Remaining life prediction method of full-motion analog machine control load loop amplifier

Abstract

The invention belongs to the technical field of flight simulation equipment maintenance, particularly relates to a residual life prediction method for a control load loop amplifier of an all-motion simulator, and aims to solve the problem of low AMP life prediction precision. The method comprises the following steps: acquiring multi-source heterogeneous data including pilot manipulation behaviors, an electric thermal state of an amplifier body and a current training scene; the unmeasurable junction temperature in the power device is synchronously estimated by processing multi-source heterogeneous data, multiple degradation characteristics are decoupled and evaluated, and a health state index is generated; and predicting the residual service life of the AMP based on the evolution trajectory of the health state index. According to the method, the real service condition of the AMP can be more comprehensively represented, and the accuracy, adaptability and engineering practicability of AMP residual life prediction are remarkably improved.

Description

A method for predicting the remaining lifetime of a load loop amplifier in a fully-motion simulator. Technical Field

[0001] This application belongs to the field of flight simulation equipment maintenance technology, and specifically relates to a method for predicting the remaining life of a control load loop amplifier in a full-motion simulator. Background Technology

[0002] Full Flight Simulator (FFS) is the core equipment for pilot training. Its Control Loading System (CLS) accurately simulates the control force of components such as the control stick and pedals of an aircraft through torque motors and servo valves, and is a key subsystem to ensure the realism of training. Among them, the control load loop amplifier (AMP), which drives the torque motor, is the power core and is subjected to complex electrical, thermal, and mechanical stress cycles over a long period of time. It is a vulnerable component in the system, and its health status directly affects the availability of the simulator and training safety.

[0003] Currently, maintenance of AMPs (Active Flight Simulators) on flight simulators mainly relies on periodic inspections or post-failure repairs. Periodic inspections, based on fixed time cycles, cannot account for the actual degree of wear and tear on the equipment, potentially leading to over-maintenance or under-maintenance, resulting in low cost-effectiveness. Post-failure repairs, on the other hand, cause training interruptions, resulting in significant downtime losses and potentially posing certain safety risks.

[0004] To address these issues, predictive maintenance technology has been introduced into the maintenance of critical equipment. In the field of power device lifetime prediction, existing technologies primarily employ physical model-based or data-driven methods. Physical model methods calculate fatigue damage using parameters such as junction temperature cycling. However, junction temperature, a critical parameter determining the lifetime of power devices, cannot be directly measured due to packaging limitations. Relying on indirect measurements such as case temperature introduces significant lag and error, leading to inaccurate remaining lifetime assessments. While data-driven methods can learn degradation patterns from historical data, they often operate like a "black box," failing to decouple complex and interconnected physical failure mechanisms such as bond wire fatigue and solder layer cracking, resulting in insensitivity to early, subtle degradation information. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, namely the low accuracy of AMP lifetime prediction, this application proposes a method for predicting the remaining lifetime of a loop amplifier for a fully-motion simulator's controlled load, comprising:

[0006] Acquire multi-source heterogeneous data related to the operating status of the control load loop amplifier (AMP). This multi-source heterogeneous data includes at least pilot control behavior data, AMP operating data, and simulated training scenario data. Process the multi-source heterogeneous data using a dual-path degradation feature extraction model combined with a junction temperature inversion network constrained by a physical model to estimate the internal junction temperature of the power device and decouple and evaluate various degradation features, generating a health status index. Based on the evolution trajectory of the health status index, use a dynamic degradation model to predict the remaining lifetime of the AMP. The degradation rate of the dynamic degradation model is adaptively adjusted according to real-time junction temperature fluctuations and operating load.

[0007] As a preferred implementation, multiple degradation characteristics are decoupled and evaluated, including: analyzing the high-frequency transient components of voltage or current signals in AMP operating data through a first processing path to obtain a bond wire health feature vector, wherein the bond wire health feature vector is used to characterize degradation characteristics related to bond wire fatigue; analyzing the time evolution characteristics of the thermal impedance spectrum derived from AMP operating data through a second processing path to obtain a solder layer health feature vector, wherein the solder layer health feature vector is used to characterize degradation characteristics related to solder layer fatigue; and using an attention mechanism to dynamically weight and fuse the bond wire health feature vector and the solder layer health feature vector to generate a health status index.

