference segment for the current real-time data comparison, ensuring that the reference parameters match the current operating condition (e.g., the intake pressure reference value of the cruise operating condition is lower than that of the takeoff operating condition, avoiding comparison errors caused by differences in operating conditions).

[0132] Step S153: For each parameter dimension, extract the value of the real-time ECU operating data, the real-time data acquisition timestamp, and the operating condition identifier corresponding to that parameter dimension. On the dynamic standard benchmark curve segment matched by that parameter dimension, find the benchmark data value, benchmark fluctuation range, and trend type that are consistent with the real-time data acquisition timestamp.

[0133] Taking real-time intake pressure data as an example, extract its value, real-time timestamp (e.g., "time T7"), and "cruise condition" identifier; in the "cruise condition segment" of the intake pressure dynamic standard reference curve, find the reference data value (e.g., "reference value E"), reference fluctuation range (e.g., "fluctuation range R4"), and trend type (e.g., "stable trend segment") that is consistent with T7; similarly, perform the same operation on real-time common rail pressure data and real-time propeller speed data to obtain the corresponding reference parameters and trend types respectively.

[0134] Step S154: Calculate the difference between the real-time data value of this parameter dimension and the corresponding benchmark data value, determine whether the difference is within the benchmark fluctuation range, and combine it with the feature difference degree in the target fault feature corresponding to this parameter dimension. If the difference exceeds the benchmark fluctuation range and the feature difference degree is greater than the set difference threshold, then it is determined that the real-time ECU operating data of this parameter dimension is abnormal.

[0135] Taking real-time common rail pressure data as an example, calculate the difference between the real-time value and the baseline data value (e.g., "real-time value F - baseline value G = difference H"); determine whether the difference H is within the baseline fluctuation range (e.g., "fluctuation range R5"): if the difference H is within R5, the real-time data of this parameter dimension is considered normal; if the difference H exceeds R5, further examine the feature difference degree of this parameter dimension in the target fault feature set (e.g., "feature difference degree of common rail pressure peak deviation feature dimension"); if the feature difference degree is greater than the set difference threshold (indicating that the abnormality of this parameter dimension is highly correlated with the fault state), then the real-time common rail pressure data is considered abnormal; if the feature difference degree is less than the set difference threshold (indicating that it may be a temporary fluctuation), then the abnormality is not considered for the time being, and subsequent data is continuously monitored.

[0136] Step S155: If an anomaly is determined, record the anomaly parameter identifier, real-time ECU operating data value, baseline data value, difference, real-time data acquisition timestamp, operating condition identifier, and corresponding feature difference degree for that parameter dimension, forming a single-dimensional anomaly record.

[0137] For example, when an anomaly is determined in the real-time intake pressure data, the anomaly parameter is recorded as “ECU intake pressure data - fluctuation frequency characteristics”, the real-time value is “a certain value I”, the baseline data value is “baseline value J”, the difference is “difference K”, the real-time timestamp is “T8”, the operating condition is “cruise condition”, and the characteristic difference degree is “a certain difference degree L”. The above information is integrated into a single-dimensional anomaly record. Each record contains the key basis for the anomaly determination, which is convenient for subsequent traceability and analysis.

[0138] Step S156: Analyze the correlation between the dimensions of each abnormal parameter. Based on the correlation weight of each dimension in the target fault feature set, calculate the correlation strength between the abnormal parameter dimensions. If the correlation strength is greater than the correlation threshold, it is determined that the abnormal parameter belongs to the related abnormality caused by the same abnormal event, and a related abnormality group is formed.

[0139] Step S1561: Extract the correlation weight matrix between each parameter dimension from the target fault feature set. The rows and columns of the correlation weight matrix correspond to different parameter dimensions. The element values ​​of the correlation weight matrix represent the correlation weight between two corresponding parameter dimensions. The correlation weight is calculated based on the functional dependency of the parameters in engine operation and historical abnormal correlation data.

[0140] The rows and columns of the correlation weight matrix are "intake pressure characteristic dimension", "common rail pressure characteristic dimension" and "propeller speed characteristic dimension", respectively. The matrix element values ​​are the correlation weights between each pair of dimensions: for example, the element value of "intake pressure-common rail pressure" has a high weight (because both are directly affected by the exhaust valve regulation, the functional dependence is strong, and the frequency of co-occurrence in historical anomalies is high), the element value of "intake pressure-propeller speed" has a medium weight (the functional dependence is relatively strong, and the frequency of co-occurrence of historical anomalies is moderate), and the element value of "common rail pressure-propeller speed" has a medium weight, forming a complete correlation weight matrix.

[0141] Step S1562: Extract all parameter dimension identifiers with anomalies from the single-dimensional anomaly records to form an anomaly parameter dimension list. Combine the anomaly parameter dimensions in the anomaly parameter dimension list in pairs to generate anomaly parameter combination pairs.

