Lightweight interpretable aerial equipment abnormal state perception method and system

Abstract

The application discloses a kind of lightweight explainable aviation equipment abnormal state perception method and system, belong to aircraft health management and deep learning technical field.The application constructs 1+1 branch's LiteInception light weight feature extraction network, designs and aligns two-stage cascaded diagnosis process with aviation maintenance process: first stage adopts LiteInception and hybrid architecture of Transformer to execute fault detection binary classification, and priority is guaranteed recall rate;Second stage adopts pure LiteInception architecture to execute fault identification multi-classification to abnormal sample, and priority is guaranteed accuracy and explainability.The application significantly compresses model size, adapts edge device real-time inference needs, can output the diagnosis evidence chain that can be traced to specific sensor, specific time period, considers diagnosis efficiency, precision and compliance, and provides complete solution for general aircraft fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of aircraft health management and deep learning technology, specifically relating to a lightweight and interpretable method and system for perceiving abnormal states of aircraft equipment, and particularly to intelligent analysis technology of aircraft sensor data based on multivariate time series classification. Background Technology

[0002] As a core component of aircraft predictive and health management (PHM), fault diagnosis plays a vital role in reducing lifecycle maintenance costs and ensuring flight safety by proactively monitoring and assessing the condition of critical subsystems.

[0003] In recent years, deep learning methods have made significant progress in the fields of temporal analysis and fault diagnosis, with temporal classification deep learning models, such as InceptionTime, demonstrating excellent accuracy. Existing methods mainly employ a single-stage end-to-end design, mixing normal samples with various fault samples for multi-class classification. In terms of lightweighting, methods such as MobileNet deep separable convolution, ShuffleNet feature channel shuffling, and EfficientNet neural architecture search have been widely used in the image domain. However, existing technologies still have the following core shortcomings: The resource constraints of edge deployment are a significant issue: general aviation aircraft need to complete diagnostic decisions within their ground turnaround time (typically 30 to 60 minutes), but the computing power of airborne embedded devices or ground edge terminals is far lower than that of server-grade GPUs. InceptionTime employs an integrated strategy of five networks and includes a multi-branch parallel structure (3+1 branches) with three convolutional branches and one pooling branch, resulting in millions of parameters and billions of floating-point operations, making it difficult to meet real-time inference requirements on edge devices. This is because the original architecture prioritized accuracy in its design, neglecting the resource constraints of general aviation edge devices.

[0004] There is a disconnect between actual maintenance processes and model design: Real-world aviation maintenance follows a hierarchical process of monitoring, alarming, diagnosis, and decision-making, while existing deep learning methods generally adopt a single-stage end-to-end design. This leads to: when normal samples account for more than 95% in actual operation, all samples need to undergo complete multi-class inference, resulting in a large amount of unnecessary computational overhead; the severe asymmetry between the costs of missed fault detection and false alarms makes it difficult for a single model to be effectively customized for two different optimization objectives simultaneously. The reason is that existing methods directly apply machine learning paradigms to aviation diagnostic scenarios without aligning them with actual maintenance decision-making processes.

[0005] High-dimensional sensor data contains redundancy and noise interference: General aviation aircraft are equipped with dozens of sensor channels, covering heterogeneous signals with different physical dimensions such as temperature, pressure, and electrical parameters. However, not all channels contribute equally to fault diagnosis. Existing automated feature selection methods each have their shortcomings (e.g., mutual information methods ignore nonlinear relationships, gradient methods depend on specific models, and attention methods are affected by optimization preferences). Using any one of them alone is unlikely to obtain robust and reliable channel selection results.

[0006] The need for interpretability has constrained the deployment of deep learning in the field of aviation safety: aviation maintenance personnel need to know "which sensor" malfunctioned "at what time" in order to efficiently locate the root cause of the failure, and black-box prediction results are difficult to meet the compliance requirements of airworthiness certification. Existing methods lack the ability to provide a traceable chain of evidence from both the sensor channel and time dimensions.

[0007] There is a lack of flexible adjustment mechanisms between precision and recall: safety-critical scenarios require high recall to avoid false negatives, while diagnostic aids require high precision to reduce unnecessary re-examination workload. Existing methods typically use fixed training strategies to obtain a single model, lacking a flexible means to balance the two.

[0008] Existing lightweight methods suffer from structural differences with time-series data: the channel independence assumption of depthwise separable convolution in the image domain severs the physical coupling between multi-channel sensor signals in aircraft, resulting in limited direct transfer effectiveness. In the time-series domain, a systematic approach to pruning multi-branch parallel structures like InceptionTime at the inter-branch topology level has not yet been explored. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a lightweight and interpretable method and system for perceiving abnormal states of aviation equipment. While maintaining controllable loss of diagnostic accuracy, it achieves model lightweighting to adapt to the deployment of edge devices in general aviation aircraft, deeply aligns with aviation maintenance standard procedures, and provides a traceable chain of fault diagnosis evidence to meet airworthiness compliance requirements and deployment needs in different scenarios.

[0010] This invention provides a lightweight and interpretable method for detecting abnormal states in aviation equipment, used for fault diagnosis of time-series data collected by multi-channel sensors on general aviation aircraft, characterized by comprising: A lightweight feature extraction network, LiteInception, is constructed. The basic module of the LiteInception lightweight feature extraction network is a 1+1 parallel structure containing one standard one-dimensional convolutional branch and one max pooling branch. The standard one-dimensional convolutional branch first compresses the feature channel dimension through a 1×1 bottleneck convolution, and then extracts temporal trend features through one-dimensional convolution. The max pooling branch first performs max pooling with a kernel size of 3 and a stride of 1, and then adjusts the number of feature channels through a 1×1 convolution to extract temporal extreme value features. The outputs of the two branches are concatenated and fused along the feature channel dimension. The complete network is composed of multiple stacked basic modules, and residual connections are set between the basic modules. A two-stage cascaded diagnostic process is constructed. The first stage uses a hybrid architecture of LiteInception and Transformer encoder to perform fault detection binary classification, with the optimization goal of maximizing recall. The second stage uses a pure LiteInception architecture to perform fault identification multi-class classification, with the optimization goal of high classification accuracy and interpretability. The second stage inference is triggered only when the first stage determines that the input sample is abnormal.

[0011] Therefore, through the 1+1 branch LiteInception lightweight architecture, InceptionTime is systematically pruned at the branch topology level. While maintaining the full channel interaction capability of standard convolution and avoiding the disruption of the physical coupling relationship of multi-channel sensor signals, the number of model parameters is compressed from 2159K to 639K (compression of 70.4%), and the computational cost is reduced from 8.77G FLOPs to 2.58G. FLOPs (70.6% compression), CPU inference speed accelerated from 52.37ms to 6.35ms (8.2x speedup), peak memory reduced from 128MB to 89MB (30% reduction), performance loss controlled within 3%, meeting the real-time inference needs of general aviation edge devices; the two-stage cascaded diagnostic process is naturally aligned with the standard aviation maintenance monitoring, alarm, diagnosis, and decision-making process, decoupling the two cost-asymmetric optimization goals of fault detection and fault identification. Normal samples are screened out in the first stage, eliminating the need for the second stage calculation, realizing on-demand allocation of computing resources. At the same time, the models in each stage can be updated independently, and when a new fault type is added, only the second stage model needs to be retrained, significantly reducing maintenance costs.

[0012] This includes a multi-method fusion sensor channel selection step: using three methods—mutual information evaluation, gradient analysis evaluation, and compressed excitation attention weight evaluation—to independently score the fault diagnosis contribution of each channel in the original multi-channel sensor data; and using cross-validation based on multi-method consistency criteria, physical causality criteria, and information redundancy criteria to select a subset of high-value sensor channels.

