An aircraft actuator fault prediction method based on future guided learning

Abstract

The application discloses an aircraft actuator fault prediction method based on future guidance learning and belongs to the technical field of flight control system fault prediction; the method collects actuator control instructions, actual responses, flight state parameters and state monitoring parameters, and constructs a historical observation window; in the training stage, future guidance information is constructed by using fault evolution information in a future time window, and a historical state coding network learns the correlation between current weak anomalies and future faults through cross-time domain feature constraints; in the online prediction stage, the future evolution guidance network does not participate in reasoning, and only based on current and historical observation data, a multi-step fault prediction module outputs fault occurrence probability, fault type, degradation degree and early warning level in multiple time scales. The application can perceive future fault risks in advance from current weak anomaly signals under the condition that no future data is used in the reasoning stage, and improves the advance, accuracy and real-time applicability of fault prediction.

Description

Technical Field

[0001] This invention belongs to the fields of flight control system fault prediction, actuator predictive maintenance, and artificial intelligence time series modeling technology, specifically involving a method for predicting aircraft actuator faults based on future guided learning. Background Technology

[0002] Aircraft actuators are critical actuators in the flight control system. Their function is to drive the movement of aerodynamic control surfaces such as ailerons, elevators, rudders, flaps, and spoilers according to control commands output by the flight control system, thereby achieving aircraft attitude control, trajectory tracking, and maneuvering. The operating status of actuators directly affects the response speed, control accuracy, and safety margin of the flight control system. During actual flight, aircraft actuators operate under complex loads, vibrations, temperature variations, and aerodynamic disturbances, making them susceptible to factors such as mechanical wear, hydraulic pressure fluctuations, electrical control anomalies, sensor drift, increased actuator clearances, and changes in frictional characteristics. This can lead to malfunctions such as jamming, hysteresis, bias, drift, response delay, gain loss, efficiency reduction, intermittent failure, or complete failure. If actuator malfunctions are not identified and predicted in a timely manner, it may result in increased control surface tracking errors, reduced compensation capabilities of the flight control system, and in severe cases, compromise aircraft flight safety.

[0003] Existing methods for aircraft actuator fault diagnosis and prediction mainly include threshold-based methods, mechanism-based methods, state observer-based methods, residual analysis-based methods, and machine learning or deep learning-based methods. However, most existing deep learning methods identify faults based on the current moment or historical time windows, primarily learning the mapping relationship between "current observation data" and "current fault labels," and failing to adequately utilize how current weak anomalies evolve into obvious faults in the future. Especially in the early stages of aircraft actuator faults, the fault often does not manifest as obvious amplitude anomalies, but rather as a slight increase in response delay, a slight enhancement of residual fluctuations, a slow decrease in control tracking accuracy, local friction changes, or intermittent response discontinuities. These phenomena may still be within the normal threshold range in the current window, but they may gradually evolve into obvious faults in the future. If the fault evolution information in future time windows cannot be utilized during the model training phase, the model will struggle to establish a correlation between current weak anomalies and future faults, resulting in insufficient prediction lead.

[0004] Existing technologies have at least the following problems: traditional fault diagnosis methods focus more on whether a fault has occurred and lack the ability to actively learn about the future evolution trend of the fault; existing deep learning models mainly rely on historical window features and do not make full use of the future state information available in the training data; under complex flight conditions, changes in flight state, load and control commands will affect the actuator response, causing fault characteristics to be coupled with the disturbance of the operating conditions.

[0005] Therefore, there is an urgent need to propose an aircraft actuator failure prediction technology that can make full use of future failure evolution information during the training phase and rely solely on current and historical data to complete the prediction during the online application phase. Summary of the Invention

[0006] The technical problem to be solved: To avoid the shortcomings of existing technologies, this invention provides an aircraft actuator fault prediction method based on future-guided learning. By constructing future guidance information using fault evolution information within a future time window during the training phase, and employing cross-temporal feature constraints to enable the historical state encoding network to learn the correlation between current weak anomalies and future fault trends, the future evolution guidance network is removed during the online prediction phase, and prediction is completed only based on current and historical observation data. This improves the advance and accuracy of fault prediction while ensuring real-time performance.

[0007] The technical solution of this invention is: a method for predicting aircraft actuator failures based on future guided learning, comprising: Acquire multi-source time-series data of the aircraft actuators at the current and historical times, and construct a historical observation window; The historical observation window is input into a pre-trained future guided learning model, which includes a historical state encoding network and a multi-step fault prediction module. The historical state coding network performs temporal coding on historical observation windows, extracts temporal dependencies, and outputs the current historical state representation. The multi-step fault prediction module outputs the fault prediction results of the aircraft actuators at different future prediction time scales based on the current historical state characterization. The future-guided learning model is obtained through future-guided learning training. During the training phase, future-guided information is constructed using fault evolution information within a future time window. Based on the future-guided information, the historical state encoding network is strengthened to learn the correlation between current weak anomalies and future fault evolution through cross-temporal feature constraints. During the application phase, the future-guided information does not participate in inference calculation.

[0008] A further technical solution of the present invention is: the construction of the historical observation window includes: Collect multi-source timing data during the operation of aircraft actuators. The multi-source timing data includes actuator control commands, actual actuator response data, flight status parameters, and actuator status monitoring parameters. The system performs time synchronization, outlier processing, filtering and noise reduction, operating condition normalization, feature construction, and sliding window division on multi-source time series data to obtain the historical observation window corresponding to the current moment.

[0009] A further technical solution of the present invention is as follows: the input of the historical state coding network further includes an actuator response residual sequence; the actuator response residual sequence is used to characterize the degree of deviation between the expected response and the actual response of the actuator, and is obtained by obtaining the expected response through the actuator nominal response model based on the actuator control command, the actual response data of the actuator and the aircraft flight state parameters, and then performing a difference calculation between the expected response and the actual response of the actuator; the actuator nominal response model is a response model pre-identified based on the actuator normal operation state data, and is used to calculate the expected response of the actuator according to the actuator control command and the aircraft flight state parameters.

