Civil aviation airborne server abnormal state identification method and system

Abstract

The application relates to an abnormal state recognition method of a civil aviation onboard server, which comprises the following steps: collecting onboard server running data, hardware state data, onboard environment data and flight state data, and writing the flight stage, timestamp, model identification and server number; after the multi-source monitoring data is preprocessed, statistical features, frequency domain features and time sequence dependent features are extracted to generate a multi-dimensional feature vector; the multi-dimensional feature vector is input into an onboard end lightweight self-encoding model to obtain a reconstruction error, an abnormal type, a confidence and a health score; the data uploading is determined according to the reconstruction error and the confidence; the ground end aggregates the uploaded data according to the model, flight stage and hardware batch, calculates a common mode abnormality index, trains global model parameters, and then the model is updated to the onboard end.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to a method and system for identifying abnormal states of airborne servers in civil aviation. Background Technology

[0002] As the level of informatization in civil aircraft continues to increase, the functions of airborne servers in tasks such as avionics processing, cabin information services, data caching, and edge computing are constantly expanding. Their operational status directly affects the continuity of airborne applications and the stability of onboard data processing. Existing airborne server status monitoring solutions mostly adopt single-operation indicator threshold alarms, periodic inspection records, or post-event analysis on the ground. They typically focus on isolated monitoring of a few parameters such as CPU utilization, temperature, and voltage, and do not adequately consider changes during flight phases, cabin environmental disturbances, hardware batch differences, and the correlation between multiple source signals.

[0003] Traditional methods often struggle to identify progressive anomalies caused by factors such as degraded heat dissipation, prolonged resource overload, and enhanced electromagnetic interference in a timely manner. Furthermore, they lack the ability to differentiate the reliability of identification for states bordering on thresholds or exhibiting unstable behavior. In addition, existing solutions typically rely on single-machine data, lacking the capability for cross-sectional aggregation analysis of airborne servers of the same model and batch, making it difficult to detect batch anomalies with common characteristics. As ground maintenance experience, historical operational data, and newly emerging anomaly samples accumulate, existing monitoring models and judgment rules struggle to be updated and fed back to the airborne end in a timely manner, leading to false alarms, missed alarms, and insufficient adaptability. Summary of the Invention

[0004] In response, this application provides a method and system for identifying abnormal states of airborne servers in civil aviation, in order to at least partially solve the above-mentioned technical problems.

[0005] This application provides a method for identifying abnormal states of airborne servers in civil aviation, including the following steps: Collect airborne server operation data, hardware status data, airborne environment data, and flight status data, and write the flight phase, timestamp, aircraft type identifier, and server number. Outlier removal, missing value repair, noise reduction, and normalization are performed on multi-source monitoring data to obtain effective monitoring data. Statistical features, frequency domain features, and time-series dependency features are extracted from the effective monitoring data and concatenated to generate a multi-dimensional feature vector. The multidimensional feature vector is input into the lightweight autoencoder model on the airborne end, and multiple random inactivation inferences are performed to obtain the reconstruction error, anomaly type and confidence level. A health score is generated based on the reconstruction error, anomaly type and historical trend. Based on the reconstruction error and confidence level, determine whether to upload the multidimensional feature vector, health score, anomaly type and corresponding information; The ground end aggregates the uploaded data by aircraft type, flight stage and hardware batch, calculates the common mode anomaly index, and trains global model parameters based on the uploaded sample-level multidimensional feature vector data and the aggregation results or the model parameters uploaded by each airborne end. The global model parameters or parameter differences are sent to the airborne terminal, which then updates the lightweight autoencoder model based on local data and performs the next round of abnormal state identification.

[0006] In one possible embodiment, the lightweight autoencoder model includes an input layer, an encoder, a latent representation layer, and a decoder connected in sequence, as well as an anomaly classification and discrimination branch connected to the latent representation layer; the input layer receives the multidimensional feature vector, the decoder outputs the corresponding reconstruction result, and the anomaly classification and discrimination branch outputs the anomaly type probabilities of four known anomaly categories and unknown anomalies, and determines the anomaly type based on the maximum probability; wherein, the unknown anomaly is an anomaly category that has been manually verified to exist but does not belong to the known anomaly categories, and the known anomaly categories include hardware performance degradation, resource overload, environmental disturbance, or electromagnetic interference.

[0007] In one possible embodiment, performing multiple random inactivation inferences includes: A random deactivation layer is set in the hidden layer of the encoder and the anomaly classification branch, and the random deactivation layer is kept active during the inference phase to repeatedly perform forward computation; the reconstruction error is determined based on the difference between the mean of the reconstruction results obtained from multiple forward computations and the multidimensional feature vector; the confidence level is determined based on the degree of dispersion of the reconstruction results obtained from multiple forward computations.

[0008] In one possible embodiment, generating a health score based on reconstruction error, anomaly type, and historical trend includes: The severity coefficient corresponding to the anomaly type is invoked, the degradation rate index is determined based on the reconstruction error growth rate in the continuous time window or the decline rate of the health score generated in the preceding time window, and the health score is generated based on the weighted result of the reconstruction error, the severity coefficient and the degradation rate index; wherein, the severity coefficient is preset by the anomaly type or corrected based on historical maintenance statistics.