[0008] As a preferred implementation, estimating the internal junction temperature of a power device includes: constructing a physical information neural network with measured case temperature and dissipated power as inputs and residuals of the heat transfer physical equations as constraints; and retrieving the internal junction temperature based on the physical information neural network.

[0009] As a preferred implementation, estimating the internal junction temperature of a power device further includes: predicting the change in thermal resistance parameter in the heat transfer physical equation based on the degradation characteristics of the decoupling assessment; and dynamically correcting the heat transfer physical equation using the change in thermal resistance parameter.

[0010] As a preferred implementation, the dynamic degradation model is a nonlinear Wiener process model. The adaptive adjustment process of the degradation rate includes: calculating a dynamic acceleration factor based on the fluctuation amplitude of the internal junction temperature and the conduction duration of the power device; accumulating the dynamic acceleration factor into an equivalent damage time, and using the equivalent damage time as a time variable of the nonlinear Wiener process model to adaptively adjust the degradation rate.

[0011] In a preferred embodiment, the method further includes: performing multiple remaining lifetime predictions on the same multi-source heterogeneous data to obtain multiple remaining lifetime prediction values; sorting the multiple remaining lifetime prediction values, using a preset low quantile as the lower bound and a preset high quantile as the upper bound to generate a confidence interval.

[0012] As a preferred implementation, the method further includes: inputting the remaining life prediction results into a multi-constraint optimization model, and generating a maintenance suggestion window containing specific execution times by solving an integer programming problem that minimizes the overall maintenance cost, wherein the multi-constraint optimization model incorporates the operating constraints of the simulator.

[0013] As a preferred implementation, the objective function of the integer programming problem is the comprehensive maintenance cost function, which is composed of a weighted average of replacement cost, downtime loss cost, and potential failure risk cost.

[0014] As a preferred implementation, the operational constraints include safety risk constraints and task priority constraints. The safety risk constraints are determined by the maximum acceptable failure probability during future high-intensity training tasks, and the task priority constraints are used to restrict maintenance activities from avoiding preset high-priority training periods.

[0015] As a preferred implementation, the method further includes: continuously detecting whether the currently acquired multi-source heterogeneous data deviates from the distribution range established based on the training samples; when a deviation from the distribution range is detected, triggering an alarm to indicate that the reliability of the current prediction result has decreased.

[0016] Compared with the prior art, the technical solution provided in this application has at least one of the following beneficial effects: by acquiring multi-source heterogeneous data such as pilot control behavior data, AMP operation data, and simulated training scenario data, it can more comprehensively characterize the real service conditions of the full-motion simulator AMP; through the dual-path degradation feature extraction model, degradation features can be decoupled for different failure modes, improving the early degradation identification capability; through the junction temperature inversion network constrained by the physical model, the junction temperature inside the power device that cannot be directly measured can be estimated, improving the accuracy of obtaining key life state quantities; and based on the evolution trajectory of health status indicators, a dynamic degradation model in which the degradation rate can be adaptively adjusted with real-time junction temperature fluctuations and operating load is used to predict the remaining lifetime, thereby effectively solving the problems of unmeasurable internal state, difficulty in decoupling multiple failure mechanisms, and inaccurate prediction under varying operating conditions in the prior art, significantly improving the accuracy, adaptability, and engineering practicality of AMP remaining lifetime prediction. Attached Figure Description

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 is a flowchart of a method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to an embodiment of this application; Figure 2 is a system block diagram of a system for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to an embodiment of this application; Figure 3 is a schematic diagram of the structure of a computer system for implementing the methods, systems, and electronic devices of this application. Detailed Implementation

