Ai-based onboard failure prediction and remote support method and system

Abstract

This invention provides an AI-based method and system for airborne fault prediction and remote support, relating to the field of artificial intelligence technology. It includes extracting envelope transition features from time-series state data of airborne equipment and performing dual-envelope confidence analysis using an airborne deep neural network to achieve transition fault prediction. When communication is limited, key data is selected based on the marginal information gain value of data segments and sent to the ground station. The ground station combines the received features and data, using physical equation deduction and deep neural network analysis to generate rapid and complete prediction results, which are then cross-validated. Finally, a support plan is generated when the verification is valid. This invention achieves efficient and accurate fault prediction and remote decision support under limited communication bandwidth.

Description

Technical Field

[0001] This invention relates to artificial intelligence technology, and more particularly to an AI-based method and system for predicting and remotely supporting airborne faults. Background Technology

[0002] In the field of aircraft health management and maintenance support, fault prediction and remote technical support for airborne equipment are crucial for ensuring flight safety and operational efficiency. Existing technologies typically rely on online or offline analysis of time-series status data collected from airborne equipment. However, these conventional practices have revealed several shortcomings in practical applications. Firstly, existing methods lack sufficient precision and foresight in predicting potential faults during dynamic flight envelope transitions. Flight envelope transitions often involve complex shifts in equipment operating states, and mechanisms based on fixed thresholds or simple anomaly detection struggle to effectively capture the early, subtle characteristics that foreshadow specific fault modes during these transitions. This leads to delayed fault prediction or misclassification, hindering proactive maintenance planning. Secondly, in typical aviation conditions where air-to-ground data transmission bandwidth is limited or communication is unstable due to error terms, existing data transmission strategies often lack intelligent filtering mechanisms. The common practice of periodic full transmission or simple sampling may transmit large amounts of low-value, highly redundant data while missing critical data segments essential for fault diagnosis. This not only increases the burden on communication links, but may also cause the ground end to be unable to make accurate judgments due to incomplete information, thus delaying the response of remote technical support. Summary of the Invention

[0003] This invention provides an AI-based method and system for airborne fault prediction and remote support, which can solve the problems in the prior art.

[0004] A first aspect of this invention provides an AI-based method for airborne fault prediction and remote support, comprising:

[0005] The system acquires time-series status data of airborne equipment. When the change in flight parameters exceeds a preset threshold, the second-order time derivative of the time-series status data within the switching time window is extracted as an envelope transition feature. The envelope transition feature is input into a deep neural network on the airborne side to output a double-envelope confidence distribution of the current and target flight envelopes. The corresponding fault feature databases are then weighted according to the proportion of the double-envelope confidence distribution to obtain the transition prediction result.

[0006] Calculate the marginal information gain value of each data segment in the time-series state data to the transition prediction result. When communication is restricted, select a preset number of data segments from high to low according to the marginal information gain value, and send the envelope transfer feature and the data segments to the ground end.

[0007] The ground end inputs the received envelope transfer features into the ground-side deep neural network to obtain a fast prediction result. At the same time, based on the received data segments, it uses the physical state equation of the airborne equipment to deduce the state evolution of the untransmitted data segments. The deduced complete time series data is input into the ground-side deep neural network to obtain a complete prediction result. When the fault type identifier of the fast prediction result is consistent with that of the complete prediction result, the data reconstruction is confirmed to be effective and a support plan is generated.

[0008] The steps for extracting the second-order time derivative of time-series state data within the switching time window as envelope transition features include:

[0009] The first-order time derivatives of the multidimensional flight parameters in the time-series state data within the switching time window are calculated to obtain the instantaneous rate of change sequence of each dimension.

[0010] The instantaneous rate of change sequence is subjected to second-order time differential operation to obtain acceleration change characteristics in each dimension, which characterize the dynamic evolution trend during the flight envelope transfer process;

[0011] A multi-dimensional coupling correlation matrix is ​​constructed, which describes the cooperative change pattern among different flight parameters by calculating the cross-correlation coefficient of the acceleration change characteristics in the time dimension;

[0012] The dominant coupling direction vector is extracted by performing eigenvalue decomposition on the coupling correlation matrix. The dominant coupling direction vector is then weighted and fused with the acceleration change features of the corresponding dimension to obtain the dimensionality-reduced envelope transfer features.

[0013] The steps of inputting the envelope transfer features into the airborne deep neural network to output the double-envelope confidence distribution of the current and target flight envelopes, and performing a weighted query on the corresponding fault feature databases according to the proportion of the double-envelope confidence distribution to obtain the transition prediction result include:

[0014] The envelope transfer features are input into the current envelope recognition branch and the target envelope recognition branch of the deep neural network on the airborne side, and the current flight envelope confidence vector and the target flight envelope confidence vector are output respectively to form a double envelope confidence distribution;

[0015] The current flight envelope confidence vector and the target flight envelope confidence vector are input into the envelope interaction layer in the deep neural network to calculate the transfer stress tensor, which characterizes the dynamic coupling strength between the two envelopes;

[0016] Based on the dual-envelope confidence distribution, a first candidate fault set and a second candidate fault set are retrieved from the first fault feature library corresponding to the current flight envelope and the second fault feature library corresponding to the target flight envelope, respectively.

[0017] A third set of candidate faults is obtained by retrieving the transfer fault feature library based on the transfer stress tensor. The transfer fault feature library stores fault modes that only occur during the envelope transfer process.

[0018] The first candidate fault set, the second candidate fault set, and the third candidate fault set are weighted and fused to obtain the transition prediction result. The weight coefficients are jointly determined by the double envelope confidence distribution and the transferred stress tensor.

[0019] The steps of calculating the marginal information gain value of each data segment in the time-series state data to the transition prediction result, selecting a preset number of data segments from high to low according to the marginal information gain value when communication is restricted, and transmitting the envelope transfer feature and the data segments to the ground end include:

[0020] The time-series state data is divided into multiple data segments in chronological order, and the statistical distribution characteristics of each data segment are extracted.

[0021] For each fault type in the transition prediction results, a conditional probability mapping relationship is constructed. The statistical distribution characteristics of each data segment are used as the observation input to the conditional probability mapping relationship. The difference in the predicted probability of the fault type when the data segment exists and when it does not exist is calculated.

[0022] The difference in predicted probabilities is used as the single information contribution of the data segment to the fault type. The marginal information gain value of the data segment is obtained by weighting and summing the single information contributions of the individual data segment to all fault types included in the transition prediction result. The weighting coefficient is determined by the initial probability of each fault type in the transition prediction result.

[0023] In a communication-restricted state, all data segments are sorted in descending order according to the marginal information gain value, and a preset number of data segments at the top of the sorting are selected to form a dataset to be transmitted; the envelope transfer feature and the dataset to be transmitted are sent to the ground end.

[0024] The steps for calculating the marginal information gain of the data segment on the transition prediction result include:

[0025] A causal dependency graph is constructed between flight parameters based on the time-series state data. The causal dependency graph identifies the causal precedence relationship between parameters through Granger causality test. If the historical value of parameter A has predictive ability for the current value of parameter B, a directed edge from A to B is established.

