An airborne equipment reliability prediction method based on field data

Abstract

The application discloses an airborne equipment reliability prediction method based on field data, and relates to the technical field of machine learning, comprising: inputting a potential failure label into an equipment prediction model, performing deep feature extraction through a feature enhancement layer, performing spatio-temporal attention fusion through a prediction decision layer, and generating a reliability prediction report; extracting pre-position features of a historical failure case library, associating and analyzing the pre-position features in combination with spatio-temporal context features, and obtaining a failure trigger rule set; performing simulation and modeling according to the failure trigger rule set, generating simulation data, and performing deviation quantitative analysis on the reliability prediction report and the simulation data to generate a verification report. The application realizes the early identification of pre-position signs, associated causes and trigger relationships of failures by constructing a failure trigger rule set and performing deviation quantitative verification on the reliability prediction result, improves the reliability of the prediction conclusion, reduces the risk of false positives and false negatives, and enhances the reliability of engineering application.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method for predicting the reliability of airborne equipment based on field data. Background Technology

[0002] Airborne equipment reliability prediction technology is an important research direction in the fields of aviation equipment health management, fault early warning, and maintenance support. Early related technologies mainly focused on failure mechanism modeling, life statistical analysis, and periodic test evaluation. By establishing the correspondence between factors such as stress, load, temperature, and vibration and failure probability, the remaining usable state of airborne equipment was estimated. With the continuous improvement of airborne sensing, embedded recording, digital maintenance, and airborne computing capabilities, reliability prediction has gradually shifted from single physical parameter analysis to multi-source field data-driven analysis. The research focus has also expanded from static life assessment to dynamic risk identification during operation. Existing technologies have developed schemes that utilize data such as temperature, current, vibration, operational status, and maintenance records for feature extraction, pattern recognition, and predictive modeling. Machine learning, deep learning, and time series analysis methods have been introduced to improve the ability to identify the state evolution of airborne equipment under complex operating conditions, enabling reliability prediction to develop from offline assessment to online perception, continuous prediction, and decision support.

[0003] However, existing technologies still have some shortcomings. Although existing solutions can access various types of field data, they lack depth in mining the temporal, operational, and evolutionary correlations between equipment physical parameters, equipment operational parameters, and maintenance logs. They often remain at the level of feature splicing or local statistics, making it difficult to form a unified representation that takes into account both the operational context and the fault evolution law. This results in limited prediction accuracy and adaptability under complex operating conditions. Existing solutions mostly focus on fault risk output or state classification results, and do not adequately consider the reverse tracing of fault precursor symptoms, the causal extraction of fault triggering relationships, and the consistency verification between prediction results and simulation results. This leads to weak interpretability, verifiability, and engineering application credibility of the prediction process, making it difficult to meet the application requirements of airborne equipment under high reliability and high safety constraints. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for predicting the reliability of airborne equipment based on field data to address the problems of insufficient correlation mining of multi-source field data and weak interpretability and verifiability of prediction results.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for predicting the reliability of airborne equipment based on field data, comprising: Collect multi-source field data from airborne equipment and preprocess it to extract the spatiotemporal context features of the preprocessed multi-source field data. The spatiotemporal context features are matched and analyzed with a historical failure case database to generate potential failure labels; The potential fault labels are input into the device prediction model, the feature enhancement layer performs deep feature extraction, the prediction decision layer performs spatiotemporal attention fusion, and a reliability prediction report is generated. Extract the preceding features from the historical fault case library, and combine them with spatiotemporal context features to perform correlation analysis on the preceding features to obtain the fault triggering rule set; Simulation is performed based on the fault triggering rule set to generate simulation data. The reliability prediction report and the simulation data are then subjected to deviation quantification analysis to generate a verification report.

[0007] As a preferred embodiment of the airborne equipment reliability prediction method based on field data described in this invention, the multi-source field data includes equipment physical parameters, equipment operation parameters, and maintenance logs; The preprocessing includes data cleaning, format conversion, deduplication, normalization, and outlier handling.

[0008] As a preferred embodiment of the airborne equipment reliability prediction method based on field data described in this invention, the specific steps for extracting spatiotemporal context features are as follows: Sliding window statistics are performed on the physical parameters of the equipment to obtain physical state characteristics; The detection equipment's operational parameters undergo time-series mutation analysis to obtain abnormal state characteristics; Analyze the failure frequency in the maintenance logs to obtain maintenance frequency characteristics; The physical state characteristics, anomaly characteristics, and maintenance frequency characteristics are structurally integrated to generate spatiotemporal context characteristics.

[0009] As a preferred embodiment of the airborne equipment reliability prediction method based on field data described in this invention, the historical fault case database is constructed using the following steps. Spatiotemporal context features are spatiotemporally aligned to construct a spatiotemporal correlation matrix, and correlation mining is performed on the spatiotemporal correlation matrix to generate potential fault correlation patterns. Cross-validate potential fault association patterns and maintenance frequency characteristics, and perform structured processing to generate a historical fault case library.

[0010] As a preferred embodiment of the airborne equipment reliability prediction method based on field data described in this invention, the specific steps for generating potential fault labels are as follows: Dynamic window similarity matching is performed between spatiotemporal context features and a historical failure case library to generate a candidate failure mode set; The co-occurrence frequency of the candidate fault mode set is statistically analyzed and the operating condition consistency calibration is performed to obtain potential fault labels.