[0013] Therefore, the multi-method fusion sensor channel selection strategy integrates three complementary perspectives: statistical correlation, model sensitivity, and optimization preference. It can robustly and reliably screen out a subset of high-value channels from high-dimensional heterogeneous sensor data, effectively eliminating redundancy and noise interference. After reducing the sensor channels from 23 dimensions to 15 dimensions, the model Macro F1 improved from 69.58% to 78.23%, and the accuracy improved from 69.69% to 77.58%, achieving an improvement in diagnostic performance without a decrease.

[0014] This also includes a model training step based on knowledge distillation: using a pre-trained multi-branch InceptionTime as the teacher model and the LiteInception lightweight feature extraction network as the student model, the precision and recall of the model are adjusted by switching training modes; the training modes include a direct training mode that uses only hard-label cross-entropy loss, and a distillation training mode that uses a hybrid loss that combines hard-label cross-entropy loss and soft-label KL divergence.

[0015] Therefore, by employing a knowledge distillation training strategy, the same lightweight model can flexibly adjust between precision and recall simply by switching training methods, adapting to different deployment scenarios. After distillation training, the model recall is significantly improved by approximately 2.4 percentage points, and the precision-recall gap is reduced from 4.46 percentage points to 0.55 percentage points. The direct training mode is suitable for assisted diagnostic scenarios requiring manual review, achieving high precision to reduce unnecessary re-examination workload. The distillation training mode is suitable for safety-critical early warning scenarios where the cost of missed detections far outweighs that of false positives, achieving high recall to avoid missed fault detections.

[0016] This includes a two-layer interpretability analysis step: integrating four attribution methods—input gradient method, occlusion sensitivity analysis, gradient-weighted class activation mapping, and integral gradient method—to perform a two-layer attribution analysis on the diagnostic results; summing along the time dimension to obtain the importance vector of each sensor channel, and taking the intersection of the top-ranked channels from multiple attribution methods to determine the key faulty sensor; summing along the sensor channel dimension to obtain the time importance curve to determine the critical time period of the anomaly.

[0017] Therefore, by integrating four attribution methods based on different principles, and performing two-layer attribution analysis and cross-validation in the sensor channel and time dimensions, a diagnostic report with evidence chain traceable to specific sensors and specific time periods can be output. This effectively solves the black box problem of deep learning models, provides aviation maintenance personnel with a basis for locating the root cause of failures, and meets the compliance requirements of airworthiness certification.

[0018] The complete network is composed of six basic modules stacked together, with residual connections set between every three basic modules.

[0019] Therefore, the configuration of stacking 6 basic modules and setting residual connections for every 3 modules provides the best cost performance for the model. While significantly compressing the model size, it can fully extract the deep features in the time series data, alleviate the gradient vanishing problem caused by stacking deep networks, and ensure the accuracy and performance of fault diagnosis and the stability of training.

[0020] The two-layer interpretability analysis step also includes noise perturbation robustness verification: applying Gaussian noise of multiple intensity levels to the correctly classified samples with high confidence, observing the changing trend of the attribution distribution as the model confidence decreases, and verifying the causal relationship between the attribution results and the model decision mechanism.

[0021] Therefore, by verifying the robustness of noise perturbation, the causal relationship between the attribution results and the model decision-making mechanism can be effectively verified, spurious attributions can be avoided, the credibility of interpretability analysis results can be further enhanced, and the reliability of the diagnostic evidence chain can be guaranteed.

[0022] This invention also provides a lightweight, interpretable anomalous state perception system for aviation equipment, used for fault diagnosis of time-series data collected by multi-channel sensors on general aviation aircraft, including: The LiteInception feature extraction module is based on a 1+1 parallel structure consisting of a standard one-dimensional convolutional branch and a max-pooling branch. The standard one-dimensional convolutional branch first compresses the feature channel dimension through a 1×1 bottleneck convolution, and then extracts temporal trend features through a one-dimensional convolution. The max-pooling branch first performs max pooling with a kernel size of 3 and a stride of 1, and then adjusts the number of feature channels through a 1×1 convolution to extract temporal extreme value features. The outputs of the two branches are concatenated and fused along the feature channel dimension. The complete network is composed of multiple stacked basic units, with residual connections between the basic units. The two-stage cascaded diagnostic engine includes a first-stage fault detection module and a second-stage fault identification module. The first-stage fault detection module adopts a hybrid architecture of LiteInception and Transformer encoder to perform binary classification of fault detection, with the optimization goal of maximizing recall. The second-stage fault identification module adopts a pure LiteInception architecture to perform multi-class classification of fault identification, with the optimization goal of high classification accuracy and interpretability. The input of the second-stage fault identification module is only connected to the abnormal output of the first-stage fault detection module.

[0023] Therefore, the LiteInception feature extraction module achieves lightweight and efficient extraction of temporal features, which can be adapted to the deployment and real-time inference requirements of edge devices in general aviation aircraft; the two-stage cascaded diagnostic engine decouples fault detection and fault identification, is deeply aligned with the aviation maintenance process, and only abnormal samples trigger the second stage of inference, which greatly reduces the overhead of invalid calculations. At the same time, the two modules can be independently iterated and updated, reducing the system's operation and maintenance upgrade costs.

[0024] The system also includes a sensor channel selection module, which uses three methods—mutual information evaluation, gradient analysis evaluation, and compressed excitation attention weight evaluation—to independently score the fault diagnosis contribution of each sensor channel. Through cross-validation using multi-method consistency criteria, physical causality criteria, and information redundancy criteria, a subset of high-value sensor channels is selected.

[0025] Therefore, the sensor channel selection module can robustly select a subset of high-value channels from the original high-dimensional multi-channel sensor data, effectively filter redundant information and noise interference, reduce the computational overhead of subsequent feature extraction and diagnosis processes, and improve the fault diagnosis accuracy of the model.

[0026] It also includes a knowledge distillation training module, which uses a pre-trained multi-branch InceptionTime as the teacher model and the network of the LiteInception feature extraction module as the student model. The precision and recall of the model are adjusted by switching training modes. The training modes include a direct training mode that uses only hard-label cross-entropy loss and a distillation training mode that uses a hybrid loss that combines hard-label cross-entropy loss and soft-label KL divergence.

[0027] Therefore, the knowledge distillation training module provides the system with flexible precision-recall adjustment capabilities, enabling the system to adapt to the high recall requirements of safety-critical early warning scenarios and the high precision requirements of auxiliary diagnosis scenarios by switching training modes, thereby improving the system's scenario adaptability.

[0028] It also includes an interpretability analysis module, which integrates four attribution methods: input gradient method, occlusion sensitivity analysis, gradient weighted class activation mapping, and integral gradient method, to perform a two-layer attribution analysis on the diagnostic results; it sums along the time dimension to obtain the importance vector of each sensor channel, and takes the intersection of the top-ranked channels of multiple attribution methods to determine the key fault sensors; it sums along the sensor channel dimension to obtain the time importance curve, and determines the critical time period of the anomaly.

[0029] Therefore, the interpretability analysis module can generate a fault diagnosis evidence chain that can be traced back to a specific sensor and a specific time period, solving the black box problem of deep learning models, providing maintenance personnel with a clear basis for fault root cause location, and meeting the compliance requirements of airworthiness certification in the aviation field. Attached Figure Description

[0030] Figure 1 A flowchart illustrating the lightweight and interpretable anomalous state perception method for aviation equipment according to an embodiment of the present invention is shown. Detailed Implementation

[0031] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the following description, the same reference numerals are used for the same parts, and repeated descriptions are omitted. Furthermore, the drawings are merely schematic diagrams, and the proportions of the parts or the shapes of the parts may differ from the actual figures.