[0142] If a single-dimensional anomaly record contains both "intake pressure anomaly" and "common rail pressure anomaly," then the list of anomaly parameter dimensions is "[intake pressure characteristic dimension, common rail pressure characteristic dimension]." Each pair of anomaly parameter dimensions in this list is combined to generate a unique anomaly parameter combination pair, namely, "intake pressure characteristic dimension - common rail pressure characteristic dimension." If a single-dimensional anomaly record simultaneously contains "intake pressure anomaly," "common rail pressure anomaly," and "propeller speed anomaly," then the list of anomaly parameter dimensions is "[intake pressure characteristic dimension, common rail pressure characteristic dimension, propeller speed characteristic dimension]." Each pair of anomaly parameter dimensions is combined to generate three sets of anomaly parameter combination pairs: "intake pressure characteristic dimension - common rail pressure characteristic dimension," "intake pressure characteristic dimension - propeller speed characteristic dimension," and "common rail pressure characteristic dimension - propeller speed characteristic dimension." This ensures that all anomaly parameter dimensions can form bidirectional combinations without omissions or duplications.

[0143] Step S1563: For each abnormal parameter combination pair, find the matrix element value corresponding to the abnormal parameter combination pair in the association weight matrix, and use the matrix element value as the initial association strength.

[0144] Taking the "intake pressure feature dimension - common rail pressure feature dimension" combination as an example, the element position with row label "intake pressure feature dimension" and column label "common rail pressure feature dimension" is located in the association weight matrix, and the element value at that position (high weight value) is extracted. This value is used as the initial association strength of the combination. For the "intake pressure feature dimension - propeller speed feature dimension" combination, the corresponding element value (medium weight value) is also found in the matrix as its initial association strength. This process is repeated to ensure that each abnormal parameter combination can be matched with a unique initial association strength, and this strength directly reflects the basic association degree between the two in terms of functional dependence and historical abnormal association.

[0145] Step S1564: Extract the synchronization of changes in the abnormal parameter combination with the real-time ECU operating data to obtain the similarity of change trends.

[0146] Synchronization extraction needs to be based on the time series characteristics of real-time ECU operating data: taking the combination of "intake pressure characteristic dimension - common rail pressure characteristic dimension" as an example, extract the real-time data change curves of the two during abnormal periods, and analyze whether the change trends of the two curves are synchronized. For example, if the intake pressure shows a continuous downward trend during a certain period of time, and the common rail pressure also shows a continuous downward trend, and the starting time and rate of change of the decrease are similar, then the synchronicity of the two changes is high; if the common rail pressure does not change significantly or shows an upward trend when the intake pressure decreases, then the synchronicity is low.

[0147] The similarity of the change trends is obtained by calculating the similarity between two change curves (such as the overlap of curve shapes, the time difference of the inflection points, and the proportional relationship of the change amplitudes). The range of this similarity value is consistent with the initial association strength to ensure that the units are consistent during subsequent weighted fusion. For example, high synchronicity corresponds to a high similarity value, and low synchronicity corresponds to a low similarity value.

[0148] Step S1565: Weighted fusion of the initial association strength and the similarity of the change trend to obtain the final association strength of the abnormal parameter combination pair.

[0149] When performing weighted fusion, the weights of the two factors need to be set according to their importance in determining the correlation anomaly. Since the initial correlation strength is based on historical functional dependencies and abnormal data, reflecting a long-term stable correlation, its weight is set to a higher value. The trend similarity is based on real-time data, reflecting the dynamic correlation during the current abnormal period, and its weight is set to a lower value (the sum of the weights of the two factors is 1).

[0150] Taking the combination of "intake pressure feature dimension - common rail pressure feature dimension" as an example, if the initial correlation strength is a high weight value and the trend similarity is a high similarity value, the final correlation strength is calculated as "final correlation strength = initial correlation strength × high weight + trend similarity × low weight", resulting in a high final correlation strength. If the initial correlation strength of a combination is a medium weight value, but the trend similarity is a low similarity value, the final correlation strength will be lower than the initial correlation strength. This reflects the impact of real-time dynamic correlation on the final correlation determination and avoids misjudgment caused by relying solely on historical data.

[0151] Step S1566: If the final association strength of any abnormal parameter combination pair is greater than the association strength threshold, then mark the abnormal parameter combination pair as an abnormal association pair.

[0152] The correlation strength threshold is set based on the abnormal correlation judgment criteria of the exhaust valve control system to ensure that only combinations with significant correlation are selected. For example, if the final correlation strength (high value) of the "intake pressure characteristic dimension - common rail pressure characteristic dimension" combination is greater than the correlation strength threshold, it is marked as an abnormally correlated pair; if the final correlation strength (low value) of a combination is less than the threshold, it is not marked and is judged as a non-correlated abnormal pair. The marking process requires verification of all abnormal parameter combinations one by one to ensure no omissions, and the marking result directly reflects whether the combination is an abnormally correlated pair caused by the same abnormal event.

[0153] Step S1567: Cluster all associated anomaly pairs, grouping associated anomaly pairs containing the same anomaly parameter dimension into the same associated anomaly group, and calculate the average association strength within each associated anomaly group. If the average association strength within the associated anomaly group is greater than the association strength threshold, then the associated anomaly group is confirmed as a valid associated anomaly group, and the anomaly parameters within the valid associated anomaly group are determined to be associated anomalies caused by the same anomaly event.