[0011] As a preferred embodiment of the airborne equipment reliability prediction method based on field data described in this invention, the equipment prediction model is constructed using the following specific steps. Call and initialize the residual convolutional network and multi-head Transformer to build the feature enhancement layer and prediction decision layer; A device prediction model is constructed by performing backpropagation and gradient update on the feature enhancement layer and the prediction decision layer using the focus loss function.

[0012] As a preferred embodiment of the airborne equipment reliability prediction method based on field data described in this invention, the specific steps for generating the reliability prediction report are as follows: The potential fault labels are input into the device prediction model, and the feature enhancement layer performs deep feature extraction through residual blocks to obtain high-order temporal features. The predictive decision-making layer uses a multi-head self-attention mechanism to perform spatiotemporal attention fusion and generate a risk score; High-order time-series features and risk scores are integrated through feature channels to generate a reliability prediction report.

[0013] As a preferred embodiment of the airborne equipment reliability prediction method based on field data described in this invention, the specific steps for extracting the pre-features from the historical fault case library are as follows: Reverse tracing and temporal pattern matching are performed on the spatiotemporal context features in the historical fault case library to form a feature candidate pool; The synergy between the matching feature candidate pool and the potential fault association patterns is used to obtain the preliminary features.

[0014] As a preferred embodiment of the airborne equipment reliability prediction method based on field data described in this invention, the specific steps for obtaining the fault triggering rule set are as follows: Spatial topological embedding and temporal dynamic decoupling are performed on the pre-features and spatiotemporal context features to obtain fault association features; Based on potential fault association patterns, causal rules of fault association features are extracted and causal inference verification is performed to obtain a fault triggering rule set.

[0015] As a preferred embodiment of the airborne equipment reliability prediction method based on field data described in this invention, the specific steps for generating the verification report are as follows: The feature combination and triggering conditions of the fault triggering rule set are analyzed to obtain the rule execution elements. The rule execution elements are then numerically mapped and multi-dimensional scenario simulations are performed to generate simulation data. The reliability prediction report is dynamically mapped to the simulation data to construct the prediction-simulation correlation matrix. The deviation of the prediction-simulation correlation matrix is ​​then quantitatively analyzed to generate a verification report.

[0016] The beneficial effects of this invention are as follows: By constructing a fault triggering rule set, it enables the early identification of fault precursors, associated causes, and triggering relationships, further elevating airborne equipment reliability prediction from result output to a forward-looking analysis of the fault formation process, improving the interpretability of prediction results, enhancing early warning capabilities, and improving prediction stability under complex operating conditions; by generating simulation data and performing deviation quantification verification on the reliability prediction results, it achieves cross-verification and consistency confirmation of prediction conclusions, providing verifiable comparison and correction basis for prediction results, improving the credibility of prediction conclusions, reducing the risk of false alarms and missed alarms, and enhancing the reliability of engineering applications. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for predicting the reliability of airborne equipment based on field data.

[0019] Figure 2 A flowchart for spatiotemporal context feature extraction and the construction of a historical fault case library.

[0020] Figure 3 A flowchart for generating equipment prediction models and reliability prediction reports.

[0021] Figure 4 This is a flowchart for generating and verifying fault triggering rule sets. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for predicting the reliability of airborne equipment based on field data, including the following steps: S1. Collect multi-source field data from airborne equipment and preprocess it to extract the spatiotemporal context features of the preprocessed multi-source field data. S1.1 Multi-source field data includes equipment physical parameters, equipment calculation parameters, and maintenance logs; It should be noted that the physical parameters of the equipment are acquired by multi-source sensors installed on the airborne equipment. These include temperature sensors to collect temperature data during operation, acceleration sensors to collect vibration amplitude and frequency data, pressure sensors to collect hydraulic pressure data, current transformers to collect current data in the power supply lines, and voltage sensors to collect operating voltage data. The equipment's operational parameters are extracted from the airborne equipment's own operating records. The maintenance log is filled out after each maintenance or fault handling, and the record includes the specific time of the fault occurrence, a description of the fault phenomenon, the fault code, the name and quantity of replaced parts, and the start and end times of the maintenance.

[0026] S1.2 Preprocessing includes data cleaning, format conversion, deduplication, normalization, and outlier handling; It should be noted that the collected multi-source field data, including equipment physical parameters, equipment operating parameters, and maintenance logs, undergoes data cleaning to remove incomplete, erroneous, or irrelevant data records. Format conversion is performed to ensure all data sources adhere to a unified format standard for subsequent processing. Deduplication is performed to eliminate duplicate data entries, ensuring the uniqueness of the dataset. A normalization step maps data values ​​from different sources to the same scale, preventing certain features from unduly affecting the results due to excessively large differences in magnitude. Outlier handling identifies and corrects data points that deviate from the normal range, ensuring the accuracy and reliability of the dataset.

[0027] S1.3 Perform sliding window statistics on the physical parameters of the equipment to obtain physical state characteristics; It should be noted that a fixed-length sliding time window and step size are set according to the fluctuation characteristics of the preprocessed equipment physical parameters. Data segments are extracted sequentially along the time axis. Within each sliding time window, the mean, variance, maximum value, minimum value, peak value, peak-to-peak value, and rate of change of each physical parameter (such as temperature, pressure, current, etc.) within the equipment physical parameters are obtained. All statistics are integrated according to parameter type and sliding time window order to generate structured physical state characteristics.