[0032] like Figure 1 As shown, this invention provides a lightweight and interpretable general aviation aircraft fault diagnosis method based on LiteInception, used for fault diagnosis of time-series data collected by multi-channel sensors on general aviation aircraft. It addresses the core technical problems in existing technologies, such as difficulty in adapting models to edge device deployment, disconnection from aviation maintenance processes, redundant interference from sensor data, insufficient interpretability, and difficulty in flexibly adjusting precision and recall. The core technologies of this invention work in deep synergy to form a complete technical closed loop: multi-method fusion sensor channel selection filters redundant noise from the data source, providing high-quality input for subsequent feature extraction; the LiteInception lightweight architecture provides a unified, edge-device-compatible feature extraction foundation for two-stage cascaded diagnosis; the two-stage cascaded diagnosis process fully leverages the advantages of the lightweight architecture, achieving on-demand allocation of computing resources and decoupling of optimization objectives; the knowledge distillation training strategy enables flexible adjustment of model performance without changing the network structure; and the two-layer interpretability analysis provides a compliant and traceable chain of evidence for the diagnostic results, while simultaneously forming cross-validation with the channel selection step, further improving the accuracy of fault location.

[0033] The complete technical process of this invention is executed sequentially according to the following links, with each link working in deep collaboration to form a closed loop: The first step is to perform preprocessing on the raw multi-channel sensor time series data, which includes missing value removal, linear interpolation, resampling to 2048 time steps, and Min-Max normalization. The second step involves using a multi-method fusion sensor channel selection strategy to perform channel filtering on the preprocessed data, selecting 15 high-value diagnostic channels from the original 23 sensor channels. The third step is to perform feature extraction on the filtered time series data through the 1+1 branch LiteInception lightweight architecture. This architecture consists of 6 layers of LiteInception modules stacked together, with residual connections set between modules to achieve lightweight and efficient extraction of time series features. The fourth step is to perform a two-stage cascaded diagnosis. The first stage uses a LiteInception+Transformer hybrid architecture to complete the binary classification of fault detection. Only when a sample is determined to be abnormal is the second stage of multi-class identification of fault types using a pure LiteInception architecture triggered. The fifth step involves using a knowledge distillation-based training strategy to flexibly adjust the model's precision and recall, allowing for switching between high-precision auxiliary diagnostic scenarios and high-recall safety-critical early warning scenarios. The sixth step involves performing attribution analysis on the diagnostic results using a two-layer interpretability analysis framework, ultimately outputting a diagnostic report containing the chain of evidence including the fault type, key fault sensors, and abnormal time periods.

[0034] In this embodiment, InceptionTime is the benchmark deep learning model in the field of time series classification. Its core employs a multi-branch parallel Inception module design, with time series data classification accuracy as the core optimization objective. It serves as the benchmark architecture for the teacher model used in knowledge distillation in this invention. Multi-branch InceptionTime refers to the 3+1 parallel basic module structure (3 standard one-dimensional convolutional branches + 1 max-pooling branch) used in the original InceptionTime model, along with the ensemble inference strategy consisting of 5 independently trained isomorphic InceptionTime networks. Pre-trained multi-branch InceptionTime refers to the multi-branch InceptionTime model that has undergone end-to-end supervised training and has all parameters frozen on the target general aviation aircraft fault diagnosis dataset, and is used as the teacher model for knowledge distillation in this invention.

[0035] In this embodiment, the pre-trained multi-branch InceptionTime is the teacher model in the knowledge distillation-based model training step of this invention. It is a benchmark high-precision model in the temporal classification field, used to provide a fine-grained feature learning benchmark for the LiteInception lightweight student model. This allows for flexible adjustment of the student model's precision and recall without changing its network structure. This model only participates in loss calculation during the training phase and does not participate in the final inference deployment stage of this invention.

[0036] In this implementation, the core basic unit of the multi-branch InceptionTime is a 3+1 parallel structure consisting of 3 standard one-dimensional convolutional branches and 1 max-pooling branch. The complete network is composed of multiple stacked basic modules, and finally, 5 independently trained isomorphic InceptionTime networks are integrated to output the final classification result. The number of parameters of a single 3+1 branch basic module is approximately 135K, the total number of parameters of the complete integrated model is approximately 2159K, and the total computation is approximately 8.77 GFLOPs.

[0037] Furthermore, through branch ablation experiments, it was verified that the 3+1 branch structure is the benchmark optimal structure of the original InceptionTime: reducing the model's F1 score from 3+1 branches to 2+1 branches only reduces the model's F1 score by about 0.80 percentage points; the 1+1 branch LiteInception architecture adopted in this invention can maintain about 96.5% of the F1 performance with less than 30% of the parameters of the benchmark model.

[0038] In this embodiment, the pre-training of the multi-branch InceptionTime dataset is performed using a general aviation airborne sensor dataset that is completely consistent with the fault diagnosis task of this invention. Taking the NGAFID aviation maintenance dataset as an example, the complete pre-training process is as follows: (1) Dataset preprocessing: The preprocessing steps are completely consistent with the diagnostic process of this invention: First, samples with a missing value ratio exceeding 20% ​​and samples with a time series length of less than 2000 time steps are removed; Second, NaN data points in the remaining samples are filled using linear interpolation; Third, all time series are uniformly resampled to 2048 time steps; Fourth, the numerical range of each sensor channel is standardized to the [0, 1] interval using Min-Max normalization. After preprocessing, the dataset is divided into training set and test set at a ratio of 80% / 20%.

[0039] (2) Training optimization objective: The end-to-end supervised training method is adopted, the optimization objective is standard cross-entropy loss, and the training task is aircraft fault multi-classification task, which is completely consistent with the classification category of the fault identification task in the second stage of this invention.

[0040] (3) Training completion criteria: The model's classification accuracy on the test set converges to a stable range. After training is completed, all network parameters of the model are frozen and will not be updated during the entire training process.

[0041] In this implementation, the use of pre-trained multi-branch InceptionTime fully follows the knowledge distillation teacher-student architecture rules, and is divided into two core stages: the training stage and the inference and deployment stage. The complete process is as follows: During the training phase, the core objective is to guide the lightweight student model to learn the timing features of aircraft failures using a pre-trained, high-precision teacher model. Only the student model weights are updated throughout the training phase, while the teacher model parameters remain frozen. The specific steps are as follows: (1) Model initialization and pre-processing: Complete the parameter freezing of the pre-trained multi-branch InceptionTime teacher model, and initialize the LiteInception lightweight student model proposed in this invention, fixing the optimal distillation hyperparameters verified by grid search: distillation weights = 0.7, temperature coefficient = 4.0.

[0042] (2) Dual-path synchronous forward propagation: Inputting identical time-series data from general aviation aircraft onboard multi-channel sensors, forward propagation is performed separately for the teacher model and the student model: The teacher model outputs data after temperature coefficient... The softened classification probability distribution provides soft label information on the fine-grained similarity between different fault categories; the student model synchronously outputs its own classification probability distribution for subsequent bi-branch loss calculation.

[0043] (3) Two-branch loss calculation and total loss fusion: Based on the forward propagation results, two independent loss components are calculated separately, and then fused according to weights to obtain the total loss. The formula for calculating the total loss is as follows:

[0044] in, The cross-entropy loss is calculated as the standard classification loss between the student model output and the true hard label of the sample (one-hot encoding of the fault category), which is used to constrain the basic classification accuracy of the student model. for KL Divergence loss calculates the difference between the softened probability distributions of the teacher model and the student model, used to guide the student model in learning the fine-grained feature knowledge mastered by the teacher model and smoothing the model's decision boundary. It is applied according to preset distillation weights. =0.7, the two loss components are weighted and fused to obtain the final total loss used for model iteration.