[0018] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] This application provides a method for predicting the remaining lifetime of a control load loop amplifier (AMP) in a full-motion simulator. It acquires multi-source heterogeneous data related to the AMP's operating status, including at least pilot control behavior data, AMP operating data, and simulated training scenario data. A dual-path degradation feature extraction model, combined with a junction temperature inversion network constrained by a physical model, processes the multi-source heterogeneous data to estimate the internal junction temperature of the power device and decouple and evaluate various degradation features, generating a health status index. Based on the evolution trajectory of the health status index, a dynamic degradation model is used to predict the AMP's remaining lifetime. The degradation rate of the dynamic degradation model is adaptively adjusted according to real-time junction temperature fluctuations and operating load. This application can more comprehensively characterize the actual service conditions of a full-motion simulator AMP, significantly improving the accuracy, adaptability, and engineering practicality of AMP remaining lifetime prediction.

[0021] To more clearly illustrate the remaining lifetime prediction method for the fully-motion simulator control load loop amplifier of this application, the steps in the embodiments of this application will be described in detail below with reference to Figure 1.

[0022] The first embodiment of this application provides a method for predicting the remaining lifetime of a full-motion simulator control load loop amplifier, including steps S10-S30, each step described in detail below: Step S10, acquiring multi-source heterogeneous data related to the operating status of the control load loop amplifier AMP, wherein the multi-source heterogeneous data includes at least pilot control behavior data, AMP operating data and simulation training scenario data.

[0023] Optionally, pilot control behavior data may include, but is not limited to, joystick displacement, control frequency, control direction change frequency, control duration, etc.; AMP operation data may include, but is not limited to, current, voltage, temperature, vibration, power, conduction time, etc.; simulated training scenario data may include, but is not limited to, training scenario labels such as normal flight, go-around, engine failure, turbulence, etc.

[0024] In one embodiment of this application, pilot control behavior data includes joystick displacement. Displacement velocity Displacement acceleration Frequency of manipulation Number of times the joystick crosses zero within the window / window duration. These parameters reflect the impact of the pilot's control style on the AMP load; different pilots (e.g., student / instructor, aggressive / gentle) will result in completely different current surge spectra.

[0025] In one embodiment of this application, the AMP operating data includes three-phase current. DC bus voltage Power tube case temperature Radiator temperature Vibration signals These parameters reflect the electrical and thermal stress states of the AMP itself.

[0026] In one embodiment of this application, the simulated training scenario data is read directly from the simulator host to obtain the current training scenario label. and scene duration The load spectrum of the AMP varies greatly under different scenarios. For example, the "engine failure go-around" scenario requires a huge amount of stick force, and the AMP is at full load peak state.

[0027] As one possible implementation, sensors are deployed at key locations in the AMP to collect multi-source heterogeneous data.

[0028] Specifically, a high-frequency Hall current sensor is installed at the three-phase output terminal to collect current signals; an isolated voltage probe is installed at both ends of the power device to collect voltage signals; a temperature sensor is installed on the surface of the power module to collect case temperature signals; a vibration sensor is installed on the printed circuit board to collect vibration signals; joystick displacement signals and control behavior-related signals are read from the existing sensing system of the simulator; and the current training scenario label is read from the simulator host or training management system.

[0029] In one embodiment of this application, the bandwidth of the high-frequency Hall current sensor is... Range Isolated voltage probe bandwidth Range The temperature sensor is an NTC thermistor with high accuracy. Response time The vibration sensor is a microelectromechanical system (MEMS) accelerometer with a measurement range of [missing information]. Sampling rate .

[0030] The data from all the aforementioned channels are preferably acquired synchronously to ensure timestamp alignment and facilitate subsequent multi-source fusion. For example, all signals can be acquired synchronously at a sampling rate of 50kHz.

[0031] Furthermore, the collected raw data can be segmented and preprocessed. For example, a fixed time window and sliding step size can be used to segment the data into frames, and each frame of data can be preprocessed.

[0032] As an example, a sample frame is constructed with a window length of 10 seconds and a sliding step of 5 seconds.

[0033] As an example, the preprocessing of each frame of data can be: (1) Outlier removal: sensor spike noise is removed using the 3σ principle; (2) Missing value interpolation: linear interpolation is used to fill in the temporary data loss; (3) Trend removal: DC component is removed from the vibration signal; (4) Normalization: each physical quantity is normalized separately.