[0009] In one possible embodiment, the step of determining whether to upload the multidimensional feature vector, health score, anomaly type, and corresponding information based on reconstruction error and confidence level includes: When the reconstruction error is not lower than the high error threshold, upload the multidimensional feature vector, health score, anomaly type and corresponding information. When the reconstruction error is between the low error threshold and the high error threshold and the confidence level is not higher than the confidence threshold, the multidimensional feature vector, health score, anomaly type and corresponding information are uploaded; When the reconstruction error is lower than the low error threshold, or when the reconstruction error is between the low error threshold and the high error threshold and the confidence level is higher than the confidence level threshold, the health score, timestamp, and anomaly type are saved as a local summary on the airborne end.

[0010] In one possible embodiment, the ground terminal aggregates the uploaded data by aircraft type, flight phase, and hardware batch, and calculates a common-mode anomaly index, including: Uploaded data is written into a data cube according to aircraft type, flight stage, and hardware batch; within the same time window, the occurrence ratio of the same anomaly type in airborne servers of the same aircraft type, flight stage, and hardware batch is counted, and the occurrence ratio is determined as the common mode anomaly index; when the common mode anomaly index reaches a preset threshold, a batch anomaly record associated with the aircraft type, flight stage, and hardware batch is generated.

[0011] In one possible embodiment, training global model parameters based on aggregated data includes: When the original data is available for aggregation, the ground end uses the uploaded sample-level multidimensional feature vector data and combines it with the aggregated data to perform sample screening or loss-weighted training of the lightweight autoencoder model, so as to reconstruct the weighted loss and classification loss and update the model parameters. The unknown anomaly labels in the classification loss are provided by unclassified anomaly samples that have been manually verified. The updated model parameters are verified according to the preset training rounds, and the verified model parameters are determined as the global model parameters.

[0012] In one possible embodiment, training global model parameters based on model parameters uploaded from each airborne terminal includes: When the original data is not aggregated, each airborne terminal trains local model parameters based on local data and uploads them to the ground terminal; the ground terminal performs weighted aggregation of the local model parameters according to the local sample size corresponding to each airborne terminal, and determines the aggregation result as the global model parameters.

[0013] In one possible embodiment, the airborne terminal updates the lightweight autoencoder model by combining local data, including: After receiving the global model parameters or parameter differences, the airborne terminal updates the lightweight autoencoder model with a weighted sum of the training loss on local data and the old model parameter constraints. The old model parameter constraints limit the difference between the updated parameters and the old model parameters based on parameter importance coefficients, which are determined according to the diagonal elements of the Fisher information matrix on the local normal sample set or validation sample set of the airborne terminal.

[0014] On the other hand, the application also provides an abnormal status identification system for civil aviation airborne servers, including: The system includes a data acquisition module, a preprocessing module, a feature construction module, an airborne recognition module, an upload and judgment module, a ground aggregation and training module, and a parameter distribution module. The data acquisition module is used to collect airborne server operation data, hardware status data, airborne environment data and flight status data, and write flight phase, timestamp, aircraft type identifier and server number. The preprocessing module is used to perform outlier removal, missing value repair, noise reduction, and normalization on multi-source monitoring data. The feature construction module is used to extract statistical features, frequency domain features, and time-dependent features, and generate multi-dimensional feature vectors; The airborne identification module is used to output reconstruction error, anomaly type, confidence level, and health score based on a lightweight autoencoder model; The upload determination module is used to determine whether to upload data based on reconstruction error and confidence level. The ground aggregation training module is used to aggregate uploaded data by aircraft type, flight phase and hardware batch, calculate the common mode anomaly index and train global model parameters. The parameter distribution module is used to distribute global model parameters or parameter differences to the airborne terminal to update the lightweight autoencoder model.

[0015] In another aspect, this application also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the abnormal state identification method for a civil aviation airborne server as described above.

[0016] In another aspect, this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the abnormal state identification method for a civil aviation airborne server as described above.

[0017] Another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the abnormal state identification method for a civil aviation airborne server as described above.

[0018] Compared to existing methods based on single thresholds or post-event analysis of individual machines, this invention jointly processes operational data, hardware status data, airborne environmental data, and flight status data on the airborne end. It constructs a multi-dimensional state representation through statistical features, frequency domain features, and time-series dependent features, providing a more complete reflection of the airborne server's operational status. Simultaneously, by combining a lightweight autoencoder model, multiple random inactivation inferences, and a health score generation mechanism, it differentiates and processes significant anomalies, boundary anomalies, and low-confidence states, facilitating real-time identification on the airborne end under limited computing power. Furthermore, the ground end performs aggregated analysis by aircraft type, flight stage, and hardware batch, and updates model parameters through centralized or federated training before transmitting the data back to the airborne end for continuous model correction. Therefore, it effectively addresses the needs of online identification, air-to-ground bandwidth utilization, and cross-fleet anomaly analysis. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of an abnormal state identification method for a civil aviation airborne server provided in an embodiment of this application.

[0021] Figure 2 This is a schematic diagram of the structure of an abnormal state identification system for a civil aviation airborne server provided in an embodiment of this application.

[0022] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] This embodiment applies to scenarios where airborne servers are continuously monitored during flight of civil aircraft. Airborne servers typically handle avionics processing, cabin information services, data caching, or edge computing tasks simultaneously. Their operational status is affected by computational load, heat dissipation, power supply, cabin environment, and flight phase transitions. Considering the constraints on airborne computing power and air-to-ground link bandwidth, this embodiment performs real-time identification on the airborne end, while cross-aircraft aggregation analysis and model retraining are performed on the ground end. This allows online identification, filtering and uploading, and subsequent model updates to be performed continuously within the same technical chain.