[0010] A further technical solution of the present invention is: the multi-step fault prediction module includes a shared feature mapping layer and multiple parallel prediction head branches; The shared feature mapping layer is used to receive the current historical state representation output by the historical state coding network and output shared prediction features through nonlinear mapping. The multiple parallel prediction head branches branch off from the shared prediction features. Each prediction head branch corresponds to a future prediction time scale and is used to output the fault prediction result at the corresponding time scale based on the shared prediction features. The future prediction timescale includes at least two of the following: short-term prediction timescale, medium-term prediction timescale, and long-term prediction timescale. The fault prediction results include at least one of the following for each of the following future prediction time scales: probability of future fault occurrence, fault type, degree of degradation, health status level, fault occurrence lead time, remaining available time, and warning level.

[0011] A further technical solution of the present invention is: each prediction head branch includes a fault classification output layer and a degradation degree regression output layer; The fault classification output layer is used to output the probability of future fault occurrence and fault type at the corresponding future prediction time scale; The degradation degree regression output layer is used to output the future degradation degree at the corresponding future prediction time scale.

[0012] A method for training a future-guided learning model for an aircraft actuator, the future-guided learning model being used to execute the aircraft actuator fault prediction method, comprising: Obtain a training dataset, which contains historical observation windows corresponding to multiple sample times and corresponding future guidance information. The future guidance information is constructed based on fault evolution information within a future time window of a preset length after the current time. An initial model is constructed, which includes a historical state encoding network and a multi-step fault prediction module connected in sequence, as well as a future evolution guidance network set in parallel with the historical state encoding network during the training phase, a cross-temporal feature constraint module connecting the historical state encoding network and the future evolution guidance network, and an early warning decision module for generating early warning levels. Input the historical observation window at the same moment into the historical state coding network to obtain the current historical state representation; input the future guidance information corresponding to that moment into the future evolution guidance network to obtain the future degradation representation; The cross-temporal feature constraint module applies feature space constraints to the current historical state representation and the future degradation representation, enabling the current historical state representation to learn future degradation information; at the same time, the current historical state representation is input into the multi-step fault prediction module for fault prediction. Based on the constraint loss generated by feature space constraints and the prediction loss generated by fault prediction, the model parameters are jointly optimized and iteratively updated until the preset training termination condition is met to obtain a trained future guided learning model. The future evolution guidance network in the trained future guided learning model is removed or deactivated. The early warning decision module generates at least one early warning level from among normal, attention, mild, moderate and severe warnings based on the results of multi-step fault prediction and prediction uncertainty.

[0013] A further technical solution of the present invention is that the future guidance information includes at least one of the following: future residual evolution sequence, future degradation direction vector, future fault label, future degradation degree, future health status, fault occurrence advance amount, and remaining available time; The future residual evolution sequence consists of the response residual between the actual response and the expected response of the actuator within the future time window, and is used to describe the amplitude change, slope change, fluctuation change, cumulative deviation change and continuous abnormal change of the response residual within the future time window. The future degradation direction vector is composed of the residual mean, residual maximum, residual variance, residual change slope, residual accumulation, and response delay change within the future time window. It is used to characterize the direction and extent of the actuator's evolution from a normal state to a fault state by the positive and negative change directions of each degradation feature and their normalized amplitude. The future fault label is the fault type label corresponding to the first abnormal fault state that appears within a future time window. The future degradation degree is a scalar obtained by normalizing and weighting the feature components in the future degradation direction vector, and is used to represent the comprehensive degree of the actuator state deviating from the normal state within the future time window; The future health status is divided into normal, mild deterioration, moderate deterioration, and severe deterioration based on the relationship between the future degree of deterioration and the preset deterioration threshold. The fault occurrence advance and remaining available time are determined based on the time interval between the current moment and the moment when the fault first occurs in the future or when the degree of degradation reaches the preset fault threshold.

[0014] A further technical solution of the present invention is: the future evolution guidance network includes a residual sequence coding layer, a discrete state embedding layer, a continuous feature mapping layer, and a fusion mapping layer; The future evolution guidance network is used to encode and fuse future guidance information, specifically including: The residual sequence coding layer is used to sequence encode the future residual evolution sequence in the future guidance information and extract the residual temporal evolution features within the future time window; The discrete state embedding layer is used to convert the future fault label and future health status in the future guidance information into continuous feature vectors, respectively. The continuous feature mapping layer is used to linearly map the future degradation direction vector, future degradation degree, fault occurrence advance, and remaining available time in the future guidance information; The fusion mapping layer is used to concatenate the residual temporal evolution features, the output of the discrete state embedding layer and the output of the continuous feature mapping layer, and map the concatenated features to a feature space of the same dimension as the current historical state representation, and output the future degradation representation. During the online prediction phase, the future evolution guidance network does not receive input data and does not participate in fault prediction inference calculations.

[0015] A further technical solution of the present invention is: the cross-temporal feature constraint module realizes feature space constraint through at least one of feature consistency constraint, degradation trend constraint and contrastive learning constraint; The feature consistency constraint makes the mapped current historical state representation approximate the corresponding future degenerate representation. The degradation trend constraint ensures that the degradation trend corresponding to the current historical state characterization is consistent with the future degradation direction vector; The contrastive learning constraint makes the current historical state representation and future degradation representation of the same training sample form a positive sample pair, and makes samples corresponding to different fault types, different degradation trends or different health status levels form a negative sample pair, so as to improve the feature separability between different future degradation modes.

[0016] A further technical solution of the present invention is: the loss function adopted by the joint optimization includes at least one of fault classification loss, degradation degree regression loss, feature consistency constraint loss, degradation trend constraint loss, contrastive learning loss and uncertainty calibration loss, and each loss function is calculated based on the corresponding prediction result, the true label and the difference between the current historical state representation and the future degradation representation.

[0017] Beneficial effects The beneficial effects of this invention are as follows: By constructing future guidance information during the model training phase and utilizing cross-temporal feature constraints to enable the historical state encoding network to learn the correlation between current weak anomalies and future fault evolution, this invention allows the model to identify precursory features pointing to future faults from seemingly normal weak anomaly signals during the online prediction phase. Compared to traditional methods that can only identify fault states that are currently exhibiting obvious anomalies, this invention enables the model to identify precursory features pointing to future faults when the fault is in its early evolution stage, thus providing a technical basis for improving fault prediction lead time and allowing maintenance personnel sufficient time for fault handling.