[0026] The causal centrality of each data segment is calculated based on the causal dependency graph, and the causal centrality measures the range of influence of the flight parameters contained in the data segment in the entire causal network.

[0027] The statistical distribution characteristics of each data segment are mapped to the causal centrality input conditional probability, and the conditional entropy difference between the predicted probability of fault type when the data segment exists and does not exist is calculated.

[0028] The conditional entropy difference and the causal centrality are nonlinearly weighted and fused to obtain the causal enhancement marginal information gain value. The nonlinear weighting is implemented through a gating mechanism, which amplifies the data segments corresponding to the causal key nodes and suppresses the data segments corresponding to the causal edge nodes.

[0029] The steps of inputting the received envelope transfer features into a ground-side deep neural network to obtain a fast prediction result, and simultaneously using the physical state equations of the airborne equipment to deduce the state evolution of the untransmitted data segments based on the received data segments, and inputting the deduced complete time-series data into the ground-side deep neural network to obtain a complete prediction result include:

[0030] The envelope transfer features are input into the pre-feature extraction layer of the deep neural network on the ground side to obtain the envelope abstract features. The envelope abstract features are directly input into the high-level decision module through a fast path to output fast prediction results. The fast path skips the temporal modeling layer.

[0031] The state values ​​and rates of change of each flight parameter are extracted from the received data segments as initial conditions and substituted into the physical state equation of the airborne equipment. The flight parameter evolution trajectory for the time interval corresponding to the untransmitted data segments is calculated. The flight parameter evolution trajectory and the received data segments are spliced ​​together in chronological order to obtain the reconstructed complete time series data.

[0032] The reconstructed complete time series data is input into the pre-feature extraction layer to obtain the initial time series features. The envelope abstract features are used by the feature modulation module to generate a spatial attention weight map to perform element-wise weighting on the initial time series features to obtain the modulated time series features. The modulated time series features are then input into the high-level decision module after the time series modeling layer extracts the time series dependencies, and the complete prediction results are output.

[0033] Both the rapid prediction result and the complete prediction result include a fault type identifier and a corresponding fault probability distribution.

[0034] The steps to confirm the effectiveness of data reconstruction and generate supporting solutions include:

[0035] Determine whether the rapid prediction result and the complete prediction result are consistent in fault type identification and fault probability distribution. If they are consistent, confirm that the data reconstruction is effective and generate a support plan.

[0036] When there is a discrepancy, the contribution of each time period in the reconstructed complete time series data to the prediction difference is calculated, and the time periods are sorted in descending order of contribution. A preset number of time periods at the top of the sorted list are identified as belonging to the time interval corresponding to the untransmitted data segment.

[0037] If a preset number of time periods at the top of the sorting list belong to the untransmitted interval, the original data segments of these time periods are requested to be retransmitted from the airborne terminal and the corresponding inferred data is replaced. If they belong to the transmitted interval, the boundary conditions or solution accuracy parameters of the physical state equation of the airborne equipment in these time periods are adjusted and the untransmitted data segments are re-inferred.

[0038] The corrected complete time series data is re-input into the ground-side deep neural network to obtain the corrected complete prediction result, and consistency is checked with the fast prediction result. The correction process is iteratively executed until consistency is achieved or the preset maximum number of iterations is reached.

[0039] A second aspect of the present invention provides an AI-based airborne fault prediction and remote support system, comprising:

[0040] The envelope transfer unit is used to acquire the time-series status data of airborne equipment. When the change in flight parameters exceeds the preset change threshold, the second time derivative of the time-series status data within the switching time window is extracted as the envelope transfer feature.

[0041] The dual-envelope unit is used to input the envelope transfer features into the airborne deep neural network and output the dual-envelope confidence distribution of the current flight envelope with the target. The corresponding fault feature databases are then weighted according to the proportion of the dual-envelope confidence distribution to obtain the transition prediction result.

[0042] Marginal information unit is used to calculate the marginal information gain value of each data segment in the time-series state data to the transition prediction result. When communication is restricted, a preset number of data segments are selected from high to low according to the marginal information gain value, and the envelope transfer feature and the data segments are sent to the ground end.

[0043] The data reconstruction unit is used to input the received envelope transfer features into the ground-side deep neural network to obtain a fast prediction result. At the same time, based on the received data segments, it uses the physical state equation of the airborne equipment to deduce the state evolution of the untransmitted data segments. The deduced complete time series data is input into the ground-side deep neural network to obtain a complete prediction result. When the fault type identifier of the fast prediction result is consistent with that of the complete prediction result, the data reconstruction is confirmed to be effective and a support plan is generated.

[0044] A third aspect of the present invention provides an electronic device, comprising:

[0045] processor;

[0046] Memory used to store processor-executable instructions;

[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0049] This invention inputs envelope transfer features into an airborne deep neural network, directly outputting the dual-envelope confidence distribution of the current and target envelopes, thus achieving a quantitative assessment of the probability of failure under complex transition states. Based on this distribution ratio, a weighted query is performed on the corresponding fault feature database, enabling the transition prediction results to incorporate typical fault knowledge under different envelopes, thereby improving the relevance and reliability of the prediction.

[0050] Under conditions of limited communication resources, this method achieves intelligent filtering of transmitted data by calculating the marginal information gain value of each data segment on the transition prediction result. The ground end employs a dual-path parallel prediction and data reconstruction verification mechanism to further ensure the robustness of remote support. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the AI-based airborne fault prediction and remote support method according to an embodiment of the present invention.

[0052] Figure 2 This is a flowchart for calculating the marginal information gain value based on the statistical characteristics of time-series data segments. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0055] Figure 1 This is a flowchart illustrating the AI-based airborne fault prediction and remote support method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] Acquire the time-series status data of airborne equipment. When the change in flight parameters exceeds the preset change threshold, extract the second time derivative of the time-series status data within the switching time window as the envelope transition feature.

[0057] The envelope transfer features are input into the airborne deep neural network, which outputs the dual-envelope confidence distribution of the current and target flight envelopes. The corresponding fault feature databases are then weighted according to the proportion of the dual-envelope confidence distribution to obtain the transition prediction results.

[0058] Calculate the marginal information gain value of each data segment in the time-series state data to the transition prediction result. When communication is restricted, select a preset number of data segments from high to low according to the marginal information gain value, and send the envelope transfer feature and the data segments to the ground end.

[0059] The ground end inputs the received envelope transfer features into the ground-side deep neural network to obtain a fast prediction result. At the same time, based on the received data segments, it uses the physical state equation of the airborne equipment to deduce the state evolution of the untransmitted data segments. The deduced complete time series data is input into the ground-side deep neural network to obtain a complete prediction result. When the fault type identifier of the fast prediction result is consistent with that of the complete prediction result, the data reconstruction is confirmed to be effective and a support plan is generated.

[0060] In one optional implementation, the step of extracting the second-order time derivative of the time-series state data within the switching time window as an envelope transition feature includes:

[0061] The first-order time derivatives of the multidimensional flight parameters in the time-series state data within the switching time window are calculated to obtain the instantaneous rate of change sequence of each dimension.