[0025] The following detailed description, in conjunction with specific embodiments, illustrates the implementation process of the abnormal state identification method for civil aviation airborne servers described in this application. It should be noted that this embodiment is merely for explaining this application and not for limiting the scope of protection of this application. Conventional adjustments or substitutions made by those skilled in the art to each step without departing from the concept of this application should be included within the scope of protection of this application.

[0026] like Figure 1 As shown in the diagram, this application discloses a method for identifying abnormal states of a civil aviation airborne server, including the following steps: S1 collects airborne server operation data, hardware status data, airborne environment data, and flight status data, and writes the flight phase, timestamp, aircraft type identifier, and server number. S2, perform outlier removal, missing value repair, noise reduction and normalization on the multi-source monitoring data to obtain effective monitoring data, extract statistical features, frequency domain features and time-series dependency features from the effective monitoring data, and concatenate them to generate a multi-dimensional feature vector; S3, input the multidimensional feature vector into the airborne lightweight autoencoder model, perform multiple random inactivation inferences to obtain reconstruction error, anomaly type and confidence level, and generate a health score based on the reconstruction error, anomaly type and historical trend; S4. Determine whether to upload the multidimensional feature vector, health score, anomaly type and corresponding information based on the reconstruction error and confidence level. S5, the ground end aggregates the uploaded data by aircraft type, flight stage and hardware batch, calculates the common mode anomaly index, and trains global model parameters based on the uploaded sample-level multidimensional feature vector data and the aggregation results or the model parameters uploaded by each airborne end. S6, send the global model parameters or parameter differences to the airborne terminal, and the airborne terminal updates the lightweight autoencoder model in combination with local data and performs the next round of abnormal state identification.

[0027] In some embodiments, for step S1, the airborne end pre-deploys a data acquisition module, a preprocessing module, a feature construction module, a lightweight identification module, and a communication module, while the ground end deploys a communication gateway, a fleet health database, training nodes, and a parameter distribution module.

[0028] The data acquisition module connects to the airborne server monitoring interface, surrounding environment sensors, and flight status interface. The ground terminal maintains a mapping table between aircraft type and hardware batch, used for subsequent classification of uploaded data by aircraft type, flight stage, and hardware batch. The hardware batch is mapped from the server number on the ground terminal; therefore, the airborne terminal only needs to write the server number into the original record, and the ground terminal can then complete the corresponding batch information through the asset mapping table.

[0029] After the airborne terminal begins operation, the data acquisition module acquires four types of monitoring data according to a unified sampling rhythm: airborne server operating data, hardware status data, airborne environmental data, and flight status data. Airborne server operating data may include CPU utilization, memory usage, task queue length, task response latency, and process switching frequency; hardware status data may include motherboard temperature, CPU temperature, fan speed, and power supply output voltage; airborne environmental data may include cabin temperature, relative humidity, and electromagnetic interference intensity; and flight status data may include flight stage, altitude, and speed.

[0030] Each sampling record is written with the flight phase, timestamp, aircraft type identifier, and server number, forming the original monitoring record for subsequent processing. For example, the sampling frequency can be configured to 10 times per second, which is sufficient to cover changes in heat dissipation, load fluctuations, and response latency without significantly increasing the processing overhead on the airborne end.

[0031] In some embodiments, for step S2, to ensure that subsequent identification results are based on available data, the preprocessing module first performs validity processing on the original monitoring records. Validity processing includes four actions: outlier removal, missing value repair, noise reduction, and normalization. The output of the previous action is used as the input for the next action, and these actions are performed sequentially.

[0032] First, a sliding time window is maintained for each type of monitored variable. The sliding time window stores the sampled values ​​of that variable over a recent period, and the window length can be configured based on the sensor update rate and the flight stability period; for example, it can be set to 1 minute in one example. Let the sampled value at the current moment be... The window mean is The standard deviation of the window is Then the statistical deviation can be expressed as:

[0033] in, This is a numerical stability term used to avoid the denominator being zero when the standard deviation is extremely small. The preprocessing module first... The system determines statistical anomalies by checking if a value exceeds a preset threshold, and then performs a secondary check based on physical constraint rules. Physical constraint rules refer to constraints consistent with the equipment structure and operational boundaries, such as ensuring the CPU temperature is not lower than the ambient temperature of the chassis, the power supply output voltage does not exceed the rated upper limit of the interface, and the fan speed does not exhibit sudden changes exceeding mechanical inertia within adjacent short-term windows. Only when a sampled value triggers a statistical anomaly and is inconsistent with physical constraint rules is the system marked as invalid data. This reduces accidental deletions caused by transient changes in operating conditions.

[0034] For identified invalid data, the preprocessing module does not directly delete the corresponding time position from the sequence, because the feature construction stage still needs to maintain temporal continuity. The onboard unit searches for the nearest valid sample value before and after the invalid sampling point, performs linear interpolation on the short-term missing data, and records the interpolation result as follows: Its expression is:

[0035] in, The previous valid sampling time, For the next valid sampling time, and If the duration of consecutive missing data exceeds the missing duration threshold, or if the missing rate in the current sampling period exceeds the missing rate threshold, the preprocessing module will mark this period as a low-confidence data segment. Data from this low-confidence segment will still be retained in the log for ground-based traceability, but will not participate in the current round of feature extraction and airborne online identification. Long periods of missing data often correspond to sensor malfunctions, interface interruptions, or link anomalies. If interpolated results are still used instead of the original data for identification, it is easy to misinterpret acquisition failures as normal equipment status.