[0018] Compared to existing technologies, this invention does not directly use future data as input for online prediction. Instead, during the training phase, it utilizes residual evolution, degradation direction, fault state, and health state within future time windows to construct future guidance information. This allows the historical state encoding network to learn the correspondence between current weak anomalies and subsequent degradation results. Through feature consistency constraints, the current historical state representation has feature orientations corresponding to future degradation representations. Through degradation trend constraints, the current historical state representation retains its predictive ability for future degradation directions. Through contrastive learning constraints, the separability between different fault evolution modes is improved. Therefore, even without using future data during the online prediction phase, it can still identify weak anomaly signals with future fault orientation based on current and historical observation data, thereby improving the advance notice, accuracy, and stability of aircraft actuator fault prediction. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the aircraft actuator fault prediction technology based on future guided learning in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating multi-source time-series data processing and historical observation window construction in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the future guidance information construction process in an embodiment of the present invention; Figure 4 This is a schematic diagram of the future guided learning model structure in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the feature constraints of the historical state encoding network and the future evolution guidance network during the training phase in an embodiment of the present invention. Figure 6 This is a schematic diagram of the model inference process in the online prediction stage of this invention. Figure 7 This is a schematic diagram illustrating the multi-step fault prediction and early warning level generation in an embodiment of the present invention. Detailed Implementation

[0020] The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.

[0021] To address the problems existing in current technologies, this invention provides an aircraft actuator fault prediction technology based on future-guided learning, comprising the following steps: Acquire multi-source time-series data of the aircraft actuators at the current and historical times, and construct a historical observation window; The historical observation window is input into a pre-trained future guided learning model, which includes a historical state encoding network and a multi-step fault prediction module. The historical state coding network performs temporal coding on historical observation windows, extracts temporal dependencies, and outputs the current historical state representation. The multi-step fault prediction module outputs the fault prediction results of the aircraft actuators at different future prediction time scales based on the current historical state characterization. The future-guided learning model is obtained through future-guided learning training. During the training phase, future-guided information is constructed using fault evolution information within a future time window. Based on the future-guided information, the historical state encoding network is strengthened to learn the correlation between current weak anomalies and future fault evolution through cross-temporal feature constraints. During the application phase, the future-guided information does not participate in inference calculation.

[0022] The above technical solution will be further analyzed below with reference to the accompanying drawings: In one exemplary embodiment, refer to Figure 1 As shown, this invention provides an aircraft actuator fault prediction technology based on future-guided learning, comprising the following steps: Step 1: Multi-source data acquisition; Collect multi-source timing data during the operation of aircraft actuators, including actuator control commands, actual actuator response data, aircraft flight status parameters, and actuator status monitoring parameters.

[0023] The actuator control commands include at least one of the following: rudder surface deflection command, position control command, speed control command, torque control command, current control command, and hydraulic control command.

[0024] The actual response data of the actuator includes at least one of the following: actual displacement, actual angle, response speed, response acceleration, output torque, execution time, tracking error, and response delay.

[0025] The flight status parameters include at least one of the following: aircraft altitude, airspeed, Mach number, angle of attack, sideslip angle, pitch angle, roll angle, yaw angle, angular velocity, overload, flight phase, aerodynamic control surface status, and estimated external aerodynamic load.

[0026] The actuator condition monitoring parameters include at least one of the following: current, voltage, temperature, oil pressure, hydraulic flow, load, vibration, friction force estimate, clearance estimate, jamming degree estimate, and health status index.

[0027] Among them, actuator control commands are used to reflect the flight control system's desired control behavior towards the actuator; actuator actual response data are used to reflect the actual output of the actuator after executing the control commands; aircraft flight status parameters are used to describe the external flight environment and operating conditions of the actuator; and actuator status monitoring parameters are used to describe the actuator's own operating status.

[0028] Step 2: Data preprocessing and construction of historical observation windows; The acquired multi-source time-series data is subjected to time synchronization, outlier processing, filtering and noise reduction, operating condition normalization, and sliding window division.

[0029] Specifically, data from different sampling frequencies are resampled to a unified time axis, and short-term missing data are supplemented by interpolation; abnormal abrupt changes are identified using a preset normal range or sliding statistical range, and corrected using adjacent valid data; high-frequency noise is suppressed using low-pass filtering or moving average; the data are normalized according to the normal statistical range of each variable under the corresponding flight conditions, and finally, the historical observation window is divided according to the preset window length and window sliding step.

[0030] Let the current time be The multi-source observation vector is:

[0031] in, Indicates actuator control commands. This represents the actual response data of the actuator. Indicates aircraft flight status parameters, This indicates the actuator status monitoring parameters.

[0032] At the current moment Based on this, the construction length is Historical observation window:

[0033] in, Indicates the current time t For the historical observation window at the end of the time, L This indicates the number of time steps contained in the historical observation window. This represents the first multi-source observation vector in the historical observation window. This represents the second multi-source observation vector within the historical observation window. Indicates the current timet The multi-source observation vector.

[0034] This historical observation window is used to characterize the correspondence between actuator control commands and actual responses over a period of time prior to the current moment, as well as the changes in response error, response delay, and status monitoring parameters over time.

[0035] Step 3: Constructing the actuator response residual sequence; To enhance the physical interpretability of fault prediction, this invention further constructs an actuator response residual sequence.

[0036] Based on the actuator control commands, aircraft flight state parameters, and actuator nominal response model, the desired actuator response is obtained:

[0037] in, This represents the nominal response model of the actuator. Indicates model parameters.

[0038] In this embodiment, the nominal response model of the actuator adopts a linear state-space model based on normal state data identification, which is expressed as:

[0039] in, The nominal response model is represented in the first... t The state vector at each time step A Represents the state transition matrix. and These represent the input matrices corresponding to control commands and flight status parameters, respectively. C Indicates the output matrix. This represents the direct transfer matrix of control commands to output responses. This represents the direct transfer matrix of flight state parameters to the output response; the parameters of this matrix are identified based on the actuator's normal operating data by minimizing the mean square error between the expected response and the normal actual response.

[0040] The actuator response residual is expressed as:

[0041] in, This indicates the degree of deviation between the actual response and the expected response. This represents the first response calculated by the actuator nominal response model under the current control command and flight conditions. t The expected response at each moment.

[0042] Within a continuous time window, the actuator response residual sequence can be obtained:

[0043] in, Indicates the current time t This is the actuator response residual sequence at the end time. This represents the response residual at the start of the historical observation window. This represents the response residual at the next moment. This represents the response residual at the current moment.