[0062] The instantaneous rate of change sequence is subjected to second-order time differential operation to obtain acceleration change characteristics in each dimension, which characterize the dynamic evolution trend during the flight envelope transfer process;

[0063] A multi-dimensional coupling correlation matrix is ​​constructed, which describes the cooperative change pattern among different flight parameters by calculating the cross-correlation coefficient of the acceleration change characteristics in the time dimension;

[0064] The dominant coupling direction vector is extracted by performing eigenvalue decomposition on the coupling correlation matrix. The dominant coupling direction vector is then weighted and fused with the acceleration change features of the corresponding dimension to obtain the dimensionality-reduced envelope transfer features.

[0065] For example, during actual flight, when the detected change in flight parameters of the airborne equipment exceeds a preset threshold, fine-grained feature extraction needs to be performed on the time-series status data. The preset threshold is set according to different flight parameter types. For example, the altitude change threshold can be set to 50 meters per second, the speed change threshold can be set to 10 meters per second, the angle of attack change threshold can be set to 3 degrees per second, and the engine speed change threshold can be set to 5% of the rated speed. Assuming that the collected time-series status data includes flight parameters in multiple dimensions such as altitude, speed, angle of attack, and engine speed, the sampling frequency is 100Hz, and the switching time window is set to 5 seconds, then each dimension forms a sequence of 500 data points.

[0066] The first-order time derivative of the flight parameter in each dimension is calculated, and the numerical differentiation is achieved using the central difference method. For the data point x of the i-th dimension flight parameter at time t... i (t), whose first derivative is calculated as Δt is the sampling time interval of 0.01 seconds. The instantaneous rate of change sequence for each dimension is calculated, reflecting the instantaneous rate of change of each flight parameter. For example, the rate of change of altitude corresponds to the rate of climb or descent, and the rate of change of velocity corresponds to acceleration or deceleration.

[0067] Continue performing second-order time differential operations on the instantaneous rate of change sequence, and also use the central difference method to calculate the acceleration change characteristics. The second derivative characterizes the evolution rate of flight parameter changes and can capture dynamic abrupt changes during envelope transition. Forward or backward differencing is used at boundary points to ensure sequence integrity.

[0068] For the extracted acceleration change features in each dimension, a coupling correlation matrix between dimensions is constructed. Assuming there are N flight parameter dimensions, an N×N correlation matrix is ​​constructed. The element in the i-th row and j-th column of the matrix is ​​obtained by calculating the Pearson cross-correlation coefficient between the i-th and j-th dimension acceleration change features on the time axis. The calculation window for the cross-correlation coefficient is consistent with the switching time window to ensure the capture of complete coordinated change patterns. The diagonal elements of this matrix are 1, and the off-diagonal elements range from -1 to 1. Positive values ​​indicate synchronous changes, and negative values ​​indicate opposite changes.

[0069] The coupling correlation matrix is ​​decomposed using eigenvalues, and the eigenvalues ​​are sorted from largest to smallest. The eigenvectors corresponding to the top three dominant eigenvalues ​​are selected as the dominant coupling direction vectors. These eigenvectors represent the main patterns of coordinated changes in multidimensional flight parameters. Each component of the dominant coupling direction vector is used as a weight coefficient, and a weighted sum is performed with the acceleration change feature of the corresponding dimension to obtain three dimensionality-reduced comprehensive feature sequences. Specifically, the i-th component of the k-th dominant eigenvector is the weight coefficient of the i-th dimension flight parameter in the k-th comprehensive feature. The absolute value of this coefficient reflects the degree of contribution of the corresponding dimension parameter to the dominant pattern, and the sign indicates the direction of contribution. These comprehensive feature sequences retain the most significant dynamic evolution information in the original multidimensional data and achieve an effective mapping from high to low dimensions, forming the final envelope transfer feature vector, which is used by the subsequent airborne deep neural network to calculate the double-envelope confidence distribution.

[0070] The envelope transfer feature extraction method described above can significantly reduce feature dimensionality while preserving key dynamic information, thereby alleviating the computational burden on airborne deep neural networks and improving real-time inference efficiency.

[0071] In one optional implementation, the step of inputting the envelope transfer features into the airborne deep neural network to output the dual-envelope confidence distribution of the current and target flight envelopes, and performing a weighted query on the corresponding fault feature databases according to the proportion of the dual-envelope confidence distributions to obtain the transition prediction results includes:

[0072] The envelope transfer features are input into the current envelope recognition branch and the target envelope recognition branch of the deep neural network on the airborne side, and the current flight envelope confidence vector and the target flight envelope confidence vector are output respectively to form a double envelope confidence distribution;

[0073] The current flight envelope confidence vector and the target flight envelope confidence vector are input into the envelope interaction layer in the deep neural network to calculate the transfer stress tensor, which characterizes the dynamic coupling strength between the two envelopes;

[0074] Based on the dual-envelope confidence distribution, a first candidate fault set and a second candidate fault set are retrieved from the first fault feature library corresponding to the current flight envelope and the second fault feature library corresponding to the target flight envelope, respectively.

[0075] A third set of candidate faults is obtained by retrieving the transfer fault feature library based on the transfer stress tensor. The transfer fault feature library stores fault modes that only occur during the envelope transfer process.

[0076] The first candidate fault set, the second candidate fault set, and the third candidate fault set are weighted and fused to obtain the transition prediction result. The weight coefficients are jointly determined by the double envelope confidence distribution and the transferred stress tensor.

[0077] For example, the airborne deep neural network employs a hybrid architecture of convolutional neural networks and long short-term memory networks. The network input layer receives envelope transfer features, which are then processed through three one-dimensional convolutional layers to extract local temporal patterns. The kernel sizes are 3, 5, and 7, respectively, followed by batch normalization and a ReLU activation function at each layer. The output of the convolutional layers enters a bidirectional long short-term memory network layer to capture bidirectional temporal dependencies. A dual-branch structure branches off from this layer, with the current envelope recognition branch and the target envelope recognition branch each containing two fully connected layers. Finally, a confidence vector is output through a softmax activation function. The envelope interaction layer uses bilinear pooling to perform tensor multiplication operations.

[0078] The network training employs supervised learning. Training data is derived from historical flight records, with each sample containing an envelope transition feature sequence and its corresponding current envelope label, target envelope label, and actual fault type label. The loss function consists of three parts: cross-entropy loss for the current envelope classification, cross-entropy loss for the target envelope classification, and weighted cross-entropy loss for fault prediction, which are summed according to a preset weighting ratio. Gradient descent optimization is performed using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. Training continues until the validation set loss converges or the maximum number of training epochs is reached.

[0079] After acquiring the envelope transition features, the airborne deep neural network needs to simultaneously identify the current flight envelope and the target flight envelope it is about to enter. This deep neural network employs a dual-branch structure, with the envelope transition features simultaneously input into both the current envelope identification branch and the target envelope identification branch. The current envelope identification branch outputs a current flight envelope confidence vector of dimension N, where N is the total number of predefined flight envelope types, such as cruise, climb, dive, and hover. Each element in the vector ranges from 0 to 1, representing the probability of belonging to the corresponding envelope type. The target envelope identification branch also outputs a target flight envelope confidence vector of dimension N, reflecting the probability distribution of the aircraft's upcoming transition to various envelope types. These two confidence vectors together constitute a dual-envelope confidence distribution, fully describing the envelope transition state.