[0036] After outlier handling and missing value repair, the preprocessing module performs denoising and normalization on the remaining data. Denoising can be achieved using wavelet thresholding to suppress high-frequency noise without altering the low-frequency trend; subsequently, each variable is mapped to a uniform numerical range. Let the upper and lower bounds of a certain monitoring variable in the calibration set be... and Normalization results It can be represented as:

[0037] in, Similarly, for stable terms, the upper and lower bounds here can be obtained from the ground end based on historical normal data statistics, or by rolling calibration based on the most recent operating period; if the ground end updates the normalized boundary, the parameter distribution module will distribute it together with the subsequent model version update, so that the airborne end and the ground end always use the same set of calibration ranges.

[0038] The preset threshold, missing duration threshold, and missing rate threshold are calibrated by the ground terminal based on the statistical distribution of historical normal flight data and the sensor update cycle, and are distributed in a randomized configuration file.

[0039] The feature construction module extracts statistical features, frequency domain features, and time-dependent features from effective monitoring data. Statistical features reflect the basic shape of the data distribution within the current window and are typically composed of quantities such as mean, variance, maximum, minimum, and peak-to-peak values. For variables such as CPU utilization, task response latency, and motherboard temperature, statistical features can describe the average load level and fluctuation amplitude of the current window, but they cannot fully represent periodic oscillations or slow evolution trends. Therefore, frequency domain features and time-dependent features need to be further calculated.

[0040] Frequency domain characteristics are obtained by performing a frequency domain transformation on the effective monitoring data. Let the length be... The time series is corresponding spectral components It can be calculated using the Discrete Fourier Transform:

[0041] In obtaining Subsequently, the feature construction module extracts features such as the dominant frequency position, frequency band energy ratio, and spectral entropy based on the power spectrum distribution. For signals such as fan speed, power supply fluctuations, and electromagnetic interference intensity, changes in spectral components often precede obvious anomalies in the amplitude mean. Therefore, by incorporating frequency domain features into the recognition input, changes in state such as increased mechanical friction, enhanced power supply fluctuations, or aggravated electromagnetic disturbances can be reflected earlier.

[0042] The temporal dependence features originate from the Long Short-Term Memory (LSTM) network. In this embodiment, the normalized sequences of key monitoring indicators are input into the LSTM in chronological order. The LSTM processes information from previous and subsequent times through a forget gate, an input gate, and an output gate, allowing short-term fluctuations and long-term trends to be preserved simultaneously. Let the... The input vector at each time step is The previous hidden state was The state of the memory unit at the previous moment was Then its update process can be written as:

[0043] in, , , , Both the bias term and the corresponding bias term are trainable parameters. This represents the Sigmoid activation function. This represents element-wise product. The feature construction module retrieves the final hidden state. Alternatively, the pooled result of the hidden states at several time points can be used as a temporal dependency feature. Since this feature carries the sequential changes within the window, it can represent state evolutions that are not obvious when looking at a single point, such as slow heating, delayed accumulation, and fluctuating diffusion.

[0044] After obtaining the three types of features, the feature construction module performs concatenation in a preset order to form a fixed-dimensional multi-dimensional feature vector. The fixed dimension facilitates the use of a unified input structure by the airborne recognition module; for example, the multi-dimensional feature vector can be configured to have 64 dimensions, where the dimensions occupied by statistical features, frequency domain features, and temporal dependent features are determined by offline calibration at the ground end. If the number of components of a certain type of feature differs across aircraft models, the airborne end will first perform dimension alignment through a fully connected mapping layer before concatenating it with the remaining features. After concatenation, the multi-dimensional feature vector, along with the flight stage, timestamp, aircraft type identifier, and server number corresponding to that window, is sent to the airborne recognition module.

[0045] In some embodiments, for step S3, the airborne recognition module employs a lightweight autoencoder model. This model includes an input layer, an encoder, a latent representation layer, a decoder, and an anomaly classification branch connected to the latent representation layer. The input layer receives the multidimensional feature vector output from the previous stage; the encoder consists of several fully connected layers, nonlinear activation layers, and random deactivation layers, and its task is to compress the high-dimensional input into a low-dimensional latent representation; the latent representation layer outputs a compact state code; the decoder unfolds layer by layer in the opposite direction to the encoder, reconstructing the reconstruction result corresponding to the input; the anomaly classification branch originates from the latent representation layer, passes through a fully connected layer and a Softmax classification layer, and outputs the probabilities of multiple anomalies.

[0046] In one embodiment, the encoder can employ a structure of 64-dimensional input, 32-dimensional hidden layer, 16-dimensional hidden layer, and 8-dimensional latent representation, while the decoder reconstructs the output by going back to 64-dimensional representations in 8-dimensional, 16-dimensional, and 32-dimensional layers. Since this structure only shares the core representation between the latent representation layer and the classification branch, the computational load is relatively low, making it suitable for continuous operation on an onboard device.

[0047] The anomaly classification branch in the model outputs the probabilities of five types of anomalies: hardware performance degradation, resource overload, environmental disturbance, electromagnetic interference, and unknown anomalies. Hardware performance degradation mainly corresponds to persistent anomalies caused by decreased heat dissipation capacity, component aging, or changes in fan mechanical condition; resource overload mainly corresponds to latency accumulation or response jitter caused by long-term high utilization of computing resources; environmental disturbance mainly corresponds to the impact of changes in cabin temperature and humidity on the operation of the airborne server. Electromagnetic interference corresponds to signal quality anomalies caused by changes in the external electromagnetic environment; unknown anomalies are reserved for samples that have been manually verified as having anomalies but do not fully match the existing four categories. After the anomaly classification and discrimination branch outputs five probabilities, the airborne identification module selects the category corresponding to the highest probability as the anomaly type for the current window. With this setting, even if both types of anomalies are accompanied by increased temperature or increased time delay, the model can still further distinguish them by combining frequency domain features and time-series dependency features.