[0028] S1.4. The detection equipment's operational parameters undergo time-series mutation analysis to obtain abnormal state characteristics; It should be noted that the equipment operation parameters are arranged in order of timestamps to form an operation parameter time series. According to the preset detection standard, the operation parameter time series is traversed, and the deviation of the current equipment operation parameter in the operation parameter time series from the benchmark value in the preset detection standard is obtained point by point. Abnormal points that deviate from the benchmark are identified and abnormal features of abnormal points are extracted, including the timestamp of the abnormality, the deviation range, the duration and the trend changes before and after. The features of all abnormal points are integrated to form a state abnormality feature.

[0029] It should also be noted that the testing standards are based on the data distribution definition of historical equipment operation parameters.

[0030] S1.5. Statistically analyze the failure frequency in the maintenance log and obtain maintenance frequency characteristics; It should be noted that fault information, including the timestamp of the fault occurrence, fault type, faulty component, and repair completion time, should be extracted from the maintenance log. The fault information should be grouped by timestamp and fault type, and the number of faults, fault interval, and high-frequency fault periods within each group should be statistically analyzed. The statistical results should be integrated in a structured "time-type-frequency" format to generate maintenance frequency characteristics (e.g., "Hydraulic equipment faults are concentrated in high-temperature periods during a certain quarter, occurring an average of 3 times per month").

[0031] S1.6. Structure and integrate physical state characteristics, abnormal state characteristics, and maintenance frequency characteristics to generate spatiotemporal context characteristics; It should be noted that, based on the operating cycle of the airborne equipment, a unified time analysis window is set, and physical state characteristics, abnormal state characteristics, and maintenance frequency characteristics are mapped to the time analysis window. Within each time analysis window, physical state characteristics, abnormal state characteristics, and maintenance frequency characteristics are integrated by feature splicing to form a structured vector. The structured vector is then normalized to generate spatiotemporal context features.

[0032] S2. Match and analyze the spatiotemporal context features with the historical fault case library to generate potential fault labels; S2.1. Spatiotemporally align the spatiotemporal context features, construct a spatiotemporal correlation matrix, and perform correlation mining on the spatiotemporal correlation matrix to generate potential fault correlation patterns. It should be noted that time offset of spatiotemporal context features is eliminated by timestamp synchronization, and the spatial reference system of spatiotemporal context features is unified by spatial coordinate transformation to ensure that spatiotemporal context features from different sources correspond in the "time-space" dimension. Spatiotemporal context features are divided into fixed-time sliding spatiotemporal windows (e.g., every 30 minutes). The spatiotemporal context features of each spatial partition within each sliding spatiotemporal window are filled into matrix elements to form a "sliding spatiotemporal window × spatial partition" spatiotemporal correlation matrix. The Pearson correlation coefficient of spatiotemporal context features under different "sliding spatiotemporal window × spatial partition" combinations in the spatiotemporal correlation matrix is ​​obtained to quantify the linear correlation strength of different spatiotemporal context features in the spatiotemporal dimension. By setting a correlation strength threshold, features with linear correlation strength higher than the correlation strength threshold are selected. The spatiotemporal context features are combined and defined as strongly correlated feature combinations. Based on historical data of normal operation of airborne equipment, the baseline occurrence frequency of strongly correlated feature combinations is obtained (e.g., strongly correlated feature combinations occur 0.5 times per hour under normal conditions). The actual occurrence frequency of strongly correlated feature combinations is statistically analyzed. If the actual occurrence frequency is higher than the baseline occurrence frequency, it is judged as an abnormally high correlation combination. The fault indication significance of abnormally high correlation combinations is explained in combination with physical mechanisms (e.g., high temperature causes component expansion) (e.g., the high frequency co-occurrence of the "high temperature-vibration" combination may indicate bearing lubrication failure or mechanical wear). Potential fault association patterns are generated (e.g., when the abnormally high frequency of the combination of "high temperature in sliding spatiotemporal window W1-region R1 + vibration in sliding spatiotemporal window W2-region R2" occurs, there is a risk of mechanical failure in airborne equipment).

[0033] It should also be noted that the association strength threshold is defined based on the association strength distribution of historical data of airborne equipment operation, and an exemplary value range is (0.4, 0.6).

[0034] S2.2 Cross-validate potential fault association patterns and maintenance frequency characteristics and perform structured processing to generate a historical fault case library; It should be noted that the abnormally high correlation combinations in the potential fault association patterns are checked for time overlap, spatial range matching, and type association verification (such as whether the abnormally high correlation combinations correspond to mechanical wear faults) with the actual fault records in the maintenance logs to generate fault association information. The ratio of the actual occurrence frequency of the strong correlation feature combination to the actual fault frequency in the maintenance frequency feature is used as the fault matching degree. A unified historical fault case field is defined, including "Case ID", "Fault Time", "Fault Type", "Affected Area", "Abnormally Highly Correlated Combination", "Actual Fault Frequency", and "Fault Matching Degree". The processed fault association information is filled and integrated according to the field requirements to form a historical fault case library.