[0080] To capture the dynamic coupling between the two envelopes, the deep neural network incorporates a dedicated envelope interaction layer. This layer receives the current flight envelope confidence vector and the target flight envelope confidence vector as input, and calculates an N×N dimensional transfer stress tensor through tensor multiplication. The element in the i-th row and j-th column of this tensor represents the dynamic stress intensity experienced by the system when transitioning from the i-th envelope to the j-th envelope. The transfer stress tensor can quantify the load differences caused to airborne equipment by different envelope transition paths; for example, the stress generated by directly entering a rapid dive from level flight is much higher than that generated by gradual altitude adjustment.

[0081] The airborne system maintains three types of fault feature databases. The first database is associated with the current flight envelope, storing typical fault modes and their feature vectors within that envelope. The second database is associated with the target flight envelope, storing fault features that may occur under the steady-state conditions of the target envelope. The third type of transition fault feature database specifically records special faults that occur only during the transient process of envelope switching. These faults will not occur in a single steady-state envelope, such as seal failure caused by rapid pressurization or structural fatigue caused by rapid load changes.

[0082] During the retrieval process, based on the top K envelope types with the highest confidence in the current flight envelope confidence vector, corresponding fault records are extracted from the first fault feature library to form a first candidate fault set. Similarly, a second candidate fault set is obtained by retrieving the target flight envelope confidence vector from the second fault feature library. The retrieval of the transfer fault feature library is based on the transfer stress tensor. Envelope transfer paths with values ​​exceeding a preset stress threshold are selected, and transfer-specific faults associated with these paths are extracted to form a third candidate fault set. This threshold is determined based on historical fault data statistics. For example, when the stress intensity exceeds 0.6, the corresponding envelope transfer path shows a significantly increased fault incidence rate in historical data; therefore, the threshold can be set to 0.6. For different aircraft models and equipment types, this threshold can be adjusted within the range of 0.5 to 0.8 based on actual operating experience and safety margin requirements.

[0083] The final weighted fusion employs an adaptive weighting mechanism. The weight of the first candidate fault set is proportional to the maximum value of the current flight envelope confidence vector; the weight of the second candidate fault set is proportional to the maximum value of the target flight envelope confidence vector; and the weight of the third candidate fault set is proportional to the Frobenius norm of the transferred stress tensor. After normalization, these three weights are multiplied and summed with the original confidence scores of the faults in each candidate set to obtain a comprehensive fault confidence ranking. The fault type with the highest confidence score is output as the transitional prediction result.

[0084] This invention identifies the current state and transfer trend simultaneously through a dual-envelope confidence distribution, and combines the transfer stress tensor to quantify dynamic loads, which can effectively capture transient-specific fault modes and improve the accuracy and comprehensiveness of fault prediction during the envelope transfer phase.

[0085] In one optional implementation, the step of calculating the marginal information gain value of each data segment in the time-series state data to the transition prediction result, selecting a preset number of data segments from high to low according to the marginal information gain value when communication is restricted, and transmitting the envelope transfer feature and the data segments to the ground end includes:

[0086] The time-series state data is divided into multiple data segments in chronological order, and the statistical distribution characteristics of each data segment are extracted.

[0087] For each fault type in the transition prediction results, a conditional probability mapping relationship is constructed. The statistical distribution characteristics of each data segment are used as the observation input to the conditional probability mapping relationship. The difference in the predicted probability of the fault type when the data segment exists and when it does not exist is calculated.

[0088] The difference in predicted probabilities is used as the single information contribution of the data segment to the fault type. The marginal information gain value of the data segment is obtained by weighting and summing the single information contributions of the individual data segment to all fault types included in the transition prediction result. The weighting coefficient is determined by the initial probability of each fault type in the transition prediction result.

[0089] In a communication-restricted state, all data segments are sorted in descending order according to the marginal information gain value, and a preset number of data segments at the top of the sorting are selected to form a dataset to be transmitted.

[0090] The envelope transfer feature and the dataset to be transmitted are sent to the ground terminal.

[0091] Combination Figure 2 The flowchart illustrating the calculation of marginal information gain based on the statistical characteristics of time-series data segments is as follows: After acquiring complete time-series state data of the airborne equipment within the switching time window, it is divided into multiple data segments at fixed time intervals, with each data segment containing several sampling points. The time interval is selected to ensure the continuity of state parameters within the data segment, typically set to 0.5 to 2 seconds. For each data segment, its statistical distribution characteristics are extracted, including the mean, variance, peak value, and standard deviation of the rate of change of each state parameter within the segment. These statistical characteristics can characterize the state evolution pattern of the segment.

[0092] For each fault type identified in the airborne transition prediction results, a conditional probability mapping relationship is constructed. This mapping relationship uses the statistical distribution characteristics of the data segment as input variables and outputs the conditional probability of the fault type occurring. Specifically, a Bayesian network is used to establish the probabilistic dependency between features and fault types. The Bayesian network is trained to determine the conditional probability table between each statistical feature node and the fault type node. Training data comes from data segments labeled with fault types in historical flight records, and the probability distribution of each feature under each fault type is statistically analyzed. During inference, the statistical feature values ​​of the data segment to be evaluated are substituted into the network, and the posterior probability is calculated according to the Bayesian formula, i.e., the conditional probability of the fault type occurring given the feature combination. The mapping process is the Bayesian inference calculation from feature vectors to probability values. For a specific data segment, its statistical distribution characteristics are input into the mapping relationship to obtain the predicted probability of the fault type when the segment is included; then, assuming the segment is missing, interpolated data from adjacent segments are used as substitutes, and the predicted probability is calculated again. The difference between the two calculated probability values ​​is the single information contribution of the data segment to the fault type; the larger the difference, the more critical the segment is to determining the fault type.

[0093] A single data segment may simultaneously influence multiple fault types. The weighted sum of the individual information contributions of this segment across all fault types yields a comprehensive marginal information gain. The weighting coefficients are determined by the initial probability of each fault type in the transition prediction results; fault types with higher initial probabilities have greater weights, reflecting their urgency. The calculation formula involves multiplying the individual information contribution of each fault type by its corresponding initial probability and then summing the results, ensuring that the marginal information gain comprehensively reflects the overall importance of the data segment.

[0094] When a restricted air-to-ground communication link is detected, such as bandwidth dropping below 30% of normal or data transmission delay exceeding a threshold, a data filtering mechanism is activated. All data fragments are sorted in descending order based on their calculated marginal information gain values, and the top-ranked fragments are selected to form the dataset to be transmitted. The preset number is dynamically adjusted based on the currently available bandwidth, typically selecting the maximum number of fragments that can be transmitted to ensure transmission is completed within bandwidth constraints. The envelope transfer characteristics are packaged with the filtered dataset to be transmitted and sent to the ground terminal via the air-to-ground data link, enabling priority transmission of high-value information under communication-constrained conditions.