[0045] (4) Backpropagation and weight update: Based on the total loss calculation result, the backpropagation algorithm is executed. Only the network weights of the LiteInception student model are updated. The parameters of the pre-trained InceptionTime teacher model are frozen throughout the process and do not participate in the weight update. The iterative process of forward propagation, loss calculation and backpropagation is repeated until the student model converges and the distillation training is completed.

[0046] In this implementation, the inference deployment phase is the application deployment stage after training is completed, and the core logic is as follows: (1) Deployment model: Only the trained LiteInception lightweight student model is deployed, without the need to deploy the pre-trained InceptionTime teacher model, which greatly reduces the computing power and memory requirements of deployment and adapts to the resource constraints of general aviation airborne embedded devices and ground edge terminals.

[0047] (2) Reasoning task: Real-time fault diagnosis reasoning is performed on the complete flight phase airborne multi-channel sensor time series data collected after the flight ends. Two modes can be flexibly switched: the safety critical early warning scenario adopts the model version trained by distillation, which prioritizes high recall rate and follows the aviation safety principle of better to false alarm than to miss detection; the auxiliary diagnosis scenario adopts the model version trained directly with only hard labels, which prioritizes high accuracy rate and reduces unnecessary manual re-inspection workload.

[0048] (3) Output results: Output real-time fault diagnosis results. Combined with the two-stage cascaded diagnosis process of this invention, the final output includes a complete diagnosis conclusion containing fault type, key fault sensors and abnormal time window, supporting aircraft ground maintenance decisions.

[0049] In this embodiment, the benchmark performance of the pre-trained multi-branch InceptionTime model used in this invention is as follows: the total number of parameters of the complete model is approximately 2159K, the total computation is approximately 8.77G FLOPs, the CPU single-sample inference time is approximately 52.37ms, and the peak memory usage is approximately 128MB. On the NGAFID general aviation aircraft fault diagnosis dataset, this model, as a teacher model, can improve the recall of the LiteInception student model after distillation training by approximately 2.4 percentage points, reduce the gap between precision and recall from 4.46 percentage points to 0.55 percentage points, and achieve an optimal F1 score of 75.67% for the student model after distillation. The standard deviation of the F1 score in 10 repeated experiments is only 0.42%, demonstrating extremely high stability.

[0050] In this implementation, high-quality input is provided for subsequent feature extraction and diagnosis, forming a synergistic optimization with the lightweight architecture. The sensor channel selection step, which integrates multiple methods, employs three complementary methods: mutual information evaluation, gradient analysis evaluation, and compression excitation (SE) attention weight evaluation. These methods independently score the fault diagnosis contribution of each channel in the original multi-channel sensor data, comprehensively evaluating the diagnostic value of each sensor channel from three different perspectives: statistical correlation, model sensitivity, and optimization preference. The specific implementation method is as follows: The complete execution architecture for this step is as follows: The input is preprocessed 23-channel multi-sensor time-series data, covering eight categories of physical quantities: electrical, fuel, lubricating oil, engine speed, cylinder head temperature, exhaust temperature, ambient temperature, and flight status; three parallel independent importance assessments are executed simultaneously, with each assessment independent of the others, and each outputting a ranking of the importance of all channels. The first path is mutual information evaluation, which extracts the 7-dimensional statistical features of the time series data of each channel, calculates the mutual information score between the statistical features and the fault category label, and completes the channel ranking; The second approach is gradient analysis and evaluation. Based on the pre-trained InceptionTime model, it calculates the absolute value of the partial derivative of the cross-entropy loss with respect to the input of each channel, and takes the average value in the batch dimension and time dimension to obtain the gradient importance score, thus completing the channel ranking. The third approach is the SE attention weight evaluation. A compression-activation (SE) module is embedded at the input of the InceptionTime model. Compression, activation, and scaling operations are performed sequentially to learn the attention weight coefficients of each channel and complete the channel ranking. After the three-way evaluation was completed, cross-validation and fusion decision-making were performed using the multi-method consistency criterion, physical causality criterion, and information redundancy criterion. Channels that ranked highly in all three channels, had a direct physical causal relationship with the fault, and had no information redundancy were retained. Finally, 8 low-value / redundant channels were eliminated, and 15 high-value diagnostic channels were selected, covering five major categories of core physical quantities: electrical system, fuel flow, lubricating oil system, cylinder head temperature, and exhaust temperature. After the selection, the diagnostic performance of the model improved rather than decreased.

[0051] In this embodiment, the sensor channel selection for multi-method fusion adopts a three-way parallel independent evaluation + cross-validation fusion decision-making process, and the complete execution logic is as follows: The input to the process is preprocessed multi-channel sensor time-series data in tensor format. ,in B For sample batch dimensions, T For time step dimension, C For sensor channel dimensions, taking the NGAFID dataset as an example, the original number of input channels... C= 23, covering eight categories of physical quantities: electrical system, fuel system, lubrication system, engine speed, cylinder temperature, exhaust temperature, environmental parameters, and flight status parameters.

[0052] For the 23 input sensor channels, three independent importance assessments are performed simultaneously: the first is a mutual information assessment based on statistical correlation, the second is a gradient analysis assessment based on model sensitivity, and the third is a compression incentive (SE) attention weight assessment based on model optimization preference. The three assessments output the independent importance score and the overall ranking results of each sensor channel, without interfering with each other.

[0053] After the three-way evaluation is completed, cross-validation and fusion decision-making are carried out through three rules: multi-method consistency criterion, physical causality criterion, and information redundancy criterion. Channels that rank highly in all three evaluations are considered to have robust diagnostic value and are retained. Channels that rank highly in only one path and perform poorly in the other two paths, or channels that have no reasonable physical causal relationship with the fault and have high information redundancy, are included in the exclusion candidate set.

[0054] After fusion decision-making, channel screening and simplification are completed, and a subset of high-value sensor channels is output. Taking the NGAFID dataset as an example, 8 low-value / redundant channels are finally excluded, and 15 core diagnostic channels are retained. The filtered channel subset only covers five categories of physical quantities that are directly related to the health status of the aircraft's core equipment: electrical system, fuel flow, lubrication system, cylinder temperature, and exhaust temperature. Redundant channels that reflect flight mission characteristics, external environmental parameters, and have no direct causal relationship with faults are eliminated.

[0055] The selected high-value channel subset will be used as input to the subsequent LiteInception lightweight feature extraction network, thereby reducing computational overhead and improving diagnostic performance.

[0056] Mutual information assessment: From a statistical correlation perspective, the degree of correlation between each sensor channel and the fault category is evaluated. The specific implementation steps are as follows: (1) Statistical feature extraction: For the time series data of each sensor channel, a 7-dimensional statistical feature vector is extracted. The 7-dimensional statistical features include mean, standard deviation, maximum value, minimum value, median, skewness and kurtosis.

[0057] (2) Mutual information score calculation: Calculate the mutual information score between the 7-dimensional statistical features extracted from each sensor channel and the corresponding fault category label of the sample. The higher the mutual information score, the stronger the correlation between the statistical features of the channel and the fault category, and the higher the contribution to fault diagnosis.

[0058] (3) Channel ranking: All sensor channels are ranked from high to low according to the calculated mutual information scores to obtain the channel importance ranking from the perspective of mutual information evaluation.

[0059] Gradient analysis and evaluation: From the perspective of model sensitivity, this evaluates the impact of changes in the input of each sensor channel on the model's diagnostic results. The specific implementation steps are as follows: (1) Basic model preparation: The InceptionTime model, which was pre-trained on the target general aviation aircraft fault diagnosis dataset, was used as the benchmark model for gradient calculation.