[0035] S2.3. Perform dynamic window similarity matching between the spatiotemporal context features and the historical fault case library to generate a set of candidate fault modes; It should be noted that the spatiotemporal context features are dynamically matched with the corresponding fields in the historical fault case library in multiple dimensions: in the time dimension, the absolute difference between the current timestamp and the corresponding fault time in the historical fault case library is obtained; in the spatial dimension, the percentage of overlap between the current "spatial partition" and the corresponding affected area in the historical fault case library is calculated; in the feature dimension, the absolute difference between the current physical state features and the corresponding abnormally high correlation combinations in the historical fault case library is obtained; the matching results of the multi-dimensional dynamic matching are weighted and fused to obtain a matching score, and highly similar historical fault cases with matching scores higher than the preset score threshold are selected, and common features (such as the physical state combination of "spatial partition Ra temperature mean > 80℃ + spatial partition Rb vibration amplitude > 10mm / s") and the corresponding fault type (such as mechanical wear) are extracted to generate a candidate fault mode set (such as "spatial partition Ra temperature mean > 80℃ + spatial partition Rb vibration amplitude > 10mm / s + timestamp abnormal current deviation, corresponding to mechanical wear fault").

[0036] It should also be noted that the scoring threshold is defined based on the statistical distribution of matching scores in the historical fault case database, with an exemplary range of values ​​being (0.6, 0.8).

[0037] When weighted and fused, the weights of each dimension in the multi-dimensional dynamic matching results are determined according to the retrospective statistical results in the historical fault case library. Specifically, historical fault cases are selected one by one from the historical fault case library, and the spatiotemporal context features of the corresponding time period before the fault occurred are extracted as retrospective samples. For each retrospective sample, single-dimensional matching is performed with the remaining historical fault cases in the historical fault case library in the time, spatial, and feature dimensions. In the time dimension, they are arranged in ascending order of the absolute difference between the current timestamp and the fault time in the historical fault case. In the spatial dimension, they are arranged in descending order of the proportion of overlap between the current spatial partition and the affected area in the historical fault case. In the feature dimension, they are arranged in descending order of the current physical state features. The absolute differences between the historical fault cases and the abnormally highly correlated combinations are arranged from smallest to largest. The differences between the corresponding matching results of adjacent historical fault cases after each dimension are obtained, and the adjacent position with the largest difference is selected as the boundary position of the front interval. The historical fault case segment before the boundary position of the front interval is determined as the front interval of the corresponding dimension. The ratio of the number of historical fault cases in the front interval of the dimension that are consistent with the fault type of the current backtracking sample to the total number of historical fault cases in the front interval is calculated as the identification contribution value of the dimension. After obtaining the identification contribution values ​​of the time dimension, spatial dimension, and feature dimension, the identification contribution values ​​are normalized to obtain the weights of the time dimension, spatial dimension, and feature dimension.

[0038] S2.4. Calculate the co-occurrence frequency of the candidate fault mode set and perform operating condition consistency calibration to obtain potential fault labels; It should be noted that the co-occurrence frequency of spatiotemporal context features in the candidate fault mode set (such as the frequency of simultaneous occurrence in mechanical wear faults) is extracted; operating parameters during the operation of airborne equipment are extracted, including equipment load rate, ambient temperature, and cooling status recorded by PLC, and aligned with the spatiotemporal context features in the candidate fault modes by timestamp to form a labeled dataset of "operating parameters-feature combinations". At the same time, according to the operating characteristics of airborne equipment and the distribution of relevant historical data, the labeled dataset is divided into different operating intervals. The co-occurrence frequency and Pearson correlation coefficient of spatiotemporal context features with the corresponding fault type in each operating interval are obtained. The differences in the correlation indicators of different operating intervals are compared. If the difference in the correlation indicators exceeds the preset threshold, it is determined that the spatiotemporal context features are not stable enough in the current operating condition, and the data of the current operating condition is excluded; if the difference in the correlation indicators of each operating interval does not exceed the preset threshold, it is confirmed that the current spatiotemporal context features can stably reflect the fault risk under different operating conditions, and are retained as potential fault labels.

[0039] It should also be noted that the indicator threshold is defined based on the distribution of historical related indicators, with an exemplary range of ±10% (if the difference between related indicators exceeds 10%, the correlation stability is deemed insufficient).