[0034] In this embodiment of the application, the normalization method is z-score normalization.

[0035] The multi-source heterogeneous data related to the AMP's operating status obtained in this application embodiment includes at least pilot control behavior data, AMP operating data, and simulated training scenario data. This overcomes the limitations of traditional methods that rely solely on single information sources such as the equipment's own voltage, current, and temperature for life assessment. Since the stress and thermal loads on the full-motion simulator AMP depend not only on its own operating status but also on the pilot's control methods and training scenarios, by introducing multi-dimensional "human-machine-environment" information, the AMP's operating status can be comprehensively described from three levels: load source, equipment response, and external mission environment. This provides a more comprehensive characterization of the AMP's actual service conditions, making the life prediction results closer to the actual degradation process and significantly improving the accuracy of remaining life prediction.

[0036] Step S20: Using a dual-path degradation feature extraction model, combined with a junction temperature inversion network constrained by a physical model, multi-source heterogeneous data is processed to estimate the internal junction temperature of the power device and decouple and evaluate various degradation features, generating a health status index.

[0037] Optionally, the multi-source heterogeneous data obtained in step S10 can be input into the dual-path degradation feature extraction model and the junction temperature inversion network constrained by the physical model to estimate the internal junction temperature of the power device and decouple and evaluate various degradation features to generate a health status index.

[0038] Among them, the health status index can be a time-series scalar representing the health level of AMP, or it can be a low-dimensional feature vector. The closer its value is to the healthy end, the better the equipment status is; the closer it is to the failure end, the more serious the equipment degradation is.

[0039] In one embodiment of this application, a physical information neural network is constructed to achieve bidirectional coupling between junction temperature inversion and model self-calibration. First, a physical information neural network is constructed with measured shell temperature and dissipated power as inputs and residuals of the heat transfer physical equations as constraints. The internal junction temperature is then inverted based on the physical information neural network.

[0040] The physical equations of heat transfer can be represented using a thermal network model, such as the Foster thermal network model or other equivalent thermal resistance-capacity network models. During model training, a joint loss function is constructed that includes a data loss term and a physical loss term, where: the data loss term is used to constrain the deviation between the model output and the calibration samples; and the physical loss term is used to constrain the model to minimize the residuals of the heat transfer equations.

[0041] By introducing physical constraints, we can prevent the network from generating estimation results that do not conform to thermal mechanisms by relying solely on data fitting, thereby improving the reliability of internal junction temperature estimation.

[0042] In one embodiment of this application, the Foster thermal network model equations are used as physical constraints for training the neural network. The input to the network is the measured shell temperature. and power dissipation The output is the virtual junction temperature. .

[0043] The physical loss term is defined as the residual of the thermal network equation: ,in, Indicates thermal resistance. The heat capacity can be initially determined from the device datasheet.

[0044] The data-driven loss term is the mean square error of the infrared thermal imager calibration data: ,in, This represents the virtual junction temperature of the i-th data point; The value represents the junction temperature reference value corresponding to the i-th data point; N represents the number of data points, which represents the total number of measurement samples used in the calibration process.

[0045] It should be noted that the junction temperature reference value is obtained during the offline calibration phase by directly measuring the power device under open or specially packaged conditions using specialized measurement equipment such as infrared thermal imagers. This value is used to train the physical information neural network. After the model has been trained and deployed, it is no longer necessary to rely on this direct measurement method; the internal junction temperature can be derived solely from the measured case temperature and dissipated power.

[0046] The total loss function is a weighted sum of the two: ,in, This represents the balance coefficient, used to control the strength of physical constraints. As an example, It can take the value 0.1.