[0060] (2) Gradient calculation: Input all samples of the training set into the pre-trained InceptionTime model, perform forward propagation and backward propagation in sequence, and calculate the absolute value of the partial derivative of the cross-entropy loss function with respect to the input of each sensor channel.

[0061] (3) Importance score calculation: The absolute value of the calculated partial derivative is averaged in the batch dimension (all samples) and the time dimension (all time steps) to obtain the gradient importance score of the corresponding sensor channel. The higher the score, the greater the impact of the input change of the channel on the model diagnosis result and the higher the contribution to fault diagnosis.

[0062] (4) Channel ranking: All sensor channels are ranked from high to low according to the calculated gradient importance score to obtain the channel importance ranking from the perspective of gradient analysis evaluation.

[0063] SE Attention Weight Evaluation: From the perspective of model optimization preferences, this evaluates the dependence of the model's diagnostic decisions on the features of each sensor channel. The specific implementation steps are as follows: (1) Module embedding and model preparation: The compression stimulus (SE) module is embedded at the input of the InceptionTime model, and the dimensionality reduction ratio of the module is set to 4. The compression stimulus module includes three core components: compression operation, stimulus operation, and scaling operation.

[0064] (2) Compression operation: Global average pooling is performed on the input multi-channel sensor feature map along the time dimension, and the complete time series of each channel is compressed into a global channel descriptor to achieve feature compression and aggregation.

[0065] (3) Activation operation: The weight coefficients of each channel are learned through two fully connected networks. The first fully connected network reduces the channel dimension to 1 / 4 of the original dimension. After activation by the ReLU activation function, the second fully connected network restores the channel dimension to the original dimension and outputs the channel weights between 0 and 1 through the Sigmoid activation function.

[0066] (4) Weight scaling and score calculation: The channel weights output by the excitation operation are linearly transformed to the interval [0.5, 1.5] to obtain the final attention weight score of each sensor channel; a weight greater than 1 means that the model actively amplifies the feature response of the channel and contributes more to fault diagnosis; a weight less than 1 means that the model suppresses the feature of the channel and contributes less to fault diagnosis.

[0067] (5) Channel ranking: All sensor channels are ranked from high to low according to the calculated attention weight scores to obtain the channel importance ranking from the perspective of compressed excitation attention weight evaluation.

[0068] After independent scoring using the three methods, a subset of high-value sensor channels was selected through cross-validation using three criteria: Multi-method consistency criterion: If a sensor channel ranks highly in all three methods, it is considered to have robust diagnostic value and is retained; if a channel ranks highly only in a single method but performs poorly in the other two methods, the channel is excluded. Physical causality criterion: Eliminate pathways that lack a reasonable physical causal relationship with the failure, including parameters that reflect flight mission characteristics rather than equipment health status, and parameters whose causal direction is the result of the failure rather than the cause of the failure; Information redundancy criterion: For two sensor channels carrying highly relevant information, prioritize retaining the channel that is more directly related to the fault mechanism and has more information, and eliminate redundant channels.

[0069] Taking the NGAFID aviation maintenance dataset as an example, this data comes from real flight records of a Cessna 172 aircraft during normal operation. The raw data contains 23 sensor channels with a sampling frequency of 1Hz, covering the electrical system (main current Amp1, auxiliary current Amp2, main voltage Volt1, critical voltage Volt2), fuel system (left remaining fuel FuelQtyL, right remaining fuel FuelQtyR, fuel flow FFlow), lubrication system (oil temperature OilT, oil pressure OilP), engine speed (RPM), cylinder temperature (CHT1 to CHT4), exhaust temperature (EGT1 to EGT4), environmental parameters (outdoor air temperature OAT), and flight status parameters (indicated airspeed IAS, vertical speed VS, altitude AltMSL, normal acceleration NormAc).

[0070] The preprocessed dataset includes: 23 original sensor channels, 15 filtered sensor channels, 2048 time steps, 11446 total samples, 5844 samples after maintenance (normal), 5602 samples before maintenance (faulty), and 19 fault categories. The training and test sets are divided at 80% and 20% respectively. In the first stage, the data is divided into "normal" and "faulty" categories. In the second stage, categories with at least 50 flights before and after maintenance are selected from 36 maintenance problem types, retaining 19 categories. TimeWarp data augmentation (intensity 0.03) is used to balance the minority classes.

[0071] After the channel selection process described above, eight redundant channels (FuelQtyL, FuelQtyR, RPM, IAS, VS, AltMSL, NormAc, and OAT) with no direct causal relationship to the fault were excluded. Fifteen high-value channels covering five core functional subsystems—electrical system, fuel flow, lubrication system, cylinder temperature, and exhaust temperature—were retained. Validation results show that reducing the channel dimension from 23 to 15 improved the model's Macro F1 score from 69.58% to 78.23% and accuracy from 69.69% to 77.58%, achieving an improvement in diagnostic performance without sacrificing it, while further reducing the computational load of subsequent feature extraction and diagnostic processes.

[0072] In this embodiment, the LiteInception lightweight feature extraction network is the core foundation for the edge deployment of this invention.

[0073] The core innovation of the LiteInception lightweight feature extraction network constructed in this invention lies in simplifying the original InceptionTime's 3+1 multi-branch parallel structure into a 1+1 parallel structure with one standard one-dimensional convolutional branch and one max pooling branch. This achieves systematic pruning at the branch topology level, rather than simply transferring lightweight methods from the image domain, fundamentally avoiding the severing of the physical coupling relationship of multi-channel sensor signals.

[0074] The internal structure and data flow of a single LiteInception basic module are as follows, and the components of the module are connected in the following order: (1) Input feature map, which is the input data source for the module; (2) 1×1 bottleneck convolution, located at the input end of the standard one-dimensional convolution branch, performs channel dimension compression on the input feature map to reduce the computational cost of subsequent convolutions; (3) A one-dimensional convolution with a preset kernel size is connected to the output of a 1×1 bottleneck convolution to perform convolution operations on the compressed feature map and extract temporal trend features; (4) Max pooling with a kernel size of 3 and a step size of 1 is located at the input end of the max pooling branch. It performs pooling processing on the input feature map and extracts temporal extreme value features. (5) 1×1 convolution, connected to the output of max pooling, adjusts the number of channels of the pooled feature map to match the output channel dimension of the convolution branch; (6) Channel splicing layer: The input end is connected to the output of one-dimensional convolution and the output of 1×1 convolution respectively, and the output features of the two branches are spliced ​​and fused along the feature channel dimension; (7) Batch normalization + ReLU activation layer, connected to the output of the channel splicing layer, performs batch normalization and nonlinear activation on the fused features; (8) Output feature map, which is the final output of the LiteInception basic module and can be input into subsequent module stacks or classification heads.

[0075] The complete LiteInception network consists of six basic modules stacked together. Residual connections are set between every three basic modules. The residual connections are matched with the main path by a 1×1 convolution and then added to the main path. This effectively alleviates the gradient vanishing problem caused by deep network stacking and ensures the model's feature extraction capability and training stability. Global average pooling and a fully connected classification head are used at the end of the network to achieve the final classification output.