[0044] The actuator response residual sequence and the historical observation window are input together into the historical state coding network to enhance the historical state coding network's ability to characterize fault precursors such as abnormal actuator control response, increased tracking error, and enhanced degradation trend.

[0045] Step 4: Constructing Future Guiding Information; The future guidance information includes at least one of the following: future residual evolution sequence, future degradation direction vector, future fault label, future degradation degree, future health status, fault occurrence advance, and remaining available time. The future residual evolution sequence consists of the response residual between the actual response and the expected response of the actuator within the future time window, and is used to describe the amplitude change, slope change, fluctuation change, cumulative deviation change and continuous abnormal change of the response residual within the future time window. The future degradation direction vector is composed of the residual mean, residual maximum, residual variance, residual change slope, residual accumulation and response delay change within the future time window, and is used to characterize the direction and intensity of the actuator's evolution from a normal state to a fault state. The future fault label is the fault type label corresponding to the first abnormal fault state that appears within a future time window. The future degradation degree is a comprehensive degradation index obtained by normalizing and weighting the components in the future degradation direction vector, and is used to quantify the degree of deviation of the actuator from the normal operating state within the future time window.

[0046] The future health status is divided into normal, mild deterioration, moderate deterioration, and severe deterioration based on the relationship between the future degree of deterioration and the preset deterioration threshold. The fault occurrence advance and remaining available time are determined based on the time interval between the current moment and the moment when the fault first occurs in the future or when the degree of degradation reaches the preset fault threshold.

[0047] During the model training phase, at the current time The length afterwards is H The time window serves as a guide for the future:

[0048] Based on the deviation between the actual and expected responses of the actuators within the future guidance window, a future residual evolution sequence is constructed:

[0049] Let the future window length be H ,but:

[0050] in, k This represents the time step index in the future residual evolution sequence. , used to represent the response residuals corresponding to different sampling times within the future window; This represents the mean of the time step indices within the future window, used to calculate the slope of the residual change. Indicates the current time t The corresponding actuator response delay Indicates the end time of the future window t+H The corresponding response latency, This represents the change in future response delay within a future window.

[0051] Based on the residual mean of the future residual evolution sequence Maximum residual residual variance Slope of residual change Cumulative residuals and future response delay changes Construct a future degradation direction vector:

[0052] The components of the future degradation direction vector are normalized, and then weighted and fused according to the contribution of each component to the fault evolution representation to obtain the future degradation degree. :

[0053] in, The vector representing the future degradation direction. j The normalized value of each component. This represents the weight of the corresponding component, satisfying... and .

[0054] The future degree of degradation is compared with a degradation threshold determined based on the statistical distribution of samples in normal operation to obtain the future health status. In this embodiment, thresholds at various levels are set based on the mean and standard deviation of the future degradation degree of normally operating samples: below The time is classified as normal, not lower than And lower than The time is classified as a mild degradation state, not lower than And below the fault determination threshold The time is classified as moderate degradation, not lower than The time is classified as a severely degraded state. Among them: The mean value representing the degree of degradation of normally functioning samples. The standard deviation represents the degree of degradation of normally functioning samples. Determined based on the distribution boundary of degradation degree in the labeled fault samples.

[0055] For the fault state label sequence within a future time window, the fault type corresponding to the first abnormal fault state is selected as the future fault label. If no abnormal fault state occurs within the future time window, the future fault label will be normal.

[0056] Let the moment when an abnormal fault label first appears within a future time window or when the degree of degradation first reaches the fault determination threshold be denoted as . Then, the lead time for failure and the remaining available time are expressed as:

[0057] in, This represents the time interval between adjacent sampling moments; when no abnormal fault state occurs within the future time window and the future degradation level does not reach the fault determination threshold, it will... Set to a time no less than the length of the future window.

[0058] The future residual evolution sequence, future degradation direction vector, future fault label, future degradation degree, future health status, fault occurrence lead time, and remaining available time together constitute the future guidance information:

[0059] Step 5: Building and training the future guided learning model; The future-guided learning model includes a historical state encoding network, a future evolution guidance network, a cross-temporal feature constraint module, a multi-step fault prediction module, and an early warning decision module.

[0060] During the model training phase, the historical observation windows corresponding to the same current time will be used. and future guidance information These form training sample pairs. The historical state encoding network outputs the current historical state representation. Future evolution guides the network to output future degradation characteristics. The multi-step fault prediction module outputs fault prediction results at different prediction time scales based on the current historical state representation.

[0061] (1) Historical state coding network: In this embodiment, the historical state encoding network includes an input mapping layer, a gated recurrent unit network, and a state mapping layer. The input mapping layer concatenates and linearly maps the control commands, actual response data, flight state parameters, state monitoring parameters, and corresponding response residuals within the historical observation window. The gated recurrent unit network processes the mapped features at each time step in chronological order and outputs the hidden state of the last time step. The state mapping layer maps the hidden state to the current historical state representation.

[0062] in, Represents a historical state coding network. This represents the actuator response residual sequence corresponding to the historical observation window.

[0063] During the training phase, the parameters of the historical state coding network are updated simultaneously by the multi-step fault prediction loss and the cross-temporal feature constraint loss; during the online prediction phase, the network parameters after training are used, and the current historical state representation is output only based on the current and historical observation data and the corresponding response residuals.

[0064] (2) Future evolution guiding network: In this embodiment, the future evolution guidance network includes a residual sequence coding layer, a discrete state embedding layer, a continuous feature mapping layer, and a fusion mapping layer.

[0065] The residual sequence coding layer performs temporal coding on the future residual evolution sequence in the future guidance information to extract the residual temporal evolution features within the future time window. The sequence coding is as follows: the response residuals at each future time step are input into the temporal coding unit in chronological order, and the hidden state at the last time step or the weighted result of the hidden states at each time step is used as the residual temporal evolution features to extract the continuous change relationship of the residuals within the future time window. The discrete state embedding layer transforms the future fault labels and future health states in the future guidance information into continuous feature vectors; The continuous feature mapping layer performs a linear mapping on the future degradation direction vector, future degradation degree, failure occurrence advance and remaining available time in the future guidance information; The fusion mapping layer concatenates the above features and maps them to a future degradation representation:

[0066] During training, the future degradation representation does not directly replace the current historical state representation. Instead, it guides the historical state encoding network to adjust its parameters through cross-temporal feature constraint loss, ensuring that the current historical state representation includes feature information related to future degradation. After training, the online deployment model only loads the parameters of the historical state encoding network, the multi-step fault prediction module, and the early warning decision module, without loading the parameters of the future evolution guidance network and the cross-temporal feature constraint module.