[0040] S3. Input the potential fault labels into the device prediction model, the feature enhancement layer performs deep feature extraction, the prediction decision layer performs spatiotemporal attention fusion, and generates a reliability prediction report. S3.1 Call and initialize the residual convolutional network and multi-head Transformer to build the feature enhancement layer and prediction decision layer; It should be noted that the initialization of the residual convolutional network involves the following steps: setting the kernel weights of each residual block according to the number of input and output channels; initializing the bias term to a zero vector; allocating independent memory space for each residual block and initializing the scaling factor and offset parameters of the batch normalization layer; and synchronously initializing the multi-head Transformer (a neural network architecture based on a self-attention mechanism) component. Specifically, this involves initializing the query weight matrix, key weight matrix, and value weight matrix of the attention head using a normal distribution; generating fixed vectors for position encoding using sine and cosine functions; generating the feedforward network weight matrix using the Kaiming initialization method and setting the bias term to a zero vector; and then calling the residual convolutional network and other machine learning architectures such as the multi-head Transformer to adjust the kernel size, stride, padding, and Transformer parameters of the residual blocks in the residual convolutional network. The hyperparameters of the Transformer, such as the number of heads, hidden layer dimensions, and feedforward network dimensions, are adjusted to be compatible with the dimensions of the input potential fault labels. The output of the residual convolutional network is adjusted to be consistent with the input dimension of the multi-head Transformer (e.g., projected to 768 dimensions) through a linear projection layer and then input into the Transformer encoder. At the same time, the output of the multi-head Transformer and the output of the residual convolutional network are residually concatenated and then weighted and fused through layer normalization and a feedforward network to generate enhanced multimodal spatiotemporal features, thus completing the construction of the feature enhancement layer. The enhanced multimodal spatiotemporal features are input into a fully connected layer, and after nonlinear transformation by the ReLU activation function, the fault type probability distribution (such as the probability of mechanical wear, electrical faults, etc.) is output through a Softmax layer. At the same time, the fault risk score is output through a linear layer, thus completing the construction of the prediction decision layer.

[0041] It should also be noted that the query weight matrix, key weight matrix, and value weight matrix are used to convert the enhanced spatiotemporal features after linear projection into query vectors, key vectors, and value vectors, respectively. For each attention head, the query weight matrix, key weight matrix, and value weight matrix are all two-dimensional parameter structures with the row direction corresponding to the input feature dimension and the column direction corresponding to the feature dimension within the head. When all attention heads are computed in parallel, the query weight matrix, key weight matrix, and value weight matrix corresponding to each attention head together constitute the overall query weight matrix, overall key weight matrix, and overall value weight matrix.

[0042] The weights in the query weight matrix, key weight matrix, and value weight matrix are generated during the multi-head Transformer initialization phase. Specifically, based on the feature dimensions after the potential fault label input and the feature segmentation results of each attention head, initial connection weights are established between the input feature position and the query vector generation position, between the input feature position and the key vector generation position, and between the input feature position and the value vector generation position. The corresponding initial connection weights are then organized into query weight matrix, key weight matrix, and value weight matrix, respectively. The initial connection weights are initial parameter values ​​randomly generated according to a normal distribution after the matrix structure is determined.

[0043] S3.2. Use the focus loss function to perform backpropagation and gradient update on the feature enhancement layer and the prediction decision layer to build a device prediction model; It should be noted that the focus loss function is used to perform forward propagation on the feature enhancement layer and the prediction decision layer to obtain a loss matrix containing weighted information; the loss matrix is ​​then backpropagated to generate the gradient tensors of the feature enhancement layer and the prediction decision layer; the Adam optimizer is used to update the gradient tensors of each layer to obtain the updated network weights; based on the updated network weights, depthwise separable convolution is used to perform hyperparameter tuning on the feature enhancement layer and the prediction decision layer, and cross-layer parameterization is performed through residual connections to output the constructed device prediction model; Next, the equipment prediction model is trained by pairing potential fault labels with corresponding spatiotemporal context features to form a prediction sample set. The prediction sample set is then stratified according to the data type of the potential fault labels, and training and validation sets are divided based on the proportion of each data type in the prediction sample set. The training set is augmented using time-series interpolation and spatial feature interpolation, and data is standardized using batch normalization to form enhanced standard samples. The Adam optimizer is used to dynamically adjust the learning rate of the enhanced standard samples, while early stopping is applied simultaneously to monitor the validation set loss and obtain intermediate parameters. The focus loss function is used to quantify the loss of the intermediate parameters and obtain the prediction error of the validation set. Training terminates when the prediction error on the validation set exceeds the convergence threshold for 10 consecutive rounds, and the trained equipment prediction model is output synchronously.

[0044] It should also be noted that the convergence threshold is defined based on the rate of change of the prediction error on the validation set, and its value ranges from 0.0001 to 0.01.

[0045] S3.3 Input the potential fault labels into the device prediction model, and the feature enhancement layer performs deep feature extraction through residual blocks to obtain high-order temporal features; It should be noted that potential fault labels and spatiotemporal context features are concatenated along the channel dimension to form a "label-feature" fusion input. Deep feature extraction is performed through multiple residual blocks, where each residual block contains a convolutional layer (such as a 3×3 temporal convolutional kernel to extract local temporal patterns), layer normalization, and a ReLU activation function. The kernel time window is divided according to the size of the temporal convolutional kernel in the residual block, and gradient vanishing is mitigated through residual connections. Deep features from local details to global dependencies are extracted step by step. After deep feature extraction through multiple residual blocks, high-order temporal features are output.

[0046] The specific steps for outputting high-order temporal features after deep feature extraction from the layer residual blocks are as follows: The bottom residual block uses a temporal convolutional layer to extract local temporal patterns (such as sudden temperature rise or vibration peaks of a single component) within a short kernel time window (e.g., 5 seconds). After performing nonlinear transformation by layer normalization and ReLU activation function, the bottom local features are output through skip connections. The middle residual block receives the bottom local features, uses a temporal convolutional layer with the same structure to expand the receptive field, integrates adjacent kernel time windows, performs normalization and nonlinear transformation by layer normalization and ReLU activation function, and then generates middle features reflecting the temperature-vibration co-trend through skip connections. The top residual block receives features from the middle layer and uses a larger temporal convolution kernel (e.g., 7×1, stride 3) to further expand the kernel time window, aggregating multi-stage temporal dependencies (e.g., the correlation pattern of continuous temperature increase in the first 30 seconds, periodic abnormal vibration frequency in the middle 20 seconds, and increasing current fluctuation amplitude in the last 50 seconds). After being normalized and nonlinearly transformed by layer normalization and ReLU activation function, high-order temporal features are output through skip connections.