[0076] Actual measurements show that a single LiteInception module has approximately 41K parameters, only 1 / 3.3 of the original InceptionTime 3+1 branch module; the total network has approximately 639K parameters, a 70.4% reduction compared to the original InceptionTime's 2159K. FLOPs decreased from 8.77G to 2.58G, a 70.6% reduction. The CPU single-sample inference time is only 6.35ms, an 8.2x speedup compared to the original model. Peak memory usage decreased from 128MB to 89MB, a 30% reduction, while performance loss was kept within 3%. This fully meets the requirements for diagnostic decisions within 30-60 minutes of general aviation aircraft ground turnaround time and the real-time inference needs of airborne / ground edge devices. Branch ablation experiments verify that changing from 3+1 to 2+1 only reduces F1 by approximately 0.80 percentage points; the 1+1 architecture maintains approximately 96.5% of the F1 performance with less than 30% of the baseline parameter count; removing the MaxPool branch (1+0 configuration) causes a sharp drop in F1 of approximately 5 percentage points, verifying the optimal cost-effectiveness of the 1+1 branch structure. In-depth network experiments have verified that a 6-layer configuration is the optimal cost-performance ratio.

[0077] In this implementation, the cascading of the two-stage diagnostic process is deeply integrated with the LiteInception lightweight architecture to achieve alignment with aviation maintenance processes.

[0078] The two-stage cascaded diagnostic architecture of this invention is fully aligned with aviation maintenance procedures. The overall architecture and execution logic are as follows: The diagnostic process is triggered by the completion of the flight and is automatically started based on the airborne sensor data of the entire flight phase. The deployment carrier is an airborne embedded device or a ground edge terminal, and the entire diagnostic process must be completed within the aircraft's ground turnaround period. The first stage is the fault detection stage, which performs a normal / abnormal binary classification task. It adopts a LiteInception + Transformer hybrid architecture, prioritizing the use of model weights trained by distillation. The core design goal is to maximize recall, adhering to the aviation safety principle of "better to have a false positive than a false negative," and only outputs two classification results: "normal" or "abnormal." The cascading trigger logic is as follows: if the sample is classified as normal in the first stage, a normal diagnosis result is directly output, and the subsequent process is terminated; if the sample is classified as abnormal in the first stage, the abnormal sample is input into the second stage, triggering the subsequent fault identification process. The second stage is the fault identification stage, which is triggered only when an anomaly is detected. It performs 19 types of fine-grained fault classification tasks and adopts a pure LiteInception architecture. The model weights can be switched between the direct training version and the distillation training version according to the scenario. The core design goal is high classification accuracy and interpretability, and the final output is interpretable fine-grained diagnostic results. The final diagnostic results output by the entire process include three core components: the fault type, the critical fault sensors, and the abnormal time window, which can directly support aircraft ground maintenance decisions.

[0079] The two-stage cascaded diagnostic process constructed in this invention is fully aligned with the hierarchical process of "monitoring, alarming, diagnosis, and decision-making" in general aviation standard maintenance. The core achievement is the decoupling of the two cost-asymmetric optimization objectives of fault detection and fault identification. At the same time, relying on the lightweight characteristics of LiteInception, it enables on-demand allocation of computing resources.

[0080] The first stage is the fault detection stage, employing a hybrid architecture of LiteInception and a Transformer encoder to perform a binary classification task of normal / abnormal data. The optimization objective is to maximize recall, strictly adhering to the aviation safety principle of erring on the side of false positives rather than false negatives. In this stage, the LiteInception module is responsible for extracting local fine features from the time-series data, while the Transformer encoder captures long-range temporal dependencies through a multi-head self-attention mechanism. The combination of these two approaches ensures both the accuracy of fault detection and a high recall rate for anomalous samples. In actual testing, the F1 score for this stage reached 81.84%, with a precision of 80.49% and a recall rate of 83.24%, differing by only 2.75 percentage points, fully meeting the requirements for safety early warning.

[0081] The second stage is the fault identification stage, which employs a pure LiteInception architecture. It performs a multi-class classification task on fault types only for samples identified as abnormal in the first stage, with the optimization goals being high classification accuracy and interpretability. This stage does not introduce a Transformer module, avoiding overfitting issues caused by excessive noise interactions introduced by the self-attention mechanism when there are many fault categories and limited data scale. Simultaneously, leveraging the lightweight nature of LiteInception, it ensures efficient inference for fault identification, with a CPU inference time of only 6.35ms and a peak memory usage of 89MB. In actual testing, this stage achieved an accuracy of 77.00% and a MacroF1 score of 75.41%.

[0082] Logically, the second stage of inference is triggered only when the first stage determines an input sample to be abnormal. Normal samples are screened out in the first stage, eliminating the need for multi-class classification calculations in the second stage. In actual general aviation aircraft operations, normal samples account for over 95%. This cascaded architecture significantly reduces the overall computational overhead of inference while allowing for independent customization of the optimization objectives for both stages. Furthermore, the models in both stages can be updated independently. When a new fault type is added, only the multi-class classification model in the second stage needs to be retrained, without adjusting the fault detection model in the first stage, greatly reducing the maintenance and iteration costs of the model.

[0083] In this implementation, the precision-recall adjustment mechanism works in deep collaboration with the LiteInception lightweight architecture to achieve flexible adjustment of model performance.

[0084] In this implementation, based on the precision-recall adjustment mechanism of knowledge distillation, a teacher-student dual-model architecture is adopted. The overall process is divided into two completely decoupled stages: the training stage and the inference deployment stage. The core execution logic is as follows: The core architecture consists of a pre-trained, fully-frozen multi-branch InceptionTime model as the teacher model and the LiteInception lightweight network proposed in this invention as the trainable student model. The teacher model only participates in the loss calculation during the training phase, providing a fine-grained feature learning benchmark for the student model, and does not participate in the final inference deployment.

[0085] The core logic of the training phase is as follows: Consistent time-series data from aircraft onboard sensors are input into both the teacher and student models, and dual-path forward propagation is performed synchronously. The teacher model outputs a classification probability distribution (soft label) softened by the temperature coefficient, while the student model simultaneously outputs its own classification probability distribution. Based on the two outputs, two independent loss components are calculated: one is the cross-entropy loss between the student model output and the true hard label of the sample; the other is the cross-entropy loss between the softened probability distributions of the teacher and student models. KLDivergence loss: The two loss components are weighted and fused according to preset distillation weights to obtain the total loss function. Backpropagation is performed based on the total loss, updating only the network weights of the student model, while the parameters of the teacher model remain frozen throughout.

[0086] The core logic of the inference deployment phase is as follows: After training is completed, only the LiteInception lightweight student model with updated weights is deployed, without the need to deploy the teacher model. The deployed model can directly perform end-to-end fault diagnosis inference on the input airborne sensor time series data. It supports flexible switching between the weight files corresponding to the direct training mode and the distillation training mode, respectively adapting to auxiliary diagnosis scenarios and safety-critical early warning scenarios. Finally, it outputs a complete diagnostic conclusion including fault type, key fault sensors, and abnormal time window.

[0087] In this implementation, the model training steps based on knowledge distillation are as follows: a pre-trained multi-branch InceptionTime network is used as the teacher model, and the LiteInception lightweight feature extraction network is used as the student model. The precision and recall of the model are adjusted by switching training modes. The training modes include a direct training mode that uses only hard-label cross-entropy loss, and a training mode that uses both hard-label cross-entropy loss and soft-label cross-entropy loss. KL Distillation training mode with divergence-integrated mixed loss. The total loss function is:

[0088] in, Output the standard cross-entropy loss (hard label loss) between the student model and the true label. The output probability distributions of the teacher and student models after softening at temperature T are compared. KL Divergence (soft label loss) The distillation weight is used to adjust the proportion of the two loss components in the total loss.

[0089] The specific calculation method for cross-entropy loss is as follows:

[0090] In the formula, Σ represents the summation calculation. C This represents the total number of fault categories. i For fault category number, One-hot encoding of the true fault category of the sample. This is the classification probability of the corresponding fault category output by the student model.