[0067] (3) Cross-time domain feature constraint module: The cross-temporal feature constraint module uses at least one of feature consistency constraints, degradation trend constraints, and contrastive learning constraints to make the current historical state representation in the feature space approach the corresponding future degradation representation and stay away from non-corresponding future degradation representations.

[0068] The cross-temporal feature constraint module uses a mapping function Mapping the current historical state representation to the future degradation representation space:

[0069] in, This represents the current historical state as mapped to the future degenerate feature space. This represents the current historical state after mapping. and These represent the weight matrix and bias vector of the mapping layer, respectively. This represents a non-linear activation function.

[0070] By employing feature consistency constraints, the mapped representation of the current historical state is made to approximate the corresponding future degenerate representation:

[0071] in, This represents the loss due to feature consistency constraints. This represents the current historical state after mapping. This represents the future degradation representation corresponding to the same training sample; the smaller the feature consistency constraint loss, the closer the current historical state representation and the corresponding future degradation representation are in the feature space.

[0072] A degradation trend constraint is adopted to ensure that the degradation trend corresponding to the current historical state is consistent with the future degradation direction vector:

[0073] in, Indicates a degradation trend constraint. This represents a degradation trend prediction function composed of fully connected mapping layers, used to predict the degradation direction vector based on the mapped current historical state representation. Specifically, it is expressed as:

[0074] in, This represents the mapping matrix for predicting degradation trends. This represents the corresponding bias parameters; the prediction degradation direction vector is minimized. Vector of actual future degradation direction The squared error between the predicted and actual future degradation directions ensures that the degradation direction represented by the current historical state aligns with the actual future degradation direction.

[0075] When using contrastive learning constraints, the current historical state representation is obtained by mapping the same training sample. With future degradation characteristics Positive sample pairs are formed, while negative samples are formed by future degradation representations of different fault types, different degradation directions, or different future health states within the same batch. The following contrastive learning loss is used to bring positive sample pairs closer together and push negative sample pairs further apart:

[0076] in, This represents the contrastive learning constraint. Represents cosine similarity. Indicates the temperature coefficient. This indicates the future degradation characteristics of candidates within the same batch.

[0077] Each loss weight is taken as a non-negative value and is determined based on the combined performance of fault classification error, degradation regression error and prediction lead on the validation set. First, each loss is normalized to a similar order of magnitude, and then the fault prediction loss is set as the main factor and the cross-time domain constraint loss is set as the auxiliary factor to avoid the auxiliary constraint exceeding the main prediction task.

[0078] For example, for a given actuator sample, there might be only a slight response delay within the current window, not yet reaching the warning threshold; however, in future windows, the response delay might continuously increase and develop into a response delay fault. Through future-guided learning, the current features of this sample can be guided to feature regions related to "future efficiency decline faults," enabling the model to identify similar weak anomalies in advance during online prediction. Simultaneously, through contrastive learning constraints, feature separability is maintained between different future fault evolution patterns. For instance, future jamming fault samples and future bias fault samples are separated in the feature space, thereby improving the accuracy of fault type prediction.

[0079] (4) Multi-step fault prediction module: The multi-step fault prediction module includes a shared feature mapping layer and multiple parallel prediction head branches. The shared feature mapping layer receives the current historical state representation. The system outputs shared prediction features. Each prediction head branch branches from the shared prediction features, corresponding to a future prediction time scale. Each prediction head branch includes a fault classification output layer and a degradation degree regression output layer, which output the fault occurrence probability, fault type, and degradation degree at the corresponding prediction time scale, respectively. During training, each prediction head branch calculates the prediction loss for its corresponding time scale and jointly updates the historical state encoding network parameters through the shared feature mapping layer. Therefore, the branches are independent at the output layer but influence each other through joint training at the shared feature layer.

[0080] Short-term, medium-term, and long-term prediction timescales are determined based on the actuator response time constant, sampling interval, and typical fault evolution cycle, satisfying the following conditions: ,in These represent the lead time steps for short-term, medium-term, and long-term forecasts, respectively.

[0081] The multi-step fault prediction module contains multiple prediction head branches, each corresponding to a future prediction time scale. For each prediction head branch, the multi-step fault prediction module outputs the probability of future fault occurrence and fault type at the time scale through the classification branch based on the current historical state representation, and outputs the future degradation degree at the time scale through the regression branch. For each prediction advance step, the failure prediction result includes at least the probability of future failure, the probability distribution of failure type, and the degree of future degradation; the future health status level is determined based on the comparison between the degree of future degradation and the health status threshold; the failure advance and remaining available time are determined based on the moment when the failure probability first reaches the warning threshold at each prediction time scale; and the warning level is jointly determined by the failure probability, the degree of degradation, and the prediction uncertainty.

[0082] The fault types include at least one of the following: actuator jamming fault, hysteresis fault, bias fault, drift fault, gain loss fault, response delay fault, efficiency degradation fault, intermittent fault, sensor malfunction fault, and complete failure fault.

[0083] For predicting the advance step size The output of the prediction head branch in the multi-step fault prediction module, corresponding to the prediction advance step size:

[0084] in, This indicates that the prediction features are shared across the corresponding prediction time scale. This indicates the weight parameters of the shared feature mapping layer. Indicates the bias parameters of the shared feature mapping layer; and Respectively represent the first i The weight parameters and bias parameters of the fault occurrence probability prediction layer corresponding to each prediction time scale; and Respectively represent the first i The weight parameters and bias parameters of the fault type prediction layer corresponding to each prediction time scale; and Respectively represent the first i The regression layer weights and biases correspond to the degradation level for each prediction time scale. These weights and biases are determined during model training. Initial values ​​are assigned randomly during the initial training phase, and then iteratively updated based on the prediction loss function and gradient optimization process to ultimately obtain parameter values ​​that meet the requirements of the prediction task. Indicates the first i The probability of failure occurrence over a future predicted timescale. Indicates the first i Probability distribution of fault types at a future prediction time scale Indicates the first i The degree of degradation over a future timescale is predicted.