[0047] S3.4 The predictive decision-making layer uses a multi-head self-attention mechanism to perform spatiotemporal attention fusion and generate a risk score; It should be noted that the high-order temporal features output by the feature enhancement layer are linearly transformed to generate Q (query) matrix, K (key) matrix, and V (value) matrix. Specifically, the high-order temporal feature vector output by the feature enhancement layer is input into three independent linear layers, and Q-space mapping, K-space mapping, and V-space mapping are performed respectively to generate the corresponding Q matrix, K matrix, and V matrix. The Q matrix is ​​used to capture the "query requirement" of the current spatiotemporal point (such as "whether the current airborne equipment status is abnormal"). The K matrix is ​​used to encode the "key information" of each spatiotemporal point (such as "feature patterns when historical anomalies occurred"). The V matrix is ​​used to store the "value content" of each spatiotemporal point (such as "high-order temporal features when an anomaly occurred"). The correlation score between each spatiotemporal point is obtained by the dot product of the Q matrix and the K matrix. The correlation score is normalized by the softmax function to generate the attention weight matrix. The attention weight matrix and the V matrix are summed element-wise to obtain the spatiotemporal attention feature. The spatiotemporal attention feature is mapped to the target dimension through a non-linear activation function and a fully connected layer to generate a risk score.

[0048] It should also be noted that, taking each current spatiotemporal point as a unit, the corresponding association score sequence is normalized so that the sum of the weight values ​​of all candidate spatiotemporal points corresponding to the current spatiotemporal point is one, and the candidate spatiotemporal point with the larger association score corresponds to the larger weight value, and the candidate spatiotemporal point with the smaller association score corresponds to the smaller weight value; the weight distribution results corresponding to each current spatiotemporal point are arranged in spatiotemporal position order to form an attention weight matrix.

[0049] S3.5 Integrate high-order time series features and risk scores through feature channels to generate a reliability prediction report; It should be noted that the IEEE precise timestamp protocol is used to align the time steps of high-order time series features with the timestamps of risk scores through bidirectional linear resampling, eliminating time offset errors. Feature optimization is performed through a dual-branch feature processing layer, specifically: channel compression is applied to high-order time series features using temporal-spatial convolution kernels, while the risk scores are smoothed using a time sliding window to eliminate short-term noise fluctuations; weighted fusion is performed through a gated attention fusion layer to generate a hybrid feature tensor, which is then input into the GeLU activation function for nonlinear feature transformation, and the dimensionality is increased through a fully connected layer to output a reliability prediction report.

[0050] S4. Extract the preceding features from the historical fault case library, and perform correlation analysis on the preceding features in combination with spatiotemporal context features to obtain the fault triggering rule set; S4.1 Perform reverse tracing and temporal pattern matching on the spatiotemporal context features in the historical fault case database to form a feature candidate pool; It should be noted that, based on the historical failure case library, spatiotemporal context features of a fixed duration (e.g., 6 hours before the failure) are extracted from the failure timestamp to form a failure pre-feature sequence. The time dimensions of different pre-feature sequences are aligned, and the consistency of failure type (e.g., more than 90% of cases within a cluster are mechanical wear failures) and frequency of occurrence of relevant historical failure cases in each failure pre-feature sequence are statistically analyzed and stored in a structured manner to form a feature candidate pool.

[0051] S4.2. Match the synergy between the candidate feature pool and the potential fault association pattern to obtain the preliminary features; It should be noted that the verification process involves verifying whether the physical state features in the feature candidate pool fully cover the strongly correlated feature combinations in the potential fault association pattern (e.g., whether the "average temperature" in the feature candidate pool corresponds to "high temperature"). It also involves checking whether the spatial partitions in the feature candidate pool are consistent with the spatial range of the potential fault association pattern (e.g., whether region Ra-Rb overlaps with region R1-R2). The spatiotemporal context features matched by the degree of coordination are then integrated with the fields of "case ID", "fault type", "affected area" and "strongly correlated feature combination" in the historical fault case library to generate preliminary features.

[0052] S4.3. Perform spatial topological embedding and temporal dynamic decoupling on the pre-features and spatiotemporal context features to obtain fault association features; It should be noted that the spatial distances between the spatial intervals involved in the pre-features are obtained to generate a spatial distance matrix. Combined with the connection relationship of airborne equipment components, the spatial distance matrix is ​​converted into a topological correlation matrix, which directly reflects the physical correlation pattern between spatial intervals. For the temporal pattern in the pre-features (such as "continuous temperature rise + sudden vibration change") and the physical state time series in the spatiotemporal context features (such as the 6-hour time series of R1 temperature and R2 vibration), the long-term trend (such as the average rate of change of temperature rising by 2℃ per hour) and short-term fluctuations (such as the sudden increase of vibration amplitude by 3mm / s at a certain time point) are extracted through a sliding window. The fault correlation features are then fused to generate fault correlation features. Specifically, the physical correlation pattern embedded in the spatial topology, the long-term trend of temporal dynamic decoupling (such as "continuous temperature rise of R1") and the short-term fluctuations (such as "sudden vibration change of R2") are aligned by time steps to generate fault correlation features.