[0095] This invention uses a marginal information gain filtering mechanism to prioritize the transmission of the most critical data segments for fault diagnosis when communication is limited, ensuring that the ground end can still obtain enough information for accurate prediction under bandwidth-constrained conditions.

[0096] In one alternative implementation, the step of calculating the marginal information gain value of the data segment for the transition prediction result includes:

[0097] A causal dependency graph is constructed between flight parameters based on the time-series state data. The causal dependency graph identifies the causal precedence relationship between parameters through Granger causality test. If the historical value of parameter A has predictive ability for the current value of parameter B, a directed edge from A to B is established.

[0098] The causal centrality of each data segment is calculated based on the causal dependency graph, and the causal centrality measures the range of influence of the flight parameters contained in the data segment in the entire causal network.

[0099] The statistical distribution characteristics of each data segment are mapped to the causal centrality input conditional probability, and the conditional entropy difference between the predicted probability of fault type when the data segment exists and does not exist is calculated.

[0100] The conditional entropy difference and the causal centrality are nonlinearly weighted and fused to obtain the causal enhancement marginal information gain value. The nonlinear weighting is implemented through a gating mechanism, which amplifies the data segments corresponding to the causal key nodes and suppresses the data segments corresponding to the causal edge nodes.

[0101] For example, when calculating the marginal information gain of a data segment on the transition prediction result, a causal dependency graph between flight parameters is first constructed from the collected time-series state data. This causal dependency graph identifies the causal lead relationship between parameters through Granger causality tests. Specifically, for multiple flight parameters contained in the time-series state data, for any parameter pair A and B, the historical value sequence of parameter A over the past τ time steps and the current value of parameter B are extracted to establish a vector autoregressive model. The vector autoregressive model takes the following form. B t Let β be the value of parameter B at time t. k and γ k For regression coefficients, The error term is defined. The regression coefficients are estimated using the least squares method. A complete model incorporating historical values ​​of parameter A and a restricted model using only historical values ​​of parameter B are established. The difference between the sums of squared residuals of the two models is used to construct an F-statistic for testing. If, after incorporating historical values ​​of parameter A, the predicted sum of squared residuals of the current value of parameter B is significantly lower than when only historical values ​​of parameter B are used, and the p-value of the F-statistic test result is less than 0.05, then parameter A is considered to have a Granger causal relationship with parameter B. A directed edge from node A to node B is established in the causal dependency graph, with the edge weight set to the normalized value of the F-statistic.

[0102] The causal centrality of each data segment is calculated based on the constructed causal dependency graph. The time-series state data is divided into multiple data segments of fixed length, each containing measurements of several flight parameters within that time period. For the i-th data segment, the set of nodes corresponding to all flight parameters is counted, and the sum of the out-degree and in-degree of these nodes in the causal dependency graph is calculated as a local influence indicator. Simultaneously, the PageRank algorithm is used to calculate the global importance score of each node, reflecting its ability to influence other nodes through causal chains. The causal centrality Ci of all parameter nodes within a data segment is obtained by weighting the global importance scores of all nodes within that data segment, with the weight being the proportion of each node's local influence.

[0103] For each data segment, its statistical distribution features are extracted, including the mean, variance, skewness, and cross-correlation matrix of each parameter within the segment. This implementation adds the skewness and cross-correlation matrix to more comprehensively characterize the distribution characteristics of the data segment. These statistical features and the causal centrality Ci are input into a pre-trained conditional probability mapping model. This model is an improved version based on the aforementioned Bayesian network mapping, adding causal centrality as an additional input node in addition to the original statistical feature nodes. During training, the extended conditional probability table is learned using historical fault samples, enabling the model to simultaneously consider statistical characteristics and causal structure information. This model, learned from historical fault samples, can estimate the conditional entropy of the fault type prediction probability distribution given the presence or absence of a data segment. The conditional entropy H(Y|Di) when the i-th data segment is included and the conditional entropy H(Y) when the i-th data segment is not included are calculated, and the difference ΔHi is the information gain brought by the data segment.

[0104] The conditional entropy difference ΔHi and causal centrality Ci are nonlinearly weighted and fused. A gating mechanism is used to implement this fusion process, defining the gating function gi = σ(w1Ci + w2ΔHi + b), where σ is the Sigmoid activation function, and w1, w2, and b are learnable parameters. The final causal enhancement marginal information gain is calculated as I. i =g i ·ΔHi+( )·αC i , where α is the scaling factor. When the flight parameter node corresponding to a data segment is in a critical position in the causal dependency graph, the causal centrality Ci is large, and the gating function outputs a higher weight, causing the segment to receive gain amplification; conversely, the data segment corresponding to the causal edge node has a smaller Ci, resulting in a lower gating weight and a gain suppression effect, thereby realizing intelligent segment selection based on causal relationships.

[0105] This invention uses a filtering mechanism that integrates causal information to more accurately locate data segments that have a fundamental impact on fault diagnosis.

[0106] In one optional implementation, the steps of inputting the received envelope transfer features into a ground-side deep neural network to obtain a fast prediction result, and simultaneously using the physical state equations of the airborne equipment to deduce the state evolution of the untransmitted data segments based on the received data segments, and inputting the deduced complete time-series data into the ground-side deep neural network to obtain a complete prediction result include:

[0107] The envelope transfer features are input into the pre-feature extraction layer of the deep neural network on the ground side to obtain the envelope abstract features. The envelope abstract features are directly input into the high-level decision module through a fast path to output fast prediction results. The fast path skips the temporal modeling layer.

[0108] The state values ​​and rates of change of each flight parameter are extracted from the received data segments as initial conditions and substituted into the physical state equation of the airborne equipment. The flight parameter evolution trajectory for the time interval corresponding to the untransmitted data segments is calculated. The flight parameter evolution trajectory and the received data segments are spliced ​​together in chronological order to obtain the reconstructed complete time series data.

[0109] The reconstructed complete time series data is input into the pre-feature extraction layer to obtain the initial time series features. The envelope abstract features are used by the feature modulation module to generate a spatial attention weight map to perform element-wise weighting on the initial time series features to obtain the modulated time series features. The modulated time series features are then input into the high-level decision module after the time series modeling layer extracts the time series dependencies, and the complete prediction results are output.

[0110] Both the rapid prediction result and the complete prediction result include a fault type identifier and a corresponding fault probability distribution.

[0111] For example, after receiving data transmitted from the airborne terminal, the ground-side system immediately initiates a dual-path prediction mechanism. The ground-side deep neural network employs an encoder-decoder architecture. The pre-feature extraction layer is the encoder, consisting of three one-dimensional convolutional layers with kernel sizes of 3, 5, and 7. Each layer is followed by batch normalization and a ReLU activation function, resulting in output channels of 64, 128, and 256, respectively. This encoder maps the input envelope transfer features to a 256-dimensional envelope abstract feature vector. The temporal modeling layer uses a bidirectional long short-term memory network, containing two LSTM layers with 256 hidden units per layer. The high-level decision module consists of two fully connected layers and a Softmax classifier. The first fully connected layer has 128 neurons, and the second layer has the same number of neurons as the total number of fault types. Network training uses a historical flight fault dataset, containing complete temporal data labeled with fault types and occurrence times, along with corresponding envelope transfer features. The cross-entropy loss function, Adam optimizer, learning rate of 0.001, and batch size of 32 are used, training continues until the validation set loss converges. Envelope transition features, as highly compressed state transition information, are first processed by a pre-feature extraction layer of a ground-side deep neural network. This pre-feature extraction layer can extract a 256-dimensional abstract envelope feature vector from the envelope transition features, which encodes the core pattern information of the flight envelope transition.