[0048] The random deactivation layer is enabled not only during the training phase but also during the inference phase, thus resulting in multiple random deactivation inferences. For example, the number of random deactivation inferences is... It can be configured to run 50 times. Each forward computation randomly masks a portion of the hidden neurons, thus the same input will yield a set of slightly different reconstruction results. The airborne recognition module first calculates the mean of these reconstruction results. Then, the reconstruction error is calculated based on the difference between the mean reconstruction result and the original multidimensional feature vector:

[0049] in, The dimension of a multidimensional feature vector is represented by its length. Represents the first input vector of the current window One portion, This represents the mean of the component across multiple reconstructions. Since the model primarily uses normal data to build a low-dimensional representation in the initial stage, the reconstruction error is usually small when the current window state is close to the normal operating mode; however, the reconstruction error will increase significantly when the state deviates from the normal mode.

[0050] To quantify the model's certainty regarding the current judgment, the airborne recognition module also generates a confidence score based on the dispersion of multiple reconstruction results. An uncertainty metric can be constructed first. :

[0051] Then map it to confidence level:

[0052] in, Indicates the first The variance of each reconstructed component in multiple forward computations This is a stable term. The closer the current input is to the state distribution already learned by the model, the more concentrated the results of multiple forward computations will be. The smaller, The closer the input is to 1, the lower the confidence level will be. Conversely, if the current input is on the edge of a known pattern or contains unseen combined features, the greater the difference between multiple forward calculations, and the lower the confidence level will be. In this way, in addition to outputting the anomaly type, the system also provides the confidence level of the judgment.

[0053] The airborne identification module further combines historical trends to generate a health score. The historical trend can be derived from the increase in reconstruction error within a continuous time window, or the decrease in health scores generated in previous time windows. Let the normalized value of the reconstruction error be... The severity coefficient corresponding to the anomaly type is The degradation rate index corresponding to the historical trend is Then the health score It can be generated according to the following formula:

[0054] in, , and There are three weighting coefficients that satisfy... . It reflects the degree to which the current window deviates from the normal mode. It reflects the severity of the anomaly category itself. This reflects the rate of deterioration in the situation. The three weighting coefficients can be obtained through calibration on a labeled validation set by the ground station, for example, by searching using a comprehensive indicator of false alarm rate, false negative rate, and warning lead time. When the fleet's operating environment changes over a long period, the ground station can also recalibrate and reissue the coefficients. Severity coefficient. The values ​​are obtained from an anomaly category lookup table. The initial values ​​are preset by maintenance experts based on the maintenance impact level and can be adjusted later based on actual maintenance statistics; degradation rate index It is obtained by normalizing the reconstruction error growth rate of the most recent windows or the decline rate of the previously generated health score, and the degradation rate index is directly calculated from the results of the previous windows.

[0055] In some embodiments, for step S4, based on the reconstruction error and confidence level, the onboard unit also needs to determine whether the current window needs to be uploaded. This involves three types of thresholds: high error threshold, low error threshold, and confidence level threshold. The high error threshold is used to filter out samples that deviate significantly from the normal pattern, the low error threshold is used to filter out boundary samples, and the confidence level threshold is used to distinguish whether the model is stable.

[0056] For example, these three thresholds can be calibrated on the ground based on the distribution of recent operational data and manual verification results; when the statistical distribution corresponding to different aircraft models and different seasonal operating conditions changes steadily, the thresholds can also be updated along with the model version. The airborne terminal performs the upload judgment according to the following logic: if the reconstruction error is not lower than the high error threshold, the current window directly triggers the upload; if the reconstruction error is between the low error threshold and the high error threshold, and the confidence level is not higher than the confidence level threshold, the upload is also triggered. In other cases, the complete window data is not uploaded; instead, a local summary is saved on the airborne terminal. The local summary is a lightweight set of results retained on the airborne terminal when no upload is triggered, typically including health score, timestamp, and anomaly type. When an upload is triggered, the airborne terminal uploads a multi-dimensional feature vector, health score, anomaly type, and corresponding information, including flight phase, timestamp, aircraft type identifier, and server number. With this processing, the air-to-ground link primarily carries samples with high anomaly levels or insufficient model understanding; stable, normal windows only retain summaries and do not continuously consume upload bandwidth.

[0057] In some embodiments, for step S5, the ground terminal receives the data uploaded by the airborne terminal through a communication gateway and writes it into a data cube in the fleet health database. The data cube is organized by aircraft type, flight phase, and hardware batch, with timestamps, server numbers, and anomaly types serving as detailed indexes within the cube cells.

[0058] The hardware batches are not collected in real-time from the airborne terminal, but are obtained by matching server numbers from an asset table on the ground. Therefore, during aggregation, server samples belonging to the same batch can be grouped together. After data entry is completed on the ground, the proportion of airborne servers of the same anomaly type, within the same aircraft type, flight phase, and hardware batch (distinguished by server number) within the same time window is calculated as the common-mode anomaly index. Let the number of airborne servers meeting the above grouping conditions within a certain time window be... The first server number is distinguished by its server ID. The exception type of each airborne server is denoted as Then the first The common-mode anomaly index of a class of anomalies can be expressed as:

[0059] in, This is an indicator function. When When the common mode anomaly threshold is reached, the common mode anomaly threshold is determined by the ground terminal according to the statistical distribution of the common mode anomaly index during historical normal operation according to a preset quantile. The ground terminal generates a batch anomaly record associated with the aircraft type, flight phase and hardware batch.