[0085] By setting multiple prediction lead times, the model can simultaneously output short-term, medium-term, and long-term fault prediction results, enabling maintenance personnel to formulate maintenance strategies based on risk changes at different time scales.

[0086] (5) Early warning decision-making module: The early warning decision module generates early warning levels based on the probability of failure, the degree of degradation, and the prediction uncertainty at each prediction time scale. If the probability of failure at each prediction time scale is below a first probability threshold and the degree of degradation is within a normal range, a normal state is output. If the probability of failure in the medium or long term reaches the first probability threshold but the probability of failure in the short term does not reach the second probability threshold, a concern or mild warning is output. If the probability of failure in the short term reaches the second probability threshold, or the degree of degradation reaches the moderate degradation threshold, a moderate warning is output. If the probability of failure in the short term reaches the third probability threshold, the degree of degradation reaches the severe degradation threshold, and the prediction uncertainty is below a preset uncertainty threshold, a severe warning is output. The first probability threshold is less than the second probability threshold, and the second probability threshold is less than the third probability threshold. P 1. Second probability threshold P 2 and the third probability threshold P 3 are used to classify different risk levels. Their values ​​are determined based on the probability distribution of failures in historical normal operation samples and failure samples, and satisfy the following:

[0087] In this embodiment, based on the statistical results of the verification samples, the first probability threshold is set to 0.3, the second probability threshold is set to 0.6, and the third probability threshold is set to 0.8.

[0088] The early warning information is output to the aircraft health management system or maintenance terminal in the form of data records containing actuator identifier, predicted time scale, probability of failure, predicted failure type, degree of degradation and early warning level.

[0089] The total loss function for model training is expressed as:

[0090] The weight coefficients of each loss term are used to balance the importance of different optimization objectives and are adjusted according to the impact of different loss terms on the model's predictive performance. In this embodiment, the following settings are used:

[0091] The aforementioned weight values ​​are adjusted based on the prediction performance of the validation set to achieve a balance between fault classification, degradation degree prediction, and consistency between historical states and future degradation characteristics.

[0092] in: This represents the fault classification loss, used to measure the difference between the predicted fault category and the actual fault category; This represents the regression loss for the degree of degradation, used to measure the error between the predicted degree of degradation and the actual degree of degradation.

[0093]

[0094] in, This indicates the number of samples in the current training batch. C Indicates the number of fault categories. This indicates that the nth sample belongs to the nth sample. c The true label of the type of fault, This indicates that the model predicts the nth sample belongs to the nth sample. c The probability of a type of failure; This represents the advance step size for the nth sample in future predictions. The true degree of degradation, This indicates the prediction result corresponding to the model.

[0095] The uncertainty calibration loss is:

[0096] in, Indicates the first i The probability of failure occurrence over a future predicted timescale. Indicates the advance step size for future forecasts The expected response.

[0097] The historical state encoding network, future evolution guidance network, cross-temporal feature constraint module, and multi-step fault prediction module are backpropagated and iteratively updated based on the total loss function.

[0098] Verify continuous lumped loss The decrease in each training cycle is less than the preset convergence threshold. Training should be stopped at this time; among them, Used to measure the duration of a model's continuous steady state, Used to determine the degree of change in the validation set loss. and The model size is determined based on the size of the training samples, the complexity of the model, and the trend of the validation set loss. It typically satisfies the following:

[0099] In this embodiment, the maximum number of training rounds is set to 200 rounds, and the continuous convergence judgment period is... Set to 20 rounds, when the total loss of the validation set decreases less than the convergence threshold for 20 consecutive rounds. Training is stopped when the condition is met; if the stopping condition is not met after 200 rounds, the model parameters corresponding to the lowest total loss of the validation set are selected as the final model.

[0100] After training, the future evolution guidance network only serves as an auxiliary branch during the training phase. It does not receive future guidance information or participate in inference computation during the online prediction phase.

[0101] Step 6: Online prediction phase; After model training is complete, the future evolution guidance network does not participate in online inference. The online prediction phase only inputs a window of historical observations acquired in real time.

[0102] The current historical state representation is obtained through a historical state coding network:

[0103] in, This represents the historical state at the current moment. This represents the historical state encoding network after training is complete. Indicates the current time t For the historical observation window at the end of the time, Indicates the current time t This is the actuator response residual sequence at the end time.

[0104] Then, the multi-step fault prediction module outputs: historical state representation. Input the shared feature mapping layer to obtain shared predicted features:

[0105] For the i For each future prediction time scale, the prediction branches are calculated as follows:

[0106] in, They represent the first i The prediction results for the probability of failure occurrence, probability distribution of failure type, and prediction of future degradation degree at each future prediction time scale. It represents the future health status obtained based on the predicted degree of degradation.

[0107] The current historical state representation is input into the shared feature mapping layer to obtain shared predictive features. Then, input the prediction head branch corresponding to each prediction advance step size to obtain:

[0108] in, The warning level is determined by comparing the predicted degree of degradation with the health status threshold; the early warning decision module then combines the probability of failure and the prediction uncertainty to generate the corresponding early warning level. Indicates the current time t The corresponding multi-step fault prediction output set includes the fault probability, fault type, degradation degree and health status prediction results at different future prediction time scales; i Indicates the index of the prediction time scale. ,in Indicates the number of time scales for prediction; It is determined by comparing the predicted degree of degradation with the health status threshold.

[0109] Since the future evolution guidance network is only used during the training phase, it does not require real future data during the online prediction phase, thus meeting the real-time fault prediction requirements of aircraft actuators.

[0110] To further illustrate the specific application of the above method in the aircraft actuator failure prediction process, the following example is taken as an example of the actuator degradation process with gradually increasing response delay.

[0111] In one exemplary embodiment, an online prediction method for aircraft actuator failures based on future-guided learning includes the following steps: This embodiment employs a joint development approach using MATLAB / Simulink and Python / PyTorch. MATLAB / Simulink is used to build a dynamic simulation model of the aircraft actuators, simulating normal operation and progressive degradation fault states, and generating time-series data containing control commands, actual responses, and state change information. Python / PyTorch is used to implement an online fault prediction algorithm based on future-guided learning, including the construction and prediction verification of a historical state encoding network, a multi-step fault prediction module.