[0112] To achieve rapid response, a fast path that skips the temporal modeling layer is designed. Envelope abstract features are directly connected to the network's high-level decision module through this path. This module consists of a fully connected layer and a Softmax classifier, capable of outputting rapid prediction results within milliseconds. The rapid prediction results include fault type identifiers, such as category numbers for flight control system response delay, data acquisition module anomalies, and sensor drift. This invention addresses fault prediction for airborne equipment during flight envelope transfer, covering fault types including those of airborne electronic devices and their subsystems such as flight control systems, data acquisition, sensors, and communication modules, rather than specific mechanical systems like hydraulics or fuel systems. This ensures consistency with the AI ​​prediction method of the invention's subject matter. The probability distribution values ​​for each fault type are also provided; for example, [0.72, 0.15, 0.08, 0.05] represent the probabilities of four fault types.

[0113] Simultaneously, the ground-based system initiates a data reconstruction process to process the received data fragments. From these fragments, instantaneous state values ​​of key flight parameters are extracted, including current values ​​of parameters such as altitude, velocity, attitude angles, and angular rates, as well as the rates of change of each parameter calculated through differential methods. These values ​​are used as initial conditions and substituted into the airborne equipment's physical state equations. The airborne equipment's physical state equations employ the aircraft's six-degree-of-freedom rigid body motion equations. This set of equations includes three translational equations and three rotational equations. The translational equations, based on Newton's second law, establish a relationship between the net external forces (including aerodynamic forces, engine thrust, and gravity) acting on the aircraft's center of mass and mass and acceleration, expressed as the acceleration of the center of mass in the three coordinate axes equals the net force in the corresponding direction divided by the aircraft's mass. The rotational equations, based on the rigid body rotation law, establish a relationship between the net torque acting on the aircraft and moment of inertia and angular acceleration, expressed as the angular acceleration about the three coordinate axes equals the net torque in the corresponding axis divided by the corresponding moment of inertia.

[0114] State variables include the aircraft's position coordinates (east, north, altitude) in the inertial coordinate system, velocity components (along the three axes of the body coordinate system), attitude angles (pitch, yaw, roll), and angular rate (rotational speed around the three axes of the body coordinate system). Control inputs include elevator deflection angle, rudder deflection angle, aileron deflection angle, and engine throttle opening. Aerodynamic forces are obtained through a lookup table method. Based on the current angle of attack, sideslip angle, Mach number, and control surface deflection angle, lift coefficient, drag coefficient, and moment coefficient are interpolated from a pre-stored aerodynamic coefficient table, and then multiplied by dynamic pressure and reference area to calculate the specific aerodynamic force and moment values. Engine thrust is obtained by interpolation from a pre-stored engine characteristic diagram based on the current altitude, speed, and throttle opening. Gravity is calculated by multiplying the aircraft's mass by gravitational acceleration, with the direction pointing towards the Earth's center.

[0115] The evolution trajectory of each flight parameter within the untransmitted time interval is calculated using numerical integration methods. For example, if data from seconds 1 to 3 and seconds 7 to 10 are received, the state changes from seconds 4 to 6 need to be deduced. A fourth-order Runge-Kutta method is used to ensure accuracy during the deduction, with the right-hand side of the above equations (the formulas for calculating each acceleration and angular acceleration) used as the derivative function. The deduced flight parameter evolution trajectory is then progressively advanced from the known initial state and control input sequence, with a step size of 0.1 seconds. The deduced flight parameter evolution trajectory is then concatenated with the actually received data segments in timestamp order to form reconstructed time-series data covering the complete time span.

[0116] The reconstructed complete temporal data is fed into the same pre-feature extraction layer, yielding 512-dimensional initial temporal features. To fuse envelope information, a feature modulation module is designed to process the abstract envelope features. This module maps the 256-dimensional envelope features into a spatial attention weight map of the same dimension as the initial temporal features through a two-layer fully connected network. Each element of the attention weight map takes a value between 0 and 1, representing the importance of the corresponding feature channel. This weight map is then multiplied element-wise with the initial temporal features to obtain the modulated temporal features, highlighting the feature dimensions related to envelope transformation.

[0117] Modulated temporal features are input into the temporal modeling layer for deep processing. The temporal modeling layer employs a bidirectional long short-term memory network, containing two LSTM layers with 256 hidden units per layer. The outputs of the forward and backward LSTMs are concatenated to form a 512-dimensional feature vector, extracting long-term dependencies and local patterns from the temporal data. This high-level temporal representation is then fed into the higher-level decision module, outputting a complete prediction result. The complete prediction result also includes fault type identification and probability distribution; however, due to the use of complete temporal information, its prediction accuracy is typically significantly higher than that of fast prediction.

[0118] The fault type identifiers of the two prediction results are compared. If the highest probability category of the rapid prediction matches the highest probability category of the complete prediction, for example, both being "flight control system response delay," then the data reconstruction based on the physical equations is confirmed to be effective, and the inferred untransmitted data fragments can accurately reflect the actual state evolution. At this point, the ground station retrieves matching maintenance guidelines, spare parts allocation suggestions, and emergency operating procedures from the support plan database based on the fault type and severity in the complete prediction results, generates a targeted remote support plan, and transmits it back to the airborne terminal.

[0119] The dual-path prediction mechanism of this invention achieves millisecond-level preliminary judgment through a fast path, ensuring rapid emergency response. Simultaneously, it provides high-precision verification through complete time-series analysis, ensuring the accuracy of the support plan. The data reconstruction method based on physical state equations effectively reduces communication bandwidth requirements and maintains prediction performance even under constrained channel conditions.

[0120] In one alternative implementation, the steps of verifying the validity of the data reconstruction and generating a supporting scheme include:

[0121] Determine whether the rapid prediction result and the complete prediction result are consistent in fault type identification and fault probability distribution. If they are consistent, confirm that the data reconstruction is effective and generate a support plan.

[0122] When there is a discrepancy, the contribution of each time period in the reconstructed complete time series data to the prediction difference is calculated, and the time periods are sorted in descending order of contribution. A preset number of time periods at the top of the sorted list are identified as belonging to the time interval corresponding to the untransmitted data segment.

[0123] If a preset number of time periods at the top of the sorting list belong to the untransmitted interval, the original data segments of these time periods are requested to be retransmitted from the airborne terminal and the corresponding inferred data is replaced. If they belong to the transmitted interval, the boundary conditions or solution accuracy parameters of the physical state equation of the airborne equipment in these time periods are adjusted and the untransmitted data segments are re-inferred.