[0060] The samples in the fleet health database come partly from multidimensional feature vectors and recognition results uploaded directly from the airborne terminal, and partly from local model parameters uploaded periodically by the airborne terminal. The ground terminal selects different training paths based on data aggregation availability. When raw data can be aggregated, centralized training is used; when raw data cannot be aggregated, federated training is used. Regardless of the chosen path, the training object remains the same lightweight autoencoder model, allowing updated parameters from the ground terminal to be directly sent back to the airborne terminal without structural transformation. The centralized training path uses uploaded sample-level multidimensional feature vectors as training input, and the aggregation results are used for common mode analysis and sample grouping, filtering, or weighting during training.

[0061] In the centralized training path, training nodes directly access uploaded sample-level multidimensional feature vector samples and combine them with aggregated results obtained by aircraft type, flight stage, and hardware batch to perform stratified sampling, sample selection, or loss weighting to update the lightweight autoencoder model. The training loss consists of reconstruction loss and classification loss.

[0062] in, This represents the reconstruction loss, reflecting the model's ability to reconstruct normal patterns; This represents the classification loss, reflecting the difference between the output of the anomaly classification branch and the known labels; This is a balancing coefficient used to adjust the relative weight of the two loss components during training. If... If the value is too small, the model will be more biased towards reconstructing normal patterns, and the distinction between abnormal categories will be weaker; if... If the value is too large, the model may sacrifice reconstruction stability and overfit the limited number of labeled samples. Therefore... It is generally obtained through validation set calibration and saved in the ground configuration along with the model version.

[0063] During training, the Adam optimizer can be used, with the learning rate and training epochs configured by the ground-side based on the sample size. After several iterations, the training nodes verify the reconstruction error distribution, classification confusion, and upload trigger hit rate on the validation set, forming new global model parameters upon confirmation. The training samples here include both manually confirmed anomalous samples and a large number of normal window samples. Therefore, the updated model retains the low-dimensional representation capability of normal states while continuing to expand the coverage of anomalous category boundaries. Among them, the training samples corresponding to unknown anomalous categories consist of samples that have been manually verified to be anomalous but cannot be classified into the four categories of hardware performance degradation, resource overload, environmental disturbance, and electromagnetic interference.

[0064] In the federated training path, each airborne terminal trains its local model parameters based on local data and uploads them to the ground terminal. The ground terminal does not receive raw monitoring data; instead, it weights and aggregates the local model parameters according to sample size to generate a new round of global model parameters. Let the... The local sample size of each airborne terminal is The model parameters obtained after local training are Then the ground end of the first The aggregation result of the wheel is:

[0065] in, This indicates the number of airborne terminals participating in this round of training. Airborne terminals with larger sample sizes have a higher proportion of their local model parameters during aggregation, thus preventing individual airborne terminals with very small sample sizes from causing excessive disturbance to the global model. Since ground-based terminals do not need to see the original monitoring data in the federated training path, it is suitable for deployment environments with strong data isolation requirements. Regardless of whether centralized or federated training is used, the training nodes output global model parameters under the same semantics, and the subsequent distribution process remains consistent.

[0066] In some embodiments, for step S6, the parameter distribution module sends the global model parameters or parameter differences to the airborne terminal. Parameter differences refer to the set of differences between the new round of global model parameters and the current model parameters on the airborne terminal; when link bandwidth is tight, sending parameter differences can reduce communication burden. After receiving the model update packet, the airborne terminal performs a constraint update based on local data. Let the local training loss be... The parameters of the old model are The updated parameters are The parameter importance coefficient is Then the local update loss at the airborne end can be expressed as:

[0067] in, These are regularization coefficients used to control the strength of constraints on old parameters. Parameter importance coefficients. The regularization coefficient is determined by the diagonal elements of the Fisher Information Matrix on the local normal or validation sample set of the airborne terminal. A larger value indicates that the parameter is more important to the existing recognition capabilities of the airborne terminal. In this way, while absorbing new knowledge from the ground, the airborne terminal retains key old parameters more effectively, preventing the model from rapidly deviating from existing normal patterns when local samples are scarce. Generally, the calibration is performed by the ground-based system before the version release; if a particular model shows significant neglect after multiple updates, it can be improved in the next model version. The value of .

[0068] After the partial update is completed, the airborne terminal replaces the old model with the new lightweight autoencoder model and performs the next round of abnormal state identification with the updated parameters. At this time, the calling relationship between the preprocessing module, feature construction module, and airborne identification module remains unchanged; only the weight parameters within the lightweight autoencoder model and the corresponding threshold, normalization boundary, and weight coefficient configurations change. In this way, the processing flow of the airborne terminal from data collection to health score generation and then to execution and upload judgment remains stable, while the ground terminal continuously corrects the model through cross-aircraft data, so that subsequent identification results gradually approach the actual operating state.