[0091] KL The specific calculation method for divergence loss is as follows:

[0092] In the formula, For temperature coefficient, The classification logic value output by the teacher model. The classification logic value output by the student model. The Softmax activation function is used. Between two probability distributions KL Divergence calculation.

[0093] This invention provides two switchable training modes, allowing the same network architecture to be adapted to two completely different application scenarios: Direct training mode: Training is performed using only hard-labeled cross-entropy loss, i.e. This makes the model tend to learn steep decision boundaries, resulting in high precision and low recall. It is suitable for auxiliary diagnostic scenarios with manual review and can effectively reduce unnecessary re-examination workload. Distillation training mode: It uses a hybrid loss combining hard label loss and soft label loss for training. Through the smooth probability distribution output by the teacher model, the student model obtains a smoother decision boundary, resulting in high recall. It is suitable for safety-critical early warning scenarios where the cost of false negatives is much greater than that of false positives, and can effectively avoid the safety risks caused by missed fault detection.

[0094] In this embodiment, the teacher model uses a pre-trained multi-branch InceptionTime model with a total of 2159K parameters. All network parameters are frozen after pre-training, and the weights are not updated throughout the training process. The student model uses the LiteInception lightweight network proposed in this invention with a total of 639K parameters, and the weights can be updated throughout the training process.

[0095] In this embodiment, the optimal distillation hyperparameter configuration, verified by grid search experiments, is adopted: distillation weight. = 0.7, temperature coefficient = 4.0. In this configuration, 70% of the total loss function comes from hard-label cross-entropy loss, and 30% comes from soft-label cross-entropy loss. KL Divergence loss can effectively improve the model's fault recall rate while ensuring the model's basic classification accuracy.

[0096] Furthermore, to verify the robustness of the parameters, this embodiment adjusts the distillation weight. With temperature coefficient A 4×4 grid search experiment was conducted, in which... The value can be {0.3, 0.5, 0.7, 0.9}. The values ​​of are {2, 4, 6, 8}. The F1 values ​​of the model in the 16 experiments range from 73.75% to 76.37%, with a maximum fluctuation of only 2.62 percentage points, which proves that the total loss function of the present invention has stable performance over a wide range of parameters.

[0097] In this embodiment, the complete training execution process is divided into two core stages: the training phase and the inference deployment phase. The complete process is as follows: Step 1: Training Phase The core objective of this stage is to train the LiteInception student model through the calculation of the total loss function and backpropagation. The specific steps are as follows: (1) Dataset preprocessing. The NGAFID aviation maintenance dataset is processed using the same preprocessing steps as the fault diagnosis process of this invention. After processing, it is divided into training set and test set in a ratio of 80% / 20%. The training set is used for model weight update and the test set is used to verify model performance.

[0098] (2) Dual-path synchronous forward propagation. The teacher model and student model are fed identical time-series data from general aviation airborne multi-channel sensors, and forward propagation calculations are performed separately for each. The teacher model outputs data after temperature coefficient calculation. = 4.0 Softened classification probability distribution, student model synchronously outputs its own classification probability distribution.

[0099] (3) Calculation of two-branch loss components. Based on the output of the forward propagation, calculate the cross-entropy loss separately. and KL Divergence loss The cross-entropy loss is calculated based on the student model output and the actual fault labels of the samples. KL The divergence loss is calculated based on the softening probability distributions of the teacher and student models.

[0100] (4) Total loss function fusion calculation. Based on the optimally configured distillation weights... = 0.7, weighted and fused the two loss components to obtain the final total loss used for model iteration.

[0101] (5) Backpropagation and weight update. Based on the total loss calculation, the backpropagation algorithm is executed to update only the network weights of the LiteInception student model. The parameters of the pre-trained InceptionTime teacher model are frozen throughout the process and do not participate in the weight update. The iterative process from step 2 to step 5 is repeated until the performance of the student model on the test set converges to a stable range, and the training is completed.

[0102] Step 2 Reasoning and Deployment Phase After training, only the trained LiteInception lightweight student model is deployed; the InceptionTime teacher model is not required, making it adaptable to the resource constraints of general aviation airborne embedded devices and ground edge terminals. Furthermore, different application scenarios can be adapted by switching training modes: the model obtained using the direct training mode has a steep decision boundary, with a precision of 77.92% and a recall of 73.46%, a difference of 4.46 percentage points between precision and recall, suitable for assisted diagnostic scenarios requiring manual review, effectively reducing unnecessary re-examination workload; the model obtained using the distillation training mode has a smooth decision boundary, with a precision of 76.40% and a recall of 75.85%, a difference of only 0.55 percentage points between precision and recall, and an improvement in recall of approximately 2.4 percentage points compared to the direct training model, suitable for safety-critical early warning scenarios where the cost of false negatives far outweighs the cost of false positives, effectively avoiding missed fault detections.

[0103] The model performance was verified by 10 repeated experiments. The best F1 value of the student model after distillation training reached 75.67%, and the standard deviation of the F1 value in 10 experiments was only 0.42%, which shows extremely high stability and reproducibility.

[0104] In this embodiment, the two-layer interpretability analysis step provides a compliant and traceable chain of evidence for the diagnostic results, while forming cross-validation with the channel selection step.

[0105] This step integrates four attribution methods based on different principles: input gradient method, occlusion sensitivity analysis, gradient-weighted class activation mapping (Grad-CAM), and integral gradient method. It performs a two-layer attribution analysis of the diagnostic results in both the sensor channel and time dimensions, and enhances the credibility of the interpretation results through multi-method cross-validation. The specific configurations of the four methods are as follows: Input gradient method: Performs one forward propagation and one backward propagation; Occlusion sensitivity analysis: 128 steps in the time dimension and 64 steps in the step size; Grad-CAM: Performs class activation mapping on the feature map output by the last LiteInception module; Integral gradient method: approximate path integration using a zero baseline and 20 interpolation points.

[0106] The specific implementation method of two-level attribution analysis is as follows: Sensor-level attribution: The two-dimensional attribution matrices calculated by each attribution method are summed along the time dimension to obtain the importance vector of each sensor channel. The intersection of the top-ranked channel sets of the four methods is taken as the key fault sensor with high confidence, and the core airborne equipment associated with the root cause of the fault is identified. Time-based attribution: The attribution matrix is ​​summed along the sensor channel dimension to obtain the time importance curve. The intervals exceeding the 95th percentile threshold are used to determine the critical time periods of anomalies and clarify the specific time range of the fault occurrence.

[0107] Simultaneously, this step sets up a noise perturbation robustness verification. Multiple levels of Gaussian noise are applied to the correctly classified samples with high confidence, and the trend of attribution distribution changes as the model confidence decreases is observed. This verifies the causal relationship between the attribution results and the model decision-making mechanism, and avoids false attributions.

[0108] Taking the experimental results of the NGAFID dataset as an example, the mean of 30 correctly classified test samples was randomly selected for each fault category. Key findings include: main current (Amp1) is a core indicator in all 19 fault categories; cylinder temperature exhibits differentiated responses, which can be used to distinguish different types of engine faults; the critical time periods for most faults are concentrated in the 30% to 90% range of flight records, and some mechanical structural faults can be detected in the early stages of flight. In noise perturbation verification, the confidence level of the original sample was 0.854. After applying 0.1 times the standard deviation of noise, the confidence level decreased to 0.626, and after applying 0.2 times the standard deviation of noise, it decreased to 0.339. The attribution distribution of all four methods degenerated from a structured pattern to a diffuse distribution, verifying the causal relationship between the attribution results and the model's decision-making mechanism. Finally, this step can generate a diagnostic report with an evidence chain traceable to "fault type—critical sensor—abnormal time period," providing clear fault location basis for aviation maintenance personnel while meeting the compliance requirements of airworthiness certification.