[0047] In one embodiment of this application, during the offline training phase, 12 months of operational data from six simulators are collected to train the physical information neural network. This data includes normal operation data, natural aging data, and accelerated aging test data. The normal operation data consists of the operational data from one normally functioning AMP; the natural aging data consists of three AMPs running until failure, recording their complete degradation trajectories; and the accelerated aging test data consists of two AMPs undergoing power cycling tests on a test bench until failure. The dataset is divided into a 70% training set, a 15% test set, and a 15% validation set. The network structure is a 3-layer fully connected network (256, 128, 64 hidden units) with the ReLU activation function. The loss function is: The Adam optimizer was used with a learning rate of 0.001. Training was conducted for 200 epochs, and the program was stopped if the validation set loss did not decrease for 10 consecutive epochs.

[0048] It should be noted that bond wire fatigue is one of the most common failure modes in power modules, typically causing bond wire breakage or detachment, leading to a drop in on-state voltage. The voltage or current signal rises and generates a spike pulse during switching transients. Therefore, in this embodiment of the application, the high-frequency transient components of the voltage or current signal in the AMP operating data are analyzed through a first processing path to obtain the bond wire health feature vector, wherein the output bond wire health feature vector is used to characterize the degradation characteristics related to bond wire fatigue.

[0049] As one possible implementation, a one-dimensional convolutional structure, an Inception structure, or a one-dimensional convolutional structure with dilated convolution can be used to extract high-frequency local patterns.

[0050] As an example, the network structure uses a 1D Inception module, containing three parallel branches: Branch 1: Convolutional kernel size 1×2, dilation rate 4, used to capture the distortion features of microsecond-level switching spikes; Branch 2: Convolutional kernel size 1×5, dilation rate 2, used to capture the rising trend of sub-millisecond-level on-state voltage drop; Branch 3: Max pooling, extracting the overall energy change trend, with high-frequency current as input. (Sampling rate ≥ 50kHz) or collector-emitter voltage Output the health feature vector of the bond line. ,in, This represents the dimension of the bonding line health feature vector.

[0051] In one embodiment of this application, a high-frequency current sequence is used. Input the 1D Inception module, where, Multi-scale temporal features are extracted through three parallel branches: A dilated convolution with a 1×2 kernel and a dilation rate of 4, equipped with 64 filters, yields a feature map of size L×64; a dilated convolution with a 1×5 kernel and a dilation rate of 2, also equipped with 64 filters, yields a feature map of size L×64; a max pooling operation with a 1×2 window and a stride of 2 outputs a feature map of size L / 2×64, which is then upsampled back to L×64. The outputs of the three branches (each of size L×64) are concatenated along the feature channel dimension to obtain a fused feature map of size L×192. Finally, global average pooling aggregates the feature maps into a 192-dimensional bond line health feature vector. .

[0052] It should be noted that solder layer fatigue typically causes thermal grease aging or solder layer cracking, leading to increased thermal resistance, deterioration of the heat conduction path, and changes in the thermal time constant. Therefore, this embodiment of the application uses a second processing path to analyze the time evolution characteristics of the thermal impedance spectrum derived from AMP running data to obtain a solder layer health feature vector, wherein the output solder layer health feature vector is used to characterize the degradation characteristics related to solder layer fatigue.

[0053] As one possible implementation method, a combination of frequency domain transformation and cyclic network can be used. For example, the thermal impedance spectrum can be calculated first, and then its evolution mode over time can be extracted through gated cyclic unit (GRU).

[0054] As an example, the network structure employs a thermal impedance analysis module with gated loop units to calculate the thermal impedance spectrum characteristics, with the virtual junction temperature as the input. Shell temperature Power dissipation The formula for calculating the characteristics of the thermal impedance spectrum is: Where F represents the Fourier transform. The thermal impedance spectrum variation pattern over time is extracted using a GRU network, and the output is a solder layer health feature vector. ,in, This represents the dimension of the solder layer health feature vector.

[0055] In one embodiment of this application, the virtual junction temperature sequence is... Shell temperature sequence Dissipated power sequence Input thermal impedance analysis module: Divides each frame of data into segments, for example, each segment could be 1 second long, calculates the Fourier transform of each segment, calculates the thermal impedance spectrum characteristics, and extracts the amplitude and phase at three frequency points: 1Hz, 10Hz, and 100Hz, to form the... The features are obtained by inputting 10 consecutive feature segments within 10 seconds into a GRU layer (128 hidden units), and taking the output at the last moment as the solder layer health feature vector. .