[0069] In actual deployment, a version management relationship is maintained between the airborne and ground terminals. Each round of model update packages typically includes model parameter version numbers, threshold configuration version numbers, and normalization boundary version numbers. Before installing a new version, the airborne terminal verifies the version matching relationship to avoid situations where model parameters have been updated but the thresholds are still using the old version, or where the normalization boundaries have changed but the recognition module is still calling the old configuration. If a link interruption occurs, the airborne terminal continues to use the current version to complete local recognition and local summary saving. Once the link is restored, it sends the local summary and the high-value window data that needs to be retransmitted in this round. After receiving the data, the ground terminal writes it into the data cube according to the original classification rules.

[0070] In summary, the airborne terminal first filters out invalid data within each monitoring window, then forms a stable multi-dimensional feature vector, and obtains reconstruction error, anomaly type, confidence level, and health score through multiple random inactivation inferences. Finally, it decides whether to upload the data based on a threshold. Upon receiving the uploaded data, the ground terminal first aggregates the data by aircraft type, flight stage, and hardware batch to generate a common-mode anomaly index. Then, it trains new global model parameters based on the availability of aggregated samples. Finally, it sends the global model parameters or parameter differences back to the airborne terminal to complete constraint updates. The airborne terminal's identification, ground terminal's aggregation, and model updates can be seamlessly integrated, preserving the airborne terminal's real-time judgment capability while incorporating cross-aircraft statistical information into the subsequent training process. This ensures that the airborne server's anomaly identification maintains relatively stable judgment accuracy and adaptability during continuous operation.

[0071] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of an abnormal state identification system 2 for a civil aviation airborne server provided in an embodiment of this application. Figure 2 As shown, the system includes: The system includes a data acquisition module 201, a preprocessing module 202, a feature construction module 203, an airborne recognition module 204, an upload and judgment module 205, a ground aggregation and training module 206, and a parameter distribution module 207. The data acquisition module is used to collect airborne server operation data, hardware status data, airborne environment data and flight status data, and write flight phase, timestamp, aircraft type identifier and server number. The preprocessing module is used to perform outlier removal, missing value repair, noise reduction, and normalization on multi-source monitoring data. The feature construction module is used to extract statistical features, frequency domain features, and time-dependent features, and generate multi-dimensional feature vectors; The airborne identification module is used to output reconstruction error, anomaly type, confidence level, and health score based on a lightweight autoencoder model; The upload determination module is used to determine whether to upload data based on reconstruction error and confidence level. The ground aggregation training module is used to aggregate uploaded data by aircraft type, flight phase and hardware batch, calculate the common mode anomaly index and train global model parameters. The parameter distribution module is used to distribute global model parameters or parameter differences to the airborne terminal to update the lightweight autoencoder model.

[0072] Those skilled in the art will clearly understand that the technical solutions of the embodiments of this application can be implemented by means of software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware that can independently complete or cooperate with other components to complete a specific function, wherein the hardware may be, for example, a field-programmable gate array (FPGA), an integrated circuit (IC), etc.

[0073] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0074] Based on the same inventive concept, this application also provides an electronic device, the method corresponding to which can be the method in the foregoing embodiments, and its problem-solving principle is similar to that method. For example... Figure 3 As shown, Figure 3 This is a schematic diagram of an electronic device structure provided in an embodiment of the present disclosure. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the methods and / or technical solutions of the foregoing embodiments of the present application.

[0075] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a processor, it performs the functions defined in the methods of this application.

[0076] Another embodiment of this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0077] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0078] The embodiments disclosed herein are preferred embodiments, but are not limited thereto. Those skilled in the art can readily grasp the spirit of the present invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. A method for identifying abnormal states of a civil aviation airborne server, characterized in that, include: Collect multi-source monitoring data, including airborne server operation data, hardware status data, airborne environment data, and flight status data, and write the flight phase, timestamp, aircraft type identifier, and server number. Outlier removal, missing value repair, noise reduction, and normalization are performed on multi-source monitoring data to obtain effective monitoring data. Statistical features, frequency domain features, and time-series dependency features are extracted from the effective monitoring data and concatenated to generate a multi-dimensional feature vector. The multidimensional feature vector is input into the lightweight autoencoder model on the airborne end, and multiple random inactivation inferences are performed to obtain reconstruction error, anomaly type, and confidence level. A health score is then generated based on the reconstruction error, anomaly type, and historical trend. Specifically, the severity coefficient is called according to the anomaly type, and the degradation rate index is determined based on the historical trend, i.e., the reconstruction error growth rate in a continuous time window, or based on the decline rate of the health score generated in a previous time window. The health score is then generated based on the weighted result of the reconstruction error, the severity coefficient, and the degradation rate index. The severity coefficient is preset by the anomaly type or corrected based on historical maintenance statistics. Based on the reconstruction error and confidence level, determine whether to upload the multidimensional feature vector, health score, anomaly type and corresponding information; The ground terminal aggregates uploaded data by aircraft type, flight stage, and hardware batch, calculates a common-mode anomaly index, and trains global model parameters based on uploaded sample-level multidimensional feature vector data and the aggregation results or model parameters uploaded by each airborne terminal. The calculation of the common-mode anomaly index includes: writing uploaded data into a data cube according to aircraft type, flight stage, and hardware batch; within the same time window, statistically analyzing the occurrence ratio of the same anomaly type in airborne servers of the same aircraft type, flight stage, and hardware batch, and determining the occurrence ratio as the common-mode anomaly index; when the common-mode anomaly index reaches a preset threshold, generating batch anomaly records associated with that aircraft type, flight stage, and hardware batch. The global model parameters or parameter differences are sent to the airborne terminal, which then updates the lightweight autoencoder model based on local data and performs the next round of abnormal state identification. The parameter differences refer to the set of differences between the new round of global model parameters and the current model parameters of the airborne terminal.