[0154] The clustering process uses a "greedy clustering method": based on all associated anomaly pairs, first select any associated anomaly pair (such as "inlet pressure - common rail pressure") as the initial group, then find other associated anomaly pairs containing "inlet pressure" or "common rail pressure" (such as "common rail pressure - propeller speed"), and add them to the group to form an associated anomaly group containing "inlet pressure, common rail pressure, propeller speed".

[0155] When calculating the average correlation strength within a group, first, the final correlation strength of all correlated anomaly pairs within the group is calculated, summed, and then divided by the number of correlated anomaly pairs to obtain the average correlation strength within the group. If this average value is greater than the correlation strength threshold, it indicates that the overall correlation of the parameter dimensions within the group is significant, confirming the group as a valid correlated anomaly group, and determining that all abnormal parameters within the group (intake pressure, common rail pressure, propeller speed) are caused by the same abnormal event (such as abnormal exhaust flow caused by exhaust valve jamming); if the average value is less than the threshold, further processing is required.

[0156] Step S1568: If the average correlation strength within a group of association anomalies is less than the correlation strength threshold, then split the association anomaly group, remove the anomaly parameter dimension with the lowest correlation strength, and recalculate the average correlation strength within the split association anomaly group until the average correlation strength within the association anomaly group meets the requirements or only one anomaly parameter dimension remains in the association anomaly group.

[0157] Taking a group of correlation anomalies containing three parameter dimensions as an example, if the average correlation strength within the group is less than the threshold, first find the pair of correlation anomalies with the lowest final correlation strength (such as "inlet pressure - propeller speed"), and remove the parameter dimension with weak correlation in the pair of correlations (such as the propeller speed feature dimension). At this time, the group of correlation anomalies is left with two dimensions: "inlet pressure" and "common rail pressure".

[0158] Recalculate the average correlation strength of the split group (only including the final correlation strength of the "intake pressure-common rail pressure" combination pair). If the new average value is greater than the threshold, the group is confirmed as a valid correlation anomaly group. If it is still less than the threshold, continue to remove the parameter dimension with the lowest correlation strength until only one parameter dimension remains in the group (at which point no correlation anomaly pair can be formed, and it is judged as a single-dimensional independent anomaly). Ensure that the final correlation anomaly groups all meet the correlation strength requirements and avoid grouping anomalous parameters with no significant correlation into one group.

[0159] Step S1569: Record the dimension of the abnormal parameters for each effective association anomaly group, the association strength of each abnormal parameter pair within the effective association anomaly group, the average association strength within the effective association anomaly group, and the basis for determining association anomalies, and include them in the anomaly information summary table.

[0160] The recorded content must be complete and traceable: For example, if a valid correlation anomaly group includes "intake pressure feature dimension and common rail pressure feature dimension", the identifiers of these two anomaly parameter dimensions, the final correlation strength of the "intake pressure-common rail pressure" combination pair within the group, and the average correlation strength within the group (i.e., the correlation strength of the combination pair) must be recorded. At the same time, the judgment criteria must be recorded (such as "the initial correlation strength is high weight, the similarity of the change trend is high, the weighted fusion is greater than the correlation threshold, and the average strength within the group meets the standard").

[0161] Fill the above information into the anomaly information summary table in a unified format, and together with the single-dimensional anomaly records and alarm-related anomaly records, they will form a complete anomaly information system.

[0162] Step S157: If the real-time ECU operating data collected by the aircraft engine contains real-time fault alarm data, then extract the alarm type identifier and alarm trigger timestamp corresponding to the fault alarm data, and associate the alarm type identifier with the analysis results of the single-dimensional abnormal records and associated abnormal groups before and after the alarm trigger timestamp to form alarm associated abnormal records.

[0163] If the real-time fault alarm data includes an "exhaust valve jamming alarm", first extract the alarm type identifier ("exhaust valve jamming") and alarm trigger timestamp (e.g., "a certain time T3"). Then search for single-dimensional abnormal records within a preset time period before and after the timestamp (e.g., "intake pressure abnormality" and "common rail pressure abnormality" records from multiple time nodes before T3 to multiple time nodes after T3) and the results of related abnormal group analysis (e.g., valid related abnormal groups containing intake pressure and common rail pressure).

[0164] The alarm type identifier is bound to these anomaly records and correlation results: for example, labeling "Exhaust gas valve jamming alarm (T3) associated anomaly: intake pressure anomaly (1 time point before T3), common rail pressure anomaly (T3), the associated anomaly group is intake pressure - common rail pressure (average correlation strength within the group meets the standard)", thus forming an alarm correlation anomaly record. This alarm correlation anomaly record can clearly define the correspondence between fault alarms and specific anomaly parameters.

[0165] Step S158: Summarize the analysis results of all single-dimensional abnormal records, associated abnormal groups, and alarm-related abnormal records, and count the number of abnormal parameter dimensions, the number of associated abnormal groups, the difference between each abnormal parameter, the duration of the abnormality, and the number of alarm associations to obtain the summary information.