[0053] It should also be noted that the steps to convert the spatial distance matrix into a topological association matrix are as follows: obtain the physical connection relationship data of airborne equipment components (such as "engine R1 and gearbox R2 are directly connected through a drive shaft"); set the edge weights of airborne equipment components with direct physical connections in the physical connection relationship data to a minimum value to indicate strong association; keep the edge weights of airborne equipment components without direct connections at their original spatial distances; divide airborne equipment components with direct physical connections into the same community, and airborne equipment components without direct physical connections into different communities; mark airborne equipment components within the same community as actual connection types such as direct drive and signal transmission, and mark airborne equipment components in different communities as having no direct connection type, thus forming a topological association matrix that reflects the real physical association between airborne equipment components.

[0054] S4.4 Based on potential fault association patterns, extract causal rules of fault association features and perform causal inference verification to obtain a fault triggering rule set; It should be noted that, based on the historical fault case database, the co-occurrence frequency of strongly correlated feature combinations in fault-related features and potential fault-related patterns is statistically analyzed in chronological order; combined with the physical mechanism of airborne equipment (e.g., "high temperature causes thermal expansion of components, triggering vibration of adjacent components"), the causal direction between fault-related features and strongly correlated feature combinations in potential fault-related patterns is determined (e.g., whether "R1 temperature rise" is a precursor to "R2 vibration mutation"); timeline backtracking is performed based on the historical fault case database to verify the reproducibility of the causal relationship; at the same time, the consistency between the causal direction and fault type matching between fault-related features and strongly correlated feature combinations in potential fault-related patterns is checked, and causal feature pairs that simultaneously meet the criteria of co-occurrence frequency higher than the historical average, causal direction consistent with the physical mechanism of airborne equipment, reproducibility, and matching consistency are selected (e.g., "R1 temperature continues to rise + R2 vibration mutation"); the causal direction and causal feature pairs are integrated to obtain the fault triggering rule set.

[0055] It should also be noted that the specific operational steps for verifying the reproducibility of causal relationships are as follows: First, select all historical failure cases involving failure-related characteristics and of the same failure type from the historical failure case database, and arrange them in chronological order to form a case time series. Extract the spatiotemporal context features of all historical failure cases in the case time series, and construct a case feature timeline by extracting a fixed-length analysis window (e.g., 4 hours before the failure) based on the failure occurrence timestamp. Then, align the case feature timelines in time, using the failure occurrence time as the endpoint, and trace the occurrence times of the long-term trend and short-term fluctuations of the time-series dynamic decoupling in the failure-related characteristics. Use the long-term trend of time-series dynamic decoupling as the causal feature and the short-term fluctuations as the result feature. Statistically calculate the percentage of cases with both causal and result features in the same historical failure case, thus completing the verification of the reproducibility of causal relationships.

[0056] S5. Perform simulation based on the fault triggering rule set, generate simulation data, and perform deviation quantification analysis between the reliability prediction report and the simulation data to generate a verification report.

[0057] S5.1 Analyze the feature combination and triggering conditions of the fault triggering rule set, obtain the rule execution elements, and perform numerical mapping and multi-dimensional scenario simulation on the rule execution elements to generate simulation data; It should be noted that feature combinations (such as "R1 temperature continues to rise + R2 vibration sudden change") and corresponding triggering conditions (such as "average temperature > 80℃ and vibration peak value > 10mm / s") are extracted from the fault triggering rule set and mapped to physical state features and abnormal state features. Based on the baseline occurrence frequency and operating parameters of strongly correlated feature combinations, multiple operating scenarios (such as "high load - high temperature" and "low load - low temperature") are divided to perform multi-dimensional scenario simulation and generate simulation data containing physical state features, abnormal state features and operating parameters.

[0058] S5.2 Dynamically map the reliability prediction report with the simulation data to construct the prediction simulation correlation matrix, and perform deviation quantification analysis on the prediction simulation correlation matrix to generate a verification report.

[0059] It should be noted that the time steps of the reliability prediction report and simulation data are aligned using the IEEE Precise Timestamp Protocol to ensure a one-to-one correspondence between the prediction and simulation time points. A prediction-simulation correlation matrix is ​​constructed according to spatial partitioning and feature dimensions, with matrix elements representing the correspondence between the predicted values ​​in the reliability prediction report and the simulated values ​​in the simulation data. Based on the deviation between the predicted and simulated values ​​and the distribution of operating parameters, the deviation distribution pattern under different operating conditions is statistically analyzed (e.g., the deviation value is 0.12 under high load). The deviation analysis results are then integrated to generate a verification report.