[0112] Step 1: Aircraft actuator data acquisition and preprocessing; In this embodiment, the aircraft actuator can be an electric servo motor, a hydraulic actuator, or an electro-hydraulic servo actuator. The actuator is used to drive the aircraft elevator, ailerons, rudder, flaps, or spoilers.

[0113] During the online prediction phase, the aircraft health management system acquires multi-source time-series data in real time from onboard data acquisition equipment, including control commands, actual responses, flight status parameters, and actuator status monitoring parameters. Control commands include target angles on control surfaces, target displacements, speed commands, or torque commands; actual responses include actual actuator angles, actual displacements, response speeds, and response accelerations; flight status parameters include altitude, airspeed, angle of attack, sideslip angle, attitude angles, angular velocity, and overload; and status monitoring parameters include current, voltage, oil pressure, temperature, vibration, and load.

[0114] Since different sensors have different sampling frequencies, this embodiment first performs time alignment on the multi-source data. For missing data, interpolation, preserving previous values, or model estimation can be used to fill in the gaps. For anomalous data, median filtering, Kalman filtering, or statistical thresholding can be used for processing. After preprocessing, a sliding window method is used to construct a historical observation window. The window length can be set according to the actuator response time, flight phase duration, and fault evolution rate.

[0115] In this embodiment, the historical observation window is set according to the dynamic characteristics of the actuator response: Historical window length: That is, input the 50 sampling points before the current time.

[0116] The historical state coding network adopts a temporal feature coding structure:

[0117] Step 2: Constructing fault precursor features based on residuals; In this embodiment, actuator response residuals are introduced to enhance the interpretability of the model. The desired response is obtained based on the control command and the actuator nominal response model:

[0118] The expected response is compared with the actual response to obtain the response residual:

[0119] When the actuator is in normal condition, the response residual usually fluctuates within a reasonable range; when the actuator experiences early degradation, the residual may show a slow increase, enhanced local fluctuations, or accumulated response delay. In this embodiment, the residual sequence is input into the model along with the original multi-source data, enabling the model to learn data-driven features and utilize physically meaningful error signals, thereby improving the reliability of fault prediction.

[0120] Step 3: Online prediction process; During the model deployment phase, the future evolution guidance network and cross-temporal feature constraint module used in the training phase are completely removed, and the system only loads the network parameters of the historical state encoding network and the multi-step fault prediction module.

[0121] During online prediction, the system inputs the historical observation window constructed in step 2 into the historical state coding network.

[0122] Subsequently, the system inputs the current historical state representation into the multi-step fault prediction module.

[0123] In this embodiment, the multi-step fault prediction module can simultaneously output fault prediction results at multiple future prediction time scales. For example, it can output the short-term, medium-term, and long-term fault probabilities. If the short-term fault probability is low but the medium-term fault probability continues to rise, it indicates that the actuator has a potential degradation trend, and a warning of concern or a mild warning can be output. If the short-term fault probability is already high and the predicted degradation level is rising rapidly, a moderate or severe warning can be output.

[0124] In this embodiment, the fault types include nine categories: normal, stuck fault, hysteresis fault, bias fault, drift fault, gain loss fault, response delay fault, efficiency degradation fault, and complete failure fault.

[0125] Step 4: Outputting prediction results and making early warning decisions; The warning level can be determined based on a comprehensive assessment of the failure probability, degradation level, remaining availability time, and prediction uncertainty. For example: when the failure probability, degradation level, and uncertainty are low, a normal warning is output; when the failure probability slightly increases or the degradation trend slightly intensifies, a warning of concern is output; when the failure probability exceeds the first threshold and the degradation level continues to rise, a mild warning is output; when the failure probability exceeds the second threshold or the remaining availability time significantly shortens, a moderate warning is output; and when the failure probability exceeds the third threshold and the model prediction uncertainty is low, a severe warning is output.

[0126] Verification results: 1. Degradation trend identification results: The response hysteresis degradation process was selected for verification.

[0127] When the actuator parameters begin to change:

[0128] The results show that the degree of degradation of the model output can gradually increase with the degradation process of the actuator.

[0129] 2. Verification of early fault prediction: Based on the fault determination threshold: =0.8, used as a criterion for severe degradation.

[0130] Simulation results:

[0131] The results show that by introducing future degradation guidance information, the model can identify the failure evolution trend earlier.

[0132] 3. Multi-step prediction results: Future forecast window:

[0133] correspond: Short-term: Mild warning; Medium-term: Moderate warning; Long-term: Severe warning trend.

[0134] Through the above simulation research and development, the following verifications were made: a simulation model of the aircraft actuator operating state was established; the method based on future degradation guidance information was verified to effectively extract the fault evolution trend; the future guidance learning model was verified to be able to predict future fault states using only historical information in the online prediction stage; compared with the method that only uses historical information, this method can detect actuator degradation trends in advance and improve the predictability of faults.

[0135] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for predicting aircraft actuator failures based on future-guided learning, characterized in that, include: Acquire multi-source time-series data of the aircraft actuators at the current and historical times, and construct a historical observation window; The historical observation window is input into a pre-trained future guided learning model, which includes a historical state encoding network and a multi-step fault prediction module. The historical state encoding network performs time-series encoding on historical observation windows, extracts temporal dependencies, and outputs a representation of the current historical state. The input to the historical state encoding network also includes an actuator response residual sequence. This sequence characterizes the deviation between the actuator's expected response and its actual response. It is derived from actuator control commands, actual actuator response data, and aircraft flight state parameters. The expected response is obtained through the actuator's nominal response model, and the difference between the expected response and the actual actuator response is calculated. The actuator's nominal response model is a response model pre-identified based on the actuator's normal operating state data, used to calculate the actuator's expected response according to the actuator control commands and aircraft flight state parameters. The multi-step fault prediction module outputs the fault prediction results of the aircraft actuators at different future prediction time scales based on the current historical state representation; the multi-step fault prediction module includes a shared feature mapping layer and multiple parallel prediction head branches. The shared feature mapping layer receives the current historical state representation output by the historical state coding network and outputs shared prediction features through nonlinear mapping. Multiple parallel prediction head branches branch from the shared prediction features, and each prediction head branch corresponds to a future prediction time scale, used to output the fault prediction result at the corresponding time scale based on the shared prediction features. The future prediction time scale includes at least two of the following: short-term prediction time scale, medium-term prediction time scale, and long-term prediction time scale. The fault prediction result includes at least one of the following for each of the multiple future prediction time scales: future fault occurrence probability, fault type, degradation degree, health status level, fault occurrence lead time, remaining available time, and warning level. The future-guided learning model is obtained through future-guided learning training. During the training phase, it uses fault evolution information within the future time window to construct future-guided information. Based on the future-guided information, it strengthens the historical state encoding network's learning of the correlation between current weak anomalies and future fault evolution through cross-temporal feature constraints. During the application phase, the future-guided information does not participate in inference calculation.