[0166] When summarizing and compiling statistics, the data is organized by category: the number of abnormal parameter dimensions (e.g., "3 dimensions are abnormal: intake pressure, common rail pressure, propeller speed"); the number of associated abnormal groups (e.g., "1 effective associated abnormal group"); the difference between each abnormal parameter (e.g., "abnormal intake pressure difference: the degree of deviation between real-time value and benchmark value, abnormal common rail pressure difference: the degree of deviation between real-time value and benchmark value"); the duration of abnormality (e.g., "abnormal intake pressure lasted from T2 to T4, with multiple time points"); and the number of alarm associations (e.g., "1 alarm (exhaust valve jamming) is associated with 2 abnormal parameter dimensions").

[0167] The above statistical results were organized into a structured summary to ensure that the data classification was clear, the numerical description was accurate, and that it could intuitively reflect the overall scale and key characteristics of the anomalies detected in this study.

[0168] Step S159: Integrate the above summary information, single-dimensional anomaly records, analysis results of associated anomaly groups, and alarm-related anomaly records to form an anomaly information summary table, and generate aircraft engine operating status detection results based on the anomaly information summary table.

[0169] The anomaly information summary table should include multiple fields: single-dimensional anomaly record fields (anomaly parameter dimension identifier, real-time value, baseline value, difference, anomaly start time, anomaly end time); associated anomaly group fields (anomaly parameter dimension within the group, associated anomaly pairs within the group, average association strength within the group, judgment result); alarm associated anomaly fields (alarm type, alarm timestamp, associated anomaly record / associated group); and summary statistics fields (total number of anomaly dimensions, total number of associated groups, number of alarm associations, main anomaly parameters).

[0170] The aircraft engine operating status detection results generated from this summary table must be presented in concise and clear technical language, explicitly describing the abnormalities: for example, "This inspection revealed an anomaly in the aircraft engine exhaust valve control system, involving three parameter dimensions: intake pressure, common rail pressure, and propeller speed. The intake pressure and common rail pressure are related anomalies caused by the same abnormal event (exhaust valve jamming), lasting from T2 to T4; the propeller speed is an independent anomaly with a small difference in value. The associated fault alarm is 'exhaust valve jamming' (triggered at T3). It is recommended to prioritize checking the exhaust valve opening adjustment function, followed by checking the propeller speed sensor status." The detection results must include anomaly parameter dimension identification, anomaly severity (difference size, duration), correlation analysis, alarm correlation, and handling suggestions to ensure that technicians can directly conduct fault diagnosis and maintenance based on the results.

[0171] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of an aircraft engine detection system 100 based on multi-task timing analysis, provided in an embodiment of this application, for performing the above-described inspection video stream processing method. The aircraft engine detection system 100 based on multi-task timing analysis may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0172] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the aircraft engine inspection system 100 based on multi-task timing analysis and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the aircraft engine inspection system 100 based on multi-task timing analysis and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may be integrated into the processor 130 and may communicate with external systems via the communication unit 110.

[0173] The processor 130 is the control center of the aircraft engine testing system 100 based on multi-task timing analysis. It connects various parts of the system via various interfaces and lines, and performs overall monitoring of the system by running or executing software programs and / or modules stored in the machine-readable storage medium 120 and calling data stored in the machine-readable storage medium 120. Optionally, the processor 130 may include one or more processing cores; for example, it may integrate an application processor and a modem processor, where the application processor primarily handles the operating system, user interface, and applications, and the modem processor primarily handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to implement the inspection video stream processing method provided in the aforementioned method embodiments.

[0174] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for detecting aircraft engines based on multi-task time series analysis, characterized in that, The method includes: Acquire a set of operational monitoring data generated by an aircraft engine during operation. The set of operational monitoring data includes ECU operational data and fault data, and the ECU operational data and fault data maintain a one-to-one correspondence with the data acquisition time series. The operation monitoring data set is subjected to cross-time period comparison representation learning processing. The difference features between normal operation status data and fault operation status data under different operating periods are extracted through the comparison representation learning model to generate a target fault feature set. The target fault feature set corresponds one-to-one with the information of each dimension of the ECU operation data. Based on the target fault feature set and the data acquisition time series, a time series dataset of the aircraft engine exhaust valve control system is constructed. A time series prediction model that integrates multi-scale attention is called to perform time series modeling processing on the time series dataset to generate multi-scale time series prediction results. Based on the multi-scale time-series prediction results, a dynamic standard reference curve for the key parameters of the aircraft engine exhaust valve control system is established. The dynamic standard reference curve corresponds to the key parameters of each dimension of the ECU operating data. The real-time ECU operating data collected from the aircraft engine is compared and analyzed in real time with the dynamic standard reference curve. By combining the correlation of target fault characteristics, abnormal information in which the real-time ECU operating data deviates from the dynamic standard reference curve is identified, and an aircraft engine operating status detection result is generated. The aircraft engine operating status detection result includes abnormal parameter dimension identifiers and abnormality degree information.