2. The method according to claim 1, characterized in that, The lightweight autoencoder model includes an input layer, an encoder, a latent representation layer, and a decoder connected in sequence, as well as an anomaly classification and discrimination branch connected to the latent representation layer. The input layer receives the multidimensional feature vector, the decoder outputs the corresponding reconstruction result, and the anomaly classification and discrimination branch outputs the anomaly type probabilities of four known anomaly categories and unknown anomalies, and determines the anomaly type based on the maximum probability. Among them, the unknown anomalies are anomalies that have been manually verified to exist but do not belong to the known anomaly categories. The known anomaly categories include hardware performance degradation, resource overload, environmental disturbance, or electromagnetic interference.

3. The method according to claim 2, characterized in that, The execution of multiple random inactivation inferences includes: A random deactivation layer is set in the hidden layer of the encoder and the anomaly classification branch, and the random deactivation layer is kept active during the inference phase to repeatedly perform forward computation; the reconstruction error is determined based on the difference between the mean of the reconstruction results obtained from multiple forward computations and the multidimensional feature vector; the confidence level is determined based on the degree of dispersion of the reconstruction results obtained from multiple forward computations.

4. The method according to claim 3, characterized in that, The step of determining whether to upload the multidimensional feature vector, health score, anomaly type, and corresponding information based on reconstruction error and confidence level includes: When the reconstruction error is not lower than the high error threshold, upload the multidimensional feature vector, health score, anomaly type and corresponding information. When the reconstruction error is between the low error threshold and the high error threshold and the confidence level is not higher than the confidence threshold, the multidimensional feature vector, health score, anomaly type and corresponding information are uploaded; When the reconstruction error is lower than the low error threshold, or when the reconstruction error is between the low error threshold and the high error threshold and the confidence level is higher than the confidence level threshold, the health score, timestamp, and anomaly type are saved as a local summary on the airborne end.

5. The method according to claim 4, characterized in that, The step of training global model parameters based on the aggregation results includes: When the original data is available for aggregation, the ground end uses the uploaded sample-level multidimensional feature vector data and the aggregation results to perform sample screening or loss-weighted training of the lightweight autoencoder model, so as to reconstruct the weighted loss and classification loss and update the model parameters. The unknown anomaly labels in the classification loss are provided by unclassified anomaly samples that have been manually verified. The updated model parameters are verified according to the preset training rounds, and the verified model parameters are determined as the global model parameters.

6. The method according to claim 5, characterized in that, The process of training global model parameters based on model parameters uploaded from each airborne terminal includes: When the original data is not aggregated, each airborne terminal trains local model parameters based on local data and uploads them to the ground terminal; the ground terminal performs weighted aggregation of the local model parameters according to the local sample size corresponding to each airborne terminal, and determines the aggregation result as the global model parameters.

7. The method according to claim 6, characterized in that, The airborne terminal updates the lightweight autoencoder model by combining local data, including: After receiving the global model parameters or parameter differences, the airborne terminal updates the lightweight autoencoder model with a weighted sum of the training loss on local data and the old model parameter constraints. The old model parameter constraints limit the difference between the updated parameters and the old model parameters based on parameter importance coefficients, which are determined according to the diagonal elements of the Fisher information matrix on the local normal sample set or validation sample set of the airborne terminal.

8. An abnormal status identification system for a civil aviation airborne server, characterized in that, include: The system includes a data acquisition module, a preprocessing module, a feature construction module, an airborne recognition module, an upload and judgment module, a ground aggregation and training module, and a parameter distribution module. The data acquisition module is used to collect multi-source monitoring data, which includes airborne server operation data, hardware status data, airborne environment data and flight status data, and writes flight phase, timestamp, aircraft type identifier and server number into the data. The preprocessing module is used to perform outlier removal, missing value repair, noise reduction, and normalization on multi-source monitoring data. The feature construction module is used to extract statistical features, frequency domain features, and time-dependent features, and generate multi-dimensional feature vectors; The airborne identification module is used to output reconstruction error, anomaly type, confidence level, and health score based on a lightweight autoencoder model. Specifically, it calls the corresponding severity coefficient according to the anomaly type, determines the degradation rate index based on the historical trend, i.e., the reconstruction error growth rate in a continuous time window, or based on the decline rate of the health score generated in a previous time window, and generates the health score based on the weighted result of the reconstruction error, the severity coefficient, and the degradation rate index. The severity coefficient is preset by the anomaly type or corrected based on historical maintenance statistics. The upload determination module is used to determine whether to upload data based on reconstruction error and confidence level. The ground aggregation training module is used to aggregate uploaded data by aircraft type, flight phase, and hardware batch, calculate the common-mode anomaly index, and train global model parameters. The calculation of the common-mode anomaly index includes writing the uploaded data into a data cube according to aircraft type, flight phase, and hardware batch; within the same time window, statistically analyzing the occurrence ratio of the same anomaly type in airborne servers of the same aircraft type, flight phase, and hardware batch, and determining the occurrence ratio as the common-mode anomaly index; when the common-mode anomaly index reaches a preset threshold, generating batch anomaly records associated with that aircraft type, flight phase, and hardware batch. The parameter distribution module is used to distribute global model parameters or parameter differences to the airborne terminal to update the lightweight autoencoder model; the parameter difference refers to the set of differences between the new round of global model parameters and the current model parameters on the airborne terminal.

Images