[0124] The corrected complete time series data is re-input into the ground-side deep neural network to obtain the corrected complete prediction result, and consistency is checked with the fast prediction result. The correction process is iteratively executed until consistency is achieved or the preset maximum number of iterations is reached.

[0125] A second aspect of the present invention provides an AI-based airborne fault prediction and remote support system, comprising:

[0126] The envelope transfer unit is used to acquire the time-series status data of airborne equipment. When the change in flight parameters exceeds the preset change threshold, the second time derivative of the time-series status data within the switching time window is extracted as the envelope transfer feature.

[0127] The dual-envelope unit is used to input the envelope transfer features into the airborne deep neural network and output the dual-envelope confidence distribution of the current flight envelope with the target. The corresponding fault feature databases are then weighted according to the proportion of the dual-envelope confidence distribution to obtain the transition prediction result.

[0128] Marginal information unit is used to calculate the marginal information gain value of each data segment in the time-series state data to the transition prediction result. When communication is restricted, a preset number of data segments are selected from high to low according to the marginal information gain value, and the envelope transfer feature and the data segments are sent to the ground end.

[0129] The data reconstruction unit is used to input the received envelope transfer features into the ground-side deep neural network to obtain a fast prediction result. At the same time, based on the received data segments, it uses the physical state equation of the airborne equipment to deduce the state evolution of the untransmitted data segments. The deduced complete time series data is input into the ground-side deep neural network to obtain a complete prediction result. When the fault type identifier of the fast prediction result is consistent with that of the complete prediction result, the data reconstruction is confirmed to be effective and a support plan is generated.

[0130] For example, in the process of confirming the validity of data reconstruction and generating support solutions, after receiving the rapid prediction results and the complete prediction results, the ground station first performs a consistency determination between the two. This determination includes two dimensions: the matching of fault type identifiers and the similarity of fault probability distributions. The matching of fault type identifiers is achieved by directly comparing the fault codes output by the rapid prediction results and the complete prediction results. When the main fault type codes of the two are the same, the types are determined to be consistent. The consistency of fault probability distributions is evaluated by calculating the KL divergence of the two probability vectors. The KL divergence is calculated as the relative entropy of the two probability distributions P and Q. Specifically, for each component i of P, the product of the i-th component of P and the logarithm of the i-th component of P is calculated, and these products of all components are summed to obtain the first summation result; then the product of the i-th component of P and the logarithm of the i-th component of Q is calculated, and these products of all components are summed to obtain the second summation result; finally, the second summation result is subtracted from the first summation result. When the KL divergence is less than 0.15, the distributions are considered to be consistent. Only when the fault type identifier and probability distribution simultaneously meet the consistency condition is the data reconstruction confirmed to be effective. Then, based on the complete prediction results, the standard maintenance procedure corresponding to the fault code is retrieved, and a support plan including diversion suggestions, emergency operation guidelines and ground preparation checklist is generated in combination with the current flight location information.

[0131] When a discrepancy arises between the rapid prediction result and the complete prediction result, it is necessary to pinpoint the time period causing the prediction difference. Gradient backtracking analysis is performed on the complete prediction network to calculate the gradient magnitude of the input data for each time period on the final prediction output. Specifically, the partial derivatives of the loss function of the network output layer with respect to the fault type probability are calculated with respect to the data at each time step of the input layer. The absolute values ​​of these partial derivatives are used as the contribution of that time period to the prediction difference. All time periods are sorted from largest to smallest contribution, and the top three time periods are identified as key sources of difference. The time indices of these three time periods are checked to determine if they fall within the time interval corresponding to untransmitted data segments. The determination method is as follows: the start and end timestamps of the key time periods are matched with the list of timestamps of data segments sent by the airborne terminal. If more than 50% of the data points within a time period are not in the transmitted list, then that time period is determined to belong to the untransmitted interval.

[0132] For critical time periods belonging to the untransmitted range, the ground station generates a targeted retransmission request message, containing the precise time range and data channel identifier to be retransmitted. Upon receiving the request, the airborne station reads the raw sensor data for the corresponding time period from its onboard memory and transmits it back to the ground station via the data link. The ground station replaces the data for the corresponding time period previously derived from the physical state equations with the received raw data segments, forming partially updated complete time-series data. For critical time periods belonging to the already transmitted range, this indicates a deviation in the derivation process itself. In this case, the solution parameters of the physical state equations are adjusted, specifically: the boundary conditions at the starting point of the time period are changed from interpolated estimates to the measured values ​​of the nearest neighbor data points actually transmitted; and the numerical integration step size is reduced from the default 0.1 seconds to 0.02 seconds to improve solution accuracy before re-deriving all untransmitted data segments.

[0133] The corrected complete time-series data is then input again into the ground-side deep neural network to obtain the corrected complete prediction result. This corrected result is then compared with the rapid prediction result using the aforementioned method for consistency verification. If inconsistency remains, the iterative process of contribution calculation, interval determination, and data correction is repeated. To prevent infinite loops, the maximum number of iterations is set to 5. If consistency is still not achieved after 5 iterations, the system generates an uncertainty warning and simultaneously sends the rapid prediction result, the final corrected complete prediction result, and the confidence difference between the two to ground maintenance experts for manual review, ensuring the reliability of critical safety decisions.

[0134] This invention reduces additional communication overhead while ensuring prediction accuracy through a directional retransmission strategy; the iterative correction process ensures the reliability of prediction results in complex flight envelope transfer scenarios.

[0135] A third aspect of the present invention provides an electronic device, comprising:

[0136] processor;

[0137] Memory used to store processor-executable instructions;

[0138] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0139] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0140] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based method for airborne fault prediction and remote support, characterized in that, include: Acquire the time-series status data of airborne equipment. When the change in flight parameters exceeds the preset change threshold, extract the second time derivative of the time-series status data within the switching time window as the envelope transition feature. The envelope transfer features are input into the airborne deep neural network, which outputs the dual-envelope confidence distribution of the current and target flight envelopes. The corresponding fault feature databases are then weighted according to the proportion of the dual-envelope confidence distribution to obtain the transition prediction results. Calculate the marginal information gain value of each data segment in the time-series state data to the transition prediction result. When communication is restricted, select a preset number of data segments from high to low according to the marginal information gain value, and send the envelope transfer feature and the data segments to the ground end. The ground end inputs the received envelope transfer features into the ground-side deep neural network to obtain a fast prediction result. At the same time, based on the received data segments, it uses the physical state equation of the airborne equipment to deduce the state evolution of the untransmitted data segments. The deduced complete time series data is input into the ground-side deep neural network to obtain a complete prediction result. When the fault type identifier of the fast prediction result is consistent with that of the complete prediction result, the data reconstruction is confirmed to be effective and a support plan is generated.

2. The method according to claim 1, characterized in that, The steps for extracting the second-order time derivative of time-series state data within the switching time window as envelope transition features include: The first-order time derivatives of the multidimensional flight parameters in the time-series state data within the switching time window are calculated to obtain the instantaneous rate of change sequence of each dimension. The instantaneous rate of change sequence is subjected to second-order time differential operation to obtain acceleration change characteristics in each dimension, which characterize the dynamic evolution trend during the flight envelope transfer process; A multi-dimensional coupling correlation matrix is ​​constructed, which describes the cooperative change pattern among different flight parameters by calculating the cross-correlation coefficient of the acceleration change characteristics in the time dimension; The dominant coupling direction vector is extracted by performing eigenvalue decomposition on the coupling correlation matrix. The dominant coupling direction vector is then weighted and fused with the acceleration change features of the corresponding dimension to obtain the dimensionality-reduced envelope transfer features.