2. The aircraft engine detection method based on multi-task time series analysis according to claim 1, characterized in that, The acquisition of the operational monitoring data set generated by the aircraft engine during operation includes: Historical operating data and historical fault data from multiple consecutive operating periods are extracted from the ECU storage module of the aircraft engine, and the historical operating data and historical fault data from all periods are combined to form an initial data set. The initial data set is divided into a subset of intake pressure data, a subset of intake temperature data, a subset of propeller speed data, a subset of common rail pressure data, and a subset of fault alarm data. Each data subset contains the correspondence between the corresponding dimension data and the data acquisition time series, and each data subset covers all continuous running segments. Each data subset is cleaned in different time periods. After each cleaned data subset is standardized, the standardized intake pressure data subset, standardized intake temperature data subset, standardized propeller speed data subset, standardized common rail pressure data subset, and standardized fault alarm data subset of all time periods are integrated across time periods according to the data acquisition time sequence to form the operation monitoring data set.

3. The aircraft engine detection method based on multi-task time series analysis according to claim 1, characterized in that, The process involves performing cross-time period comparative representation learning on the operational monitoring data set. This involves extracting the differential features between normal operating state data and fault operating state data under different operating periods using a comparative representation learning model, generating a target fault feature set, including: From the operation monitoring data set, a normal operation status data subset and a fault operation status data subset are separated for each time period according to the operation period. The normal operation status data subset includes ECU operation data of each dimension corresponding to the no-fault alarm data, and the fault operation status data subset includes ECU operation data of each dimension corresponding to the fault alarm data. Each time period corresponds to a set of normal operation status data subset and fault operation status data subset. A subset of normal operating status data from all time periods is input into the feature encoding layer of the contrastive representation learning model for feature mapping processing to generate a normal status feature vector for each time period. The mean of the normal status feature vectors from all time periods is calculated to obtain a unified normal status feature vector across time periods. A subset of fault operation status data from all time periods is input into the feature encoding layer of the contrastive representation learning model. Feature mapping is performed on the fault data from each time period to generate a fault status feature vector for each time period. The mean of the fault status feature vectors from all time periods is calculated to obtain a unified fault status feature vector across time periods. The Detector model is called to train the cross-time period unified normal state feature vector and the cross-time period unified fault state feature vector. The hyperparameters of the Detector model are adjusted by the time-segmented cross-validation method so that the Detector model can distinguish between the normal state feature vector and the fault state feature vector in different time periods. The feature difference between the cross-time period unified normal state feature vector and the cross-time period unified fault state feature vector is calculated using the trained Detector model. At the same time, the deviation between the normal state feature vector of each time period and the cross-time period unified normal state feature vector, and the deviation between the fault state feature vector of each time period and the cross-time period unified fault state feature vector are also calculated. Extract the feature dimensions whose feature difference is greater than a set threshold and whose deviation is less than the deviation threshold in each time period, and use them as the target feature dimensions. For each target feature dimension, extract the first feature value distribution range of the cross-time unified normal state feature vector and the second feature value distribution range of the cross-time unified fault state feature vector under that target feature dimension, and calculate the difference region and the proportion of the difference region between the first feature value distribution range and the second feature value distribution range. The difference region, feature difference degree, difference region ratio, and identification information of the target feature dimension in the ECU operating data are integrated to generate a target fault feature set.

4. The aircraft engine detection method based on multi-task time series analysis according to claim 3, characterized in that, The process of calling the Detector model to train the cross-time period unified normal state feature vector and the cross-time period unified fault state feature vector, and adjusting the hyperparameters of the Detector model using a time-segmented cross-validation method, so that the Detector model can distinguish between normal state feature vectors and fault state feature vectors in different time periods, includes: The normal state feature vector and the fault state feature vector of each time period are combined to form a sample subset of that time period. A category label and a time period label are added to the samples in each sample subset. The category label includes a normal state label and a fault state label, and the time period label identifies the runtime segment to which the sample belongs. The sample subsets from all time periods are combined to form the training sample set of the Detector model. The training sample set of the Detector model is divided into multiple consecutive sub-sample sequences according to the time period order. Each sub-sample sequence contains sample subsets from at least two time periods. A time-segmented cross-validation strategy is adopted, in which one sub-sample sequence is selected as the validation sub-sequence and the remaining sub-sample sequences are used as training sub-sequences. The training sub-sequences are then input into the Detector model to start the training process of the Detector model. During the training process of the Detector model, the classifier module of the Detector model learns the class distinction boundary and time period adaptation parameters at the same time, enabling the Detector model to distinguish between normal and fault states and adapt to the feature differences of different time periods, and periodically calculates the training loss value of the training subsequence. When the training loss value is reduced to the set training threshold, the validation subsequence is input into the Detector model, the validation classification result output by the Detector model is obtained, the classification accuracy of the sample in the validation subsequence in each time period is calculated, and the average classification accuracy of all time periods is calculated. The hyperparameters of the Detector model are adjusted based on the classification accuracy and average classification accuracy for each time period. The hyperparameters of the Detector model include the number of hidden layer neurons, activation function type, regularization coefficient, and time period adaptation coefficient. If the classification accuracy of any time period is lower than the time period accuracy threshold, the weight of the time period adaptation coefficient corresponding to that time period is increased, and the Detector model is retrained; if the average classification accuracy is lower than the overall accuracy threshold, the number of hidden layer neurons or the activation function type of the Detector model are adjusted to reduce training loss. Each sub-sample sequence is used as a validation sub-sequence, and the rest are used as training sub-sequences. The training, validation, and hyperparameter tuning process of the Detector model is repeated to obtain multiple sets of hyperparameter combinations and the corresponding classification accuracy and average classification accuracy for each time period. The hyperparameter combination with the highest average classification accuracy and the smallest difference in classification accuracy across different time periods is selected as the final hyperparameters of the Detector model. The Detector model is then trained using the entire training sample set. At the same time, a time-adaptive loss function is introduced to reduce the impact of feature differences across different time periods on the classification results, resulting in the trained Detector model.