[0056] In this embodiment, the bond line health feature vector is processed through two independent fully connected layers. and solder layer health feature vector Mapped to the same dimension.

[0057] As an example, dimension alignment can be performed using the following formula: ;in, express Activation function Represents the weight matrix. This represents the bias vector.

[0058] ;in, Represents the weight matrix. This represents the bias vector.

[0059] Understandably, the weight matrix and bias vector automatically optimized by the fully connected layer are used to optimize the bond line health feature vector. and solder layer health feature vector Perform linear transformations on each, mapping them to the same dimension, i.e., 128 dimensions, and then combine them. The activation function introduces nonlinear activation so that the two can be integrated in a common space.

[0060] Furthermore, an attention mechanism is employed to dynamically weight and fuse the bond line health feature vector and the solder layer health feature vector to generate a health status index.

[0061] An attention mechanism is used to dynamically fuse dual-path features, enabling the network to automatically adjust the attention weights of different failure modes according to the current operating conditions.

[0062] As an example, the formula for dynamic weighted fusion through the attention mechanism can be: ;in, This represents the fused feature vector and attention weights. Attention weight vector , This represents the sigmoid function. This indicates element-wise multiplication. and These are learnable parameters in the attention mechanism, used to learn how to generate an effective weight distribution from the concatenated features. Represents the weight matrix. This represents the bias vector.

[0063] In one embodiment of this application, the bond line health feature vector is... and solder layer health feature vector The concatenated input is then fed into the attention network: .

[0064] Since the dominant failure modes of AMP may differ under different training conditions, the attention mechanism can automatically adjust the weights of different feature components according to the current input conditions, so that the final health status index retains the sensitivity to different failure modes and has a better comprehensive representation ability.

[0065] Furthermore, the feature vectors will be fused. Input to a fully connected layer, output unit 1, activate the sigmoid function, and obtain the health status index: ,in, This represents the sigmoid activation function. This represents the weight matrix of the fully connected layer, with a dimension of 1×128. This represents the bias term of the fully connected layer.

[0066] The high-dimensional fused feature vector is transformed into an intuitive scalar as a health status indicator by using an activation function, quantifying the health level of AMP. The output value ranges from [0,1]. Indicates complete health. This indicates complete failure. In practical applications, health status indicators... It will decrease over time, reflecting the aging or degradation of the equipment.

[0067] Furthermore, since thermal resistance parameters will age due to solder layer aging, this embodiment of the application performs reverse calibration on the virtual junction temperature and the obtained health status indicators through thermal resistance parameter correction.

[0068] Specifically, based on the degradation characteristics of the decoupling assessment, the change in thermal resistance parameter in the heat transfer physical equation is predicted; and the change in thermal resistance parameter is used to dynamically correct the heat transfer physical equation.

[0069] As degradation characteristics such as solder layer fatigue and thermal interface aging intensify, the thermal resistance of the device changes over service time. By constructing a thermal resistance parameter correction subnetwork, the change in thermal resistance parameter can be predicted based on the current degradation characteristics, and the updated thermal resistance parameter can be fed back into the junction temperature inversion process to recalculate the internal junction temperature and update the physical loss term.

[0070] As an example, the change in thermal resistance can be predicted using the following formula: ;in, This represents the change in thermal resistance. and These are the learnable weights and biases of a neural network. Indicates weight, represents the bias, and ReLU represents the ReLU activation function.

[0071] The observable fusion feature vector was established using the above formula. To the change in thermal resistance that cannot be directly observed The mapping relationship between them.

[0072] Furthermore, the updated thermal resistance parameters are fed back into the heat transfer physics equations used for junction temperature inversion to recalculate the internal junction temperature and update the physical loss terms.

[0073] As an example, the updated thermal resistance parameter can be determined using the following formula: ;in, This indicates the updated thermal resistance parameter. This represents the initial thermal resistance.