3. The method according to claim 1, characterized in that, The steps of inputting the envelope transfer features into the airborne deep neural network to output the double-envelope confidence distribution of the current and target flight envelopes, and performing a weighted query on the corresponding fault feature databases according to the proportion of the double-envelope confidence distribution to obtain the transition prediction result include: The envelope transfer features are input into the current envelope recognition branch and the target envelope recognition branch of the deep neural network on the airborne side, and the current flight envelope confidence vector and the target flight envelope confidence vector are output respectively to form a double envelope confidence distribution; The current flight envelope confidence vector and the target flight envelope confidence vector are input into the envelope interaction layer in the deep neural network to calculate the transferred stress tensor; Based on the dual-envelope confidence distribution, a first candidate fault set and a second candidate fault set are retrieved from the first fault feature library corresponding to the current flight envelope and the second fault feature library corresponding to the target flight envelope, respectively. A third set of candidate faults is obtained by retrieving the transfer fault feature library based on the transfer stress tensor. The transfer fault feature library stores fault modes that only occur during the envelope transfer process. The first candidate fault set, the second candidate fault set, and the third candidate fault set are weighted and fused to obtain the transition prediction result. The weight coefficients are jointly determined by the double envelope confidence distribution and the transferred stress tensor.

4. The method according to claim 1, characterized in that, The steps for calculating the marginal information gain of each data segment in the time-series state data for the transition prediction result include: The time-series state data is divided into multiple data segments in chronological order, and the statistical distribution characteristics of each data segment are extracted. For each fault type in the transition prediction results, a conditional probability mapping relationship is constructed. The statistical distribution characteristics of each data segment are used as the observation input to the conditional probability mapping relationship. The difference in the predicted probability of the fault type when the data segment exists and when it does not exist is calculated. The difference in predicted probabilities is used as the individual information contribution of the data segment to the fault type. The marginal information gain value of the data segment is obtained by weighted summation of the individual information contributions of the data segment to all fault types included in the transition prediction result.

5. The method according to claim 4, characterized in that, The steps for calculating the marginal information gain of the data segment on the transition prediction result include: A causal dependency graph is constructed between flight parameters based on the time-series state data. The causal dependency graph identifies the causal precedence relationship between parameters through Granger causality test. If the historical value of parameter A has predictive ability for the current value of parameter B, a directed edge from A to B is established. The causal centrality of each data segment is calculated based on the causal dependency graph, and the causal centrality measures the range of influence of the flight parameters contained in the data segment in the entire causal network. The statistical distribution characteristics of each data segment are mapped to the causal centrality input conditional probability, and the conditional entropy difference between the predicted probability of fault type when the data segment exists and does not exist is calculated. The conditional entropy difference and the causal centrality are nonlinearly weighted and fused to obtain the causal enhancement marginal information gain value. The nonlinear weighting is implemented through a gating mechanism, which amplifies the data segments corresponding to the causal key nodes and suppresses the data segments corresponding to the causal edge nodes.

6. The method according to claim 1, characterized in that, The steps of inputting the received envelope transfer features into a ground-side deep neural network to obtain a fast prediction result, and simultaneously using the physical state equations of the airborne equipment to deduce the state evolution of the untransmitted data segments based on the received data segments, and inputting the deduced complete time-series data into the ground-side deep neural network to obtain a complete prediction result include: The envelope transfer features are input into the pre-feature extraction layer of the deep neural network on the ground side to obtain the envelope abstract features. The envelope abstract features are directly input into the high-level decision module through a fast path to output fast prediction results. The fast path skips the temporal modeling layer. The state values ​​and rates of change of each flight parameter are extracted from the received data segments as initial conditions and substituted into the physical state equation of the airborne equipment. The flight parameter evolution trajectory for the time interval corresponding to the untransmitted data segments is calculated. The flight parameter evolution trajectory and the received data segments are spliced ​​together in chronological order to obtain the reconstructed complete time series data. The reconstructed complete time series data is input into the pre-feature extraction layer to obtain initial time series features. The envelope abstract features are used by the feature modulation module to generate a spatial attention weight map to perform element-wise weighting on the initial time series features to obtain modulated time series features. The modulated time series features are then processed by the time series modeling layer to extract time series dependencies and input into the high-level decision module to output complete prediction results.

7. The method according to claim 6, characterized in that, The steps to confirm the effectiveness of data reconstruction and generate supporting solutions include: Determine whether the rapid prediction result and the complete prediction result are consistent in fault type identification and fault probability distribution. If they are consistent, confirm that the data reconstruction is effective and generate a support plan. When there is a discrepancy, the contribution of each time period in the reconstructed complete time series data to the prediction difference is calculated, and the time periods are sorted in descending order of contribution. A preset number of time periods at the top of the sorted list are identified as belonging to the time interval corresponding to the untransmitted data segment. If a preset number of time periods at the top of the sorting list belong to the untransmitted interval, the original data segments of these time periods are requested to be retransmitted from the airborne terminal and the corresponding inferred data is replaced. If they belong to the transmitted interval, the boundary conditions or solution accuracy parameters of the physical state equation of the airborne equipment in these time periods are adjusted and the untransmitted data segments are re-inferred. The corrected complete time series data is re-input into the ground-side deep neural network to obtain the corrected complete prediction result, and consistency is checked with the fast prediction result. The correction process is iteratively executed until consistency is achieved or the preset maximum number of iterations is reached.

8. An AI-based airborne fault prediction and remote support system, used to implement the method of any one of claims 1-7, characterized in that, include: The envelope transfer unit is used to acquire the time-series status data of airborne equipment. When the change in flight parameters exceeds the preset change threshold, the second time derivative of the time-series status data within the switching time window is extracted as the envelope transfer feature. The dual-envelope unit is used to input the envelope transfer features into the airborne deep neural network and output the dual-envelope confidence distribution of the current flight envelope with the target. The corresponding fault feature databases are then weighted according to the proportion of the dual-envelope confidence distribution to obtain the transition prediction result. Marginal information unit is used to calculate the marginal information gain value of each data segment in the time-series state data to the transition prediction result. When communication is restricted, a preset number of data segments are selected from high to low according to the marginal information gain value, and the envelope transfer feature and the data segments are sent to the ground end. The data reconstruction unit is used to input the received envelope transfer features into the ground-side deep neural network to obtain a fast prediction result. At the same time, based on the received data segments, it uses the physical state equation of the airborne equipment to deduce the state evolution of the untransmitted data segments. The deduced complete time series data is input into the ground-side deep neural network to obtain a complete prediction result. When the fault type identifier of the fast prediction result is consistent with that of the complete prediction result, the data reconstruction is confirmed to be effective and a support plan is generated.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images