5. The aircraft engine detection method based on multi-task time series analysis according to claim 1, characterized in that, The process involves constructing a time-series dataset for an aircraft engine exhaust valve control system based on the target fault feature set and the data acquisition time series. A time-series prediction model incorporating multi-scale attention is then used to perform time-series modeling processing on the dataset, generating multi-scale time-series prediction results, including: Extract feature dimensions related to the aircraft engine exhaust valve control system from the target fault feature set. The feature dimensions include intake pressure feature dimension, intake temperature feature dimension, propeller speed feature dimension, and common rail pressure feature dimension. At the same time, calculate the association weight between each feature dimension related to the aircraft engine exhaust valve control system and the exhaust valve control logic. Based on the correlation weight, the feature dimensions related to the aircraft engine exhaust valve control system are screened, and the feature dimensions with a correlation weight greater than the weight threshold are retained as candidate feature dimensions. The ECU operating data corresponding to the candidate feature dimensions are arranged in the order of the data acquisition time series to form a single-dimensional time series data sequence for each candidate feature dimension. All single-dimensional time series data sequences are aligned and integrated according to the data acquisition time nodes, so that each data acquisition time node corresponds to a set of time series data for each candidate feature dimension. At the same time, an identifier of the runtime segment to which the time series data belongs is added to each set of time series data, forming a time series dataset of the aircraft engine exhaust valve control system. The time series dataset of the aircraft engine exhaust valve control system is divided into a training dataset, a validation dataset, and a test dataset. The training dataset is used for parameter training of the time series prediction model that integrates multi-scale attention. The validation dataset is used for hyperparameter tuning of the time series prediction model that integrates multi-scale attention. The test dataset is used for verifying the final prediction accuracy of the time series prediction model that integrates multi-scale attention. The training dataset is input into the time series prediction model that integrates multi-scale attention. The encoder module of the time series prediction model that integrates multi-scale attention contains multiple attention layers at different time scales, which perform attention modeling on time series data of various scale periods and extract time series dependencies at different scales. By weighted fusion of the features output by the attention layers at each time scale in the fusion layer of the time series prediction model that integrates multi-scale attention, and by using hyperparameter optimization methods to adjust the learning rate, number of iterations, batch size and scale parameters of each attention layer of the time series prediction model that integrates multi-scale attention, the prediction error of the time series prediction model that integrates multi-scale attention on the training dataset is gradually reduced. Input the validation dataset into the adjusted time series prediction model that integrates multi-scale attention, obtain the validation prediction results output by the adjusted time series prediction model that integrates multi-scale attention, calculate the error value between the validation prediction results and the actual data in the validation dataset, and if the error value is less than the set error threshold, then use the test dataset to perform the final validation of the time series prediction model that integrates multi-scale attention. If the prediction error of the test dataset is also less than the set error threshold, the trained time series prediction model is output. The trained time series prediction model is then called to perform time series modeling processing on the time series dataset of the aircraft engine exhaust valve control system, and outputs multi-scale time series prediction results containing the data change trends of each candidate feature dimension of various time periods.