[0074] In one embodiment of this application, the measured shell temperature is... and power dissipation Input the trained physical information neural network to obtain the initial junction temperature. , will fuse feature vectors The input is fed into a thermal resistance parameter correction subnetwork (2 fully connected layers, 64 and 32 hidden units, ReLU activation function at output layer) to obtain the change in thermal resistance. This updates the thermal resistance parameters. Furthermore, using the updated thermal resistance, the heat network equations are resolved to obtain a re-estimated junction temperature. ;in, This indicates a re-estimation of the junction temperature. .

[0075] Repeat the process of thermal resistance correction and junction temperature reestimation a preset number of times, such as 3 times, and output the updated internal junction temperature.

[0076] The physical model parameters are self-calibrated online through a reverse calibration process, enabling the junction temperature inversion results to better follow the changes in the actual aging state.

[0077] By employing a junction temperature inversion network constrained by a physical model, the junction temperature of the power devices inside an AMP is estimated. Compared to existing technologies that typically only use externally measurable parameters such as case temperature to approximate the thermal state of the device, this application introduces the physical heat transfer laws into the neural network modeling process. This allows the internal junction temperature, a key variable determining the thermal fatigue life of the device, to be obtained with high reliability through inversion, thus solving the problem of life assessment distortion caused by the inability to directly measure the internal state of power devices. Since the junction temperature more accurately reflects the level of thermal stress accumulation than the case temperature, the health assessment and life prediction results constructed based on this internal state are more reliable.

[0078] Furthermore, a dual-pathway degradation feature extraction model explicitly correlated with physical failure mechanisms is used to specifically extract and decouple the features corresponding to different failure modes during AMP degradation. Since typical AMP degradation modes include at least bond wire fatigue and solder layer fatigue, and different failure modes exhibit different evolutionary patterns in electrical and thermal characterization, this application, through mechanism-correlated dual-pathway processing, can effectively retain the feature sensitivity of different degradation modes, reduce the problem of mutual masking of failure features, thereby improving the ability to identify early, weak degradation signs and enhancing the engineering interpretability of the prediction process.

[0079] Step S30: Based on the evolution trajectory of health status indicators, a dynamic degradation model is used to predict the remaining service life of the AMP. The degradation rate of the dynamic degradation model is adaptively adjusted according to real-time junction temperature fluctuations and operating load.

[0080] Optionally, the evolution trajectory can be a continuous sequence of changes in health status indicators over time or with the accumulation of equivalent damage. The degradation rate of the dynamic degradation model is adaptively adjusted according to real-time junction temperature fluctuations and operating load, so that the model can adapt to load changes caused by different training scenarios and different manipulation behaviors.

[0081] As an example, the obtained health status indicators are arranged in chronological order to form an evolutionary trajectory of the health status indicators. In other embodiments, the evolutionary trajectory can also be constructed based on equivalent damage accumulation.

[0082] As one possible implementation, the dynamic degradation model is a nonlinear Wiener process model, and the adaptive adjustment process of the degradation rate includes: calculating a dynamic acceleration factor based on the fluctuation amplitude of the internal junction temperature and the conduction duration of the power device.

[0083] As an example, the dynamic acceleration factor is calculated based on the ratio of the current junction temperature fluctuation amplitude to the reference junction temperature fluctuation and the ratio of the current conduction time to the reference conduction time. This dynamic acceleration factor is used to reflect the acceleration or deceleration effect of the current operating condition on the degradation process relative to the reference operating condition.

[0084] Specifically, the dynamic acceleration factor can be determined using the following formula: ;in, This indicates the current junction temperature fluctuation range. Indicates the reference junction temperature fluctuation. Indicates the current conduction time. This represents the reference conduction time, where m and n are material constants.

[0085] As one possible implementation method, the current junction temperature fluctuation range The maximum value can be subtracted from the minimum value within each switching cycle, and the typical value under rated operating conditions can be used as a reference for junction temperature fluctuation. On-time Reference conduction time A time of 0.1 seconds is acceptable; for power modules, this is typically the case. , .