[0060] This embodiment also provides a computer device applicable to the airborne equipment reliability prediction method based on field data, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the airborne equipment reliability prediction method based on field data as proposed in the above embodiment. The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0061] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the airborne equipment reliability prediction method based on field data as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0062] In summary, this invention achieves early identification of pre-fault symptoms, associated causes, and triggering relationships by constructing a fault triggering rule set. This elevates airborne equipment reliability prediction from simply outputting results to providing a forward-looking analysis of the fault formation process, improving the interpretability of prediction results, enhancing early warning capabilities, and increasing prediction stability under complex operating conditions. Furthermore, by generating simulation data and performing deviation quantification verification on the reliability prediction results, this invention achieves cross-checking and consistency confirmation of prediction conclusions, providing verifiable comparison and correction criteria for the prediction results, thereby improving the credibility of prediction conclusions, reducing the risk of false alarms and missed alarms, and enhancing the reliability of engineering applications.

[0063] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the reliability of airborne equipment based on field data, characterized in that: include, Collect multi-source field data from airborne equipment and preprocess it to extract the spatiotemporal context features of the preprocessed multi-source field data. The spatiotemporal context features are matched and analyzed with a historical failure case database to generate potential failure labels; The potential fault labels are input into the device prediction model, the feature enhancement layer performs deep feature extraction, the prediction decision layer performs spatiotemporal attention fusion, and a reliability prediction report is generated. Extract the preceding features from the historical fault case library, and combine them with spatiotemporal context features to perform correlation analysis on the preceding features to obtain the fault triggering rule set; Simulation is performed based on the fault triggering rule set to generate simulation data. The reliability prediction report and the simulation data are then subjected to deviation quantification analysis to generate a verification report.

2. The airborne equipment reliability prediction method based on field data as described in claim 1, characterized in that: The multi-source field data includes equipment physical parameters, equipment operating parameters, and maintenance logs; The preprocessing includes data cleaning, format conversion, deduplication, normalization, and outlier handling.

3. The airborne equipment reliability prediction method based on field data as described in claim 2, characterized in that: The specific steps for extracting spatiotemporal context features are as follows. Sliding window statistics are performed on the physical parameters of the equipment to obtain physical state characteristics; The detection equipment's operational parameters undergo time-series mutation analysis to obtain abnormal state characteristics; Analyze the failure frequency in the maintenance logs to obtain maintenance frequency characteristics; The physical state characteristics, anomaly characteristics, and maintenance frequency characteristics are structurally integrated to generate spatiotemporal context characteristics.

4. The airborne equipment reliability prediction method based on field data as described in claim 3, characterized in that: The specific steps for constructing the historical fault case database are as follows. Spatiotemporal context features are spatiotemporally aligned to construct a spatiotemporal correlation matrix, and correlation mining is performed on the spatiotemporal correlation matrix to generate potential fault correlation patterns. Cross-validate potential fault association patterns and maintenance frequency characteristics, and perform structured processing to generate a historical fault case library.

5. The airborne equipment reliability prediction method based on field data as described in claim 4, characterized in that: The specific steps for generating potential fault labels are as follows: Dynamic window similarity matching is performed between spatiotemporal context features and a historical failure case database to generate a candidate failure mode set; The co-occurrence frequency of the candidate fault mode set is statistically analyzed and the operating condition consistency calibration is performed to obtain potential fault labels.

6. The airborne equipment reliability prediction method based on field data as described in claim 5, characterized in that: The specific construction steps of the device prediction model are as follows. Call and initialize the residual convolutional network and multi-head Transformer to build the feature enhancement layer and prediction decision layer; A device prediction model is constructed by performing backpropagation and gradient update on the feature enhancement layer and the prediction decision layer using the focus loss function.

7. The airborne equipment reliability prediction method based on field data as described in claim 6, characterized in that: The specific steps for generating the reliability prediction report are as follows: The potential fault labels are input into the device prediction model, and the feature enhancement layer performs deep feature extraction through residual blocks to obtain high-order temporal features. The predictive decision-making layer uses a multi-head self-attention mechanism to perform spatiotemporal attention fusion and generate a risk score; High-order time-series features and risk scores are integrated through feature channels to generate a reliability prediction report.

8. The airborne equipment reliability prediction method based on field data as described in claim 7, characterized in that: The specific steps for extracting the pre-features from the historical fault case database are as follows: Reverse tracing and temporal pattern matching are performed on the spatiotemporal context features in the historical fault case library to form a feature candidate pool; The synergy between the matching feature candidate pool and the potential fault association patterns is used to obtain the preliminary features.

9. The airborne equipment reliability prediction method based on field data as described in claim 8, characterized in that: The specific steps for obtaining the fault triggering rule set are as follows: Spatial topological embedding and temporal dynamic decoupling are performed on the pre-features and spatiotemporal context features to obtain fault association features; Based on potential fault association patterns, causal rules of fault association features are extracted and causal inference verification is performed to obtain a fault triggering rule set.

10. The airborne equipment reliability prediction method based on field data as described in claim 9, characterized in that: The specific steps for generating the verification report are as follows: The feature combination and triggering conditions of the fault triggering rule set are analyzed to obtain the rule execution elements. The rule execution elements are then numerically mapped and multi-dimensional scenario simulations are performed to generate simulation data. The reliability prediction report is dynamically mapped to the simulation data to construct the prediction-simulation correlation matrix. The deviation of the prediction-simulation correlation matrix is ​​then quantitatively analyzed to generate a verification report.

Images