6. The aircraft engine detection method based on multi-task time series analysis according to claim 1, characterized in that, The process of establishing dynamic standard reference curves for key parameters of the aircraft engine exhaust valve control system based on the multi-scale time-series prediction results includes: The predicted data sequences of each candidate feature dimension for each time period of various scales are extracted from the multi-scale time series prediction results. Each candidate feature dimension corresponds to multiple sets of predicted data sequences for different time periods. For each candidate feature dimension, multiple sets of predicted data sequences are subjected to trend fusion processing. The weighted average of multiple sets of predicted data at the same data collection time node is calculated to obtain the fused predicted data sequence for each candidate feature dimension. The weighting weight is determined according to the error of the prediction results in each time period. The smaller the error, the larger the weighting weight. Trend analysis is performed on the fused prediction data sequence for each candidate feature dimension to identify stable trend segments, fluctuating trend segments, and trend transition nodes in the fused prediction data sequence as the data acquisition time changes. The trend transition nodes are the dividing points between stable trend segments and fluctuating trend segments. For a stable trend segment, calculate the mean and standard deviation of the fused prediction data within the stable trend segment. Use the mean as the benchmark value of the stable trend segment and the standard deviation as the benchmark fluctuation range of the stable trend segment. At the same time, record the start and end times of the stable trend segment. For the fluctuation trend segment, the moving average and moving standard deviation of the fused prediction data within the fluctuation trend segment are calculated. The size of the moving window is dynamically adjusted according to the fluctuation frequency, and a dynamic benchmark value and dynamic fluctuation range that change over time are generated according to the data collection time interval. For trend transition nodes, calculate the rate of change of data before and after the trend transition node, determine the benchmark transition value at the trend transition node, and smoothly connect the benchmark value of the stable trend segment and the dynamic benchmark value of the fluctuating trend segment at the trend transition node. The stable trend segment baseline value, stable trend segment baseline fluctuation range, fluctuation trend segment dynamic baseline value, fluctuation trend segment dynamic fluctuation range, trend transition node and baseline transition value of each candidate feature dimension are associated with the corresponding data collection time series to form the dynamic baseline data sequence of that candidate feature dimension. The dynamic benchmark data sequence is connected in chronological order according to the data acquisition time sequence to generate a dynamic standard benchmark curve corresponding to the candidate feature dimension. The horizontal axis of the dynamic standard benchmark curve is the data acquisition time, and the vertical axis is the benchmark data value, benchmark fluctuation range and trend transformation node identifier of the candidate feature dimension. Set dynamic update trigger conditions. When the amount of newly added aircraft engine operation data reaches the set threshold or when a trend type not included in the newly added data appears, the dynamic standard baseline curve update process is triggered. Repeat the above steps to recalculate the baseline value and fluctuation range, so as to realize the dynamic adjustment of the dynamic standard baseline curve. The above-mentioned dynamic benchmark data sequence construction and dynamic standard benchmark curve generation operations are performed on all candidate feature dimensions to obtain dynamic standard benchmark curves corresponding to the key parameters of each dimension of ECU operating data.

7. The aircraft engine detection method based on multi-task time series analysis according to claim 6, characterized in that, The dynamic update trigger condition is set such that when the amount of newly added aircraft engine operating data reaches a set threshold or when a trend type not included in the newly added data appears, the dynamic standard baseline curve update process is triggered. The above steps are repeated to recalculate the baseline value and fluctuation range, thereby realizing the dynamic adjustment of the dynamic standard baseline curve, including: Define the data volume trigger conditions and trend type trigger conditions for dynamic updates; To address the data volume triggering conditions, a new data volume threshold is set. The new data volume threshold is determined based on the historical data collection frequency and the number of data dimensions. The amount of aircraft engine ECU operating data added since the last dynamic standard baseline curve update is counted. When the data volume reaches the new data volume threshold, a data volume update signal is triggered. To address the triggering conditions for trend types, a trend type library is established. This library contains common trend types such as stable trends, linear fluctuation trends, periodic fluctuation trends, and random fluctuation trends, along with their corresponding feature parameters. The time-series trend of newly added data is identified, and the trend feature parameters of the newly added data are extracted and compared with the feature parameters in the trend type library. If the trend feature parameters of the newly added data differ from the feature parameters of all types in the trend type library by more than the trend difference threshold, it is determined that there is a trend type not included in the newly added data, and the trend type update signal is triggered. Real-time monitoring of data volume update signals and trend type update signals; when any update signal is detected, the update process of the dynamic standard baseline curve is initiated. In the process of updating the dynamic standard baseline curve, the new data is integrated with the original time series dataset to form the updated time series dataset. All steps from multi-scale time series prediction to dynamic standard baseline curve generation are repeated, and the baseline values, baseline fluctuation ranges and trend transition nodes of each candidate feature dimension are recalculated. The differences between the dynamic standard baseline curves before and after the update are compared, and the change in the baseline value at the same time point is calculated. If the change is less than the update difference threshold, the original dynamic standard baseline curve is retained; if the change is greater than the update difference threshold, the newly generated dynamic standard baseline curve is used to replace the original dynamic standard baseline curve. Record the update time, update trigger conditions, differences in benchmark data before and after the update, and characteristics of newly added data for each dynamic standard baseline curve, forming a baseline curve update log; Store the updated dynamic standard baseline curve and baseline curve update log to the specified location.