[0109] The lightweight and interpretable aerospace equipment abnormal state perception system provided by the embodiments of the present invention is used to perform the above-mentioned fault diagnosis method. The system includes a sensor channel selection module, a LiteInception feature extraction module, a two-stage cascaded diagnostic engine, a knowledge distillation training module, and an interpretability analysis module. The modules work together to complete the fault diagnosis of time series data of airborne sensors of general aviation aircraft.

[0110] Specifically, the output of the sensor channel selection module is connected to the input of the LiteInception feature extraction module, providing high-value sensor data input for feature extraction; the LiteInception feature extraction module provides a unified, lightweight feature extraction capability for the two-stage cascaded diagnostic engine; the first-stage output of the two-stage cascaded diagnostic engine is connected to the input of the second stage, with only abnormal samples triggering the second-stage inference; the knowledge distillation training module provides the LiteInception feature extraction module with model training and precision / recall adjustment capabilities; and the input of the interpretability analysis module is connected to the output of the two-stage cascaded diagnostic engine, performing interpretability analysis on the diagnostic results and finally outputting a diagnostic report to the human-computer interaction interface.

[0111] In actual deployment, the system's LiteInception model has a CPU single-sample inference time of only 6.35ms and a peak memory usage of 89MB. It can be deployed on mainstream embedded computing platforms or ground edge terminals, supporting flexible switching between two modes: safety-critical early warning (distillation-trained version) and assisted diagnosis (direct-trained version). Both model versions share the same network architecture, differing only in their weight files. On the NGAFID real-world aviation maintenance dataset, the system achieved an accuracy of 81.92% and a recall rate of 83.24% in the fault detection phase, and an accuracy of 77.00% in the fault identification phase, achieving an excellent balance between efficiency, accuracy, and interpretability.

[0112] The embodiments described above do not constitute a limitation on the scope of protection of this technical solution. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the above embodiments should be included within the scope of protection of this technical solution.

Claims

1. A lightweight and interpretable method for detecting abnormal states in aviation equipment, used for fault diagnosis of time-series data collected by multi-channel sensors on general aviation aircraft, characterized in that, include: A lightweight feature extraction network, LiteInception, is constructed. The basic module of the LiteInception lightweight feature extraction network is a 1+1 parallel structure containing one standard one-dimensional convolutional branch and one max pooling branch. The standard one-dimensional convolutional branch first compresses the feature channel dimension through a 1×1 bottleneck convolution, and then extracts temporal trend features through one-dimensional convolution. The max pooling branch first performs max pooling with a kernel size of 3 and a stride of 1, and then adjusts the number of feature channels through a 1×1 convolution to extract temporal extreme value features. The outputs of the two branches are concatenated and fused along the feature channel dimension. The complete network is composed of multiple stacked basic modules, and residual connections are set between the basic modules. A two-stage cascaded diagnostic process is constructed. The first stage uses a hybrid architecture of LiteInception and Transformer encoder to perform fault detection binary classification, with the optimization goal of maximizing recall. The second stage uses a pure LiteInception architecture to perform fault identification multi-class classification, with the optimization goal of high classification accuracy and interpretability. The second stage inference is triggered only when the first stage determines that the input sample is abnormal.

2. The method according to claim 1, characterized in that, It also includes a multi-method fusion sensor channel selection step: using three methods, namely mutual information evaluation, gradient analysis evaluation, and compressed excitation attention weight evaluation, to independently score the fault diagnosis contribution of each channel in the original multi-channel sensor data; and using cross-validation of multi-method consistency criteria, physical causality criteria, and information redundancy criteria, a subset of high-value sensor channels is selected.

3. The method according to claim 1, characterized in that, It also includes a model training step based on knowledge distillation: using a pre-trained multi-branch InceptionTime as the teacher model and the LiteInception lightweight feature extraction network as the student model, the precision and recall of the model are adjusted by switching training modes; the training modes include a direct training mode that uses only hard-label cross-entropy loss, and a training mode that uses both hard-label cross-entropy loss and soft-label cross-entropy loss. KL Distillation training mode for mixed loss of divergence combination.

4. The method according to claim 1, characterized in that, It also includes a two-layer interpretability analysis step: integrating four attribution methods, namely input gradient method, occlusion sensitivity analysis, gradient weighted class activation mapping, and integral gradient method, to perform two-layer attribution analysis on the diagnostic results; summing along the time dimension to obtain the importance vector of each sensor channel, and taking the intersection of the top-ranked channels of multiple attribution methods to determine the key fault sensors; summing along the sensor channel dimension to obtain the time importance curve to determine the critical time period of the anomaly.

5. The method according to claim 1, characterized in that, The complete network consists of six basic modules stacked together, with residual connections set between every three basic modules.

6. The method according to claim 4, characterized in that, The two-layer interpretability analysis step also includes noise perturbation robustness verification: applying Gaussian noise of multiple intensity levels to the correctly classified samples with high confidence, observing the changing trend of the attribution distribution as the model confidence decreases, and verifying the causal relationship between the attribution results and the model decision mechanism.

7. A lightweight and interpretable anomalous state perception system for aviation equipment, used for fault diagnosis of time-series data collected by multi-channel sensors on general aviation aircraft, characterized in that, include: The LiteInception feature extraction module is based on a 1+1 parallel structure consisting of a standard one-dimensional convolutional branch and a max-pooling branch. The standard one-dimensional convolutional branch first compresses the feature channel dimension through a 1×1 bottleneck convolution, and then extracts temporal trend features through a one-dimensional convolution. The max-pooling branch first performs max pooling with a kernel size of 3 and a stride of 1, and then adjusts the number of feature channels through a 1×1 convolution to extract temporal extreme value features. The outputs of the two branches are concatenated and fused along the feature channel dimension. The complete network is composed of multiple stacked basic units, with residual connections between the basic units. The two-stage cascaded diagnostic engine includes a first-stage fault detection module and a second-stage fault identification module. The first-stage fault detection module adopts a hybrid architecture of LiteInception and Transformer encoder to perform binary classification of fault detection, with the optimization goal of maximizing recall. The second-stage fault identification module adopts a pure LiteInception architecture to perform multi-class classification of fault identification, with the optimization goal of high classification accuracy and interpretability. The input of the second-stage fault identification module is only connected to the abnormal output of the first-stage fault detection module.

8. The system according to claim 7, characterized in that, It also includes a sensor channel selection module, which uses three methods—mutual information evaluation, gradient analysis evaluation, and compressed excitation attention weight evaluation—to independently score the fault diagnosis contribution of each sensor channel. Through cross-validation using multi-method consistency criteria, physical causality criteria, and information redundancy criteria, a subset of high-value sensor channels is selected.

9. The system according to claim 7, characterized in that, It also includes a knowledge distillation training module, which uses a pre-trained multi-branch InceptionTime as the teacher model and the network of the LiteInception feature extraction module as the student model. The module adjusts the precision and recall of the model by switching training modes. The training modes include a direct training mode using only hard-label cross-entropy loss, and a mode using both hard-label cross-entropy loss and soft-label cross-entropy loss. KL Distillation training mode for mixed loss of divergence combination.

10. The system according to claim 7, characterized in that, It also includes an interpretability analysis module, which integrates four attribution methods: input gradient method, occlusion sensitivity analysis, gradient weighted class activation mapping, and integral gradient method, to perform a two-layer attribution analysis on the diagnostic results; summing along the time dimension to obtain the importance vector of each sensor channel, and taking the intersection of the top-ranked channels of multiple attribution methods to determine the key fault sensors; summing along the sensor channel dimension to obtain the time importance curve, and determining the critical time period of the anomaly.

Images