[0086] The dynamic acceleration factor is accumulated into an equivalent damage time, and the equivalent damage time is used as a time variable in the nonlinear Wiener process model to adaptively adjust the degradation rate.

[0087] The dynamic degradation model is a nonlinear Wiener process model that uses the cumulative damage value obtained by integrating the dynamic acceleration factor over time as the time scale for its degradation trajectory. Instead of directly using natural time as the uniform time scale for the degradation model, the influence of operating conditions is converted into equivalent damage time, thus allowing the same length of actual running time to have different degradation contributions under different load conditions.

[0088] As an example, a nonlinear Wiener process can be constructed to describe the degradation trajectory of health status indicators using the following formula: ;in, The denotes the health status index at time t; 'a' represents the drift coefficient, reflecting the rate of degradation; and 'b' represents the nonlinear coefficient. This indicates accelerated degradation. Indicates deceleration and degradation; This represents cumulative damage, converting the actual load spectrum into damage time under equivalent reference conditions. Represents the diffusion coefficient, characterizing the random fluctuations of degradation; This represents standard Brownian motion.

[0089] As one possible implementation, an extended Kalman filter is used to estimate the model parameters a and b online. Specifically, the model parameters at time k are treated as a state vector. The observation equation is: ,in, This represents the health status index at time k. This represents the cumulative damage at time k. This represents observation noise, which follows a normal distribution. , This indicates the time interval between frames, such as 5 seconds.

[0090] Furthermore, Bayesian inference can be used to update the posterior distributions of model parameters a and b in real time.

[0091] Specifically, after obtaining the observation values ​​at time k Then, the parameter distribution can be updated using the following formula: Among them, the prior distribution It is the understanding of parameter distribution based on historical data before time k-1, the likelihood function. It can characterize the health status index at the newly acquired time k. The probability of occurrence under the current parameter assumptions, posterior distribution It is the updated understanding of the parameter distribution after combining prior knowledge and new evidence.

[0092] With the operation of equipment and data , ... With continuous accumulation, the model's estimation of AMP parameters a and b will become more and more accurate, enabling personalized degradation trajectory tracking and improving prediction accuracy.

[0093] A nonlinear Wiener process is employed to model the evolution trajectory of health status indicators (AMPs) to characterize the drift trend and random fluctuations during AMP degradation. By estimating or updating the model parameters online, the remaining useful life distribution corresponding to the failure threshold can be obtained, thus outputting the remaining useful life prediction result.

[0094] It should be noted that the output of the dynamic degradation model is a probability distribution, which in this embodiment is the probability distribution function of the remaining useful life. The result obtained from the probability distribution can be used as the predicted remaining useful life, or a point estimate can be further determined as the predicted remaining useful life. The point estimate can be any one of the mean, median, mode, or preset quantile.

[0095] In one embodiment of this application, the probability distribution function of the remaining useful life is: ;in, The probability distribution function representing the remaining useful life. Indicates the failure threshold. Indicates the remaining service life. , which represents the cumulative damage up to the current moment.

[0096] Remaining service life From the current health status index Initially, the price drops to the failure threshold for the first time. The time required.

[0097] As an example, failure threshold It is 0.6.

[0098] In some embodiments, when outputting the prediction results, a confidence interval for the remaining lifetime may also be output.

[0099] As one possible implementation, multiple remaining lifetime predictions are performed on the same multi-source heterogeneous data to obtain multiple remaining lifetime prediction values; the multiple remaining lifetime prediction values ​​are sorted, and a confidence interval is generated with a preset low quantile as the lower bound and a preset high quantile as the upper bound.

[0100] The uncertainty of the prediction results is quantified by using confidence intervals.

[0101] As an example, the Monte Carlo Dropout technique is used, maintaining the Dropout layer active during the inference phase with a dropout rate of 0.1, and performing N forward propagations (typically N=100) to obtain a sample set of remaining lifetime predictions. The sample set is sorted in ascending order, and the 5th percentile is selected as the lower bound and the 95th percentile as the upper bound to obtain the