8. The aircraft engine detection method based on multi-task time series analysis according to claim 1, characterized in that, The process involves comparing and analyzing real-time ECU operating data collected from the aircraft engine with the dynamic standard reference curve in real time, identifying abnormal information where the real-time ECU operating data deviates from the dynamic standard reference curve based on the correlation of target fault characteristics, and generating aircraft engine operating status detection results, including: The real-time ECU operating data collected by the aircraft engine is acquired. The real-time ECU operating data includes real-time intake pressure data, real-time intake temperature data, real-time propeller speed data, real-time common rail pressure data, and real-time fault alarm data. The real-time ECU operating data collected by the aircraft engine has a real-time data acquisition timestamp and operating condition identifier. The latest version of the dynamic standard reference curve is retrieved from the storage location, and the dynamic standard reference curve segments that match the current operating conditions are selected based on the operating condition identifier of the ECU operating data collected in real time from the aircraft engine. For each parameter dimension, extract the value of the real-time ECU operating data, the real-time data acquisition timestamp, and the operating condition identifier corresponding to that parameter dimension. On the dynamic standard benchmark curve segment matched by that parameter dimension, find the benchmark data value, benchmark fluctuation range, and trend type that are consistent with the real-time data acquisition timestamp. Calculate the difference between the real-time data value of this parameter dimension and the corresponding benchmark data value, determine whether the difference is within the benchmark fluctuation range, and combine it with the feature difference degree in the target fault features corresponding to this parameter dimension. If the difference exceeds the benchmark fluctuation range and the feature difference degree is greater than the set difference threshold, it is determined that the real-time ECU operating data of this parameter dimension is abnormal. If an anomaly is determined, record the anomaly parameter identifier, real-time ECU operating data value, baseline data value, difference, real-time data acquisition timestamp, operating condition identifier, and corresponding feature difference degree for that parameter dimension to form a single-dimensional anomaly record; Analyze the correlation between the dimensions of each abnormal parameter. Based on the correlation weight of each dimension in the target fault feature set, calculate the correlation strength between the abnormal parameter dimensions. If the correlation strength is greater than the correlation threshold, it is determined that the abnormal parameters belong to the related abnormalities caused by the same abnormal event, forming a group of related abnormalities. If the real-time ECU operating data collected by the aircraft engine contains real-time fault alarm data, then the alarm type identifier and alarm trigger timestamp corresponding to the fault alarm data are extracted. The alarm type identifier is then associated with the analysis results of the single-dimensional abnormal records and associated abnormal groups before and after the alarm trigger timestamp to form alarm-related abnormal records. The analysis results of all single-dimensional anomaly records, associated anomaly groups, and alarm-related anomaly records are summarized. The number of parameter dimensions with anomalies, the number of associated anomaly groups, the difference between each anomaly parameter, the duration of the anomaly, and the number of alarm associations are counted to obtain the summary information. The above-mentioned summary information, single-dimensional anomaly records, analysis results of associated anomaly groups, and alarm-related anomaly records are integrated to form an anomaly information summary table. Based on the anomaly information summary table, the aircraft engine operating status detection results are generated.

9. The aircraft engine detection method based on multi-task time series analysis according to claim 8, characterized in that, The analysis examines the correlation between various abnormal parameter dimensions. Based on the correlation weights of each dimension in the target fault feature set, the correlation strength between abnormal parameter dimensions is calculated. If the correlation strength is greater than a correlation threshold, the abnormal parameters are determined to belong to a group of related abnormalities caused by the same abnormal event, forming a group of related abnormalities, including: The correlation weight matrix between each parameter dimension is extracted from the target fault feature set. The rows and columns of the correlation weight matrix correspond to different parameter dimensions. The element values ​​of the correlation weight matrix represent the correlation weight between two corresponding parameter dimensions. The correlation weight is calculated based on the functional dependency of the parameters in engine operation and historical abnormal correlation data. Extract all parameter dimension identifiers that are abnormal from the single-dimensional abnormal records to form an abnormal parameter dimension list. Combine the abnormal parameter dimensions in the abnormal parameter dimension list in pairs to generate abnormal parameter combination pairs. For each abnormal parameter pair, find the corresponding matrix element value in the association weight matrix and use the matrix element value as the initial association strength. Extract the synchronization of changes in the real-time ECU operating data to the abnormal parameter combination to obtain the similarity of the change trends; The initial correlation strength and the similarity of the change trend are weighted and fused to obtain the final correlation strength of the anomaly parameter combination pair; If the final association strength of any abnormal parameter combination pair is greater than the association strength threshold, then the abnormal parameter combination pair is marked as an abnormal association pair. Cluster all associated anomaly pairs and group associated anomaly pairs containing the same anomaly parameter dimension into the same associated anomaly group. Calculate the average association strength within each associated anomaly group. If the average association strength within the associated anomaly group is greater than the association strength threshold, then the associated anomaly group is confirmed as a valid associated anomaly group, and the anomaly parameters within the valid associated anomaly group are determined to be associated anomalies caused by the same anomaly event. If the average correlation strength within a group of anomalous associations is less than the correlation strength threshold, the group of anomalous associations is split, the anomalous parameter dimension with the lowest correlation strength is removed, and the average correlation strength within a split group of anomalous associations is recalculated until the average correlation strength within a group of anomalous associations meets the requirements or only one anomalous parameter dimension remains in the group of anomalous associations. Record the dimensions of the abnormal parameters for each effective association anomaly group, the association strength of each abnormal parameter pair within the effective association anomaly group, the average association strength within the effective association anomaly group, and the criteria for determining association anomalies, and include them in the anomaly information summary table.

10. An aircraft engine testing system based on multi-task time series analysis, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the aircraft engine detection method based on multi-task timing analysis as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images