2. The aircraft actuator fault prediction method based on future-guided learning according to claim 1, characterized in that: The construction of the historical observation window includes: Collect multi-source timing data during the operation of aircraft actuators. The multi-source timing data includes actuator control commands, actual actuator response data, flight status parameters, and actuator status monitoring parameters. The system performs time synchronization, outlier processing, filtering and noise reduction, operating condition normalization, feature construction, and sliding window division on multi-source time series data to obtain the historical observation window corresponding to the current moment.

3. The aircraft actuator fault prediction method based on future guided learning according to claim 2, characterized in that: Each prediction head branch includes a fault classification output layer and a degradation degree regression output layer; The fault classification output layer is used to output the probability of future fault occurrence and fault type at the corresponding future prediction time scale; The degradation degree regression output layer is used to output the future degradation degree at the corresponding future prediction time scale.

4. A method for training a future-guided learning model for an aircraft actuator, wherein the future-guided learning model is used to execute the aircraft actuator fault prediction method according to any one of claims 1-3, characterized in that, include: Obtain the training dataset, which contains historical observation windows corresponding to multiple sample times and corresponding future guidance information. The future guidance information is constructed based on the fault evolution information within a future time window of a preset length after the current time. An initial model is constructed, which includes a historical state encoding network and a multi-step fault prediction module connected in sequence, as well as a future evolution guidance network set up in parallel with the historical state encoding network during the training phase, a cross-temporal feature constraint module connecting the historical state encoding network and the future evolution guidance network, and an early warning decision module for generating early warning levels. Input the historical observation window at the same moment into the historical state coding network to obtain the current historical state representation; input the future guidance information corresponding to that moment into the future evolution guidance network to obtain the future degradation representation; The cross-temporal feature constraint module applies feature space constraints to the current historical state representation and the future degradation representation, enabling the current historical state representation to learn future degradation information; at the same time, the current historical state representation is input into the multi-step fault prediction module for fault prediction. Based on the constraint loss generated by feature space constraints and the prediction loss generated by fault prediction, the model parameters are jointly optimized and iteratively updated until the preset training termination condition is met, resulting in a well-trained future guided learning model. The future evolution guide network in the well-trained future guided learning model is removed or deactivated. The early warning decision module generates at least one early warning level from among normal, attention, mild, moderate and severe warnings based on the results of multi-step fault prediction and prediction uncertainty.

5. The training method for a future-guided learning model of an aircraft actuator according to claim 4, characterized in that: The future guidance information includes at least one of the following: future residual evolution sequence, future degradation direction vector, future fault label, future degradation degree, future health status, fault occurrence advance, and remaining available time. The future residual evolution sequence consists of the response residual between the actual response and the expected response of the actuator within the future time window. It is used to describe the amplitude change, slope change, fluctuation change, cumulative deviation change and continuous abnormal change of the response residual within the future time window. The future degradation direction vector is composed of the residual mean, residual maximum, residual variance, residual change slope, residual accumulation, and response delay change within the future time window. It is used to characterize the direction and extent of the actuator's evolution from a normal state to a fault state by the positive and negative change directions of each degradation feature and their normalized amplitude. Future fault labels are the fault type labels corresponding to the first abnormal fault state that appears within a future time window; The future degradation degree is a scalar obtained by normalizing and weighting the feature components in the future degradation direction vector, and is used to represent the comprehensive degree of the actuator state deviating from the normal state within the future time window; Future health status is classified into normal, mild, moderate, and severe based on the relationship between the future degree of deterioration and a preset deterioration threshold. The lead time for failure and the remaining available time are determined based on the time interval between the current moment and the moment when the failure first occurs in the future or when the degree of degradation reaches the preset failure threshold.

6. The training method for a future-guided learning model of an aircraft actuator according to claim 5, characterized in that: The future evolution guidance network includes a residual sequence coding layer, a discrete state embedding layer, a continuous feature mapping layer, and a fusion mapping layer; The future evolution guidance network is used to encode and fuse future guidance information, specifically including: The residual sequence coding layer is used to sequence encode the future residual evolution sequence in the future guidance information and extract the residual temporal evolution features within the future time window; The discrete state embedding layer is used to convert the future fault labels and future health states in the future guidance information into continuous feature vectors, respectively. The continuous feature mapping layer is used to linearly map the future degradation direction vector, future degradation degree, failure occurrence advance, and remaining available time in the future guidance information; The fusion mapping layer is used to concatenate the residual temporal evolution features, the output of the discrete state embedding layer and the output of the continuous feature mapping layer, and then map the concatenated features to a feature space of the same dimension as the current historical state representation, outputting the future degradation representation; During the online prediction phase, the future evolution guidance network does not receive input data and does not participate in fault prediction inference calculations.

7. The training method for a future-guided learning model of an aircraft actuator according to claim 6, characterized in that: The cross-temporal feature constraint module implements feature space constraints through at least one of feature consistency constraints, degradation trend constraints, and contrastive learning constraints. Feature consistency constraints make the mapped current historical state representation approximate the corresponding future degenerate representation; The degradation trend constraint ensures that the degradation trend corresponding to the current historical state is consistent with the future degradation direction vector; Contrastive learning constraints pair the current historical state representation and future degradation representation of the same training sample with positive samples, and pair samples corresponding to different fault types, different degradation trends, or different health status levels with negative samples, thereby improving the feature separability between different future degradation patterns.

8. The training method for a future-guided learning model of an aircraft actuator according to claim 7, characterized in that: The loss function used in the joint optimization includes at least one of fault classification loss, degradation degree regression loss, feature consistency constraint loss, degradation trend constraint loss, contrastive learning loss, and uncertainty calibration loss. Each loss function is calculated based on the corresponding prediction result, the true label, and the difference between the current historical state representation and the future degradation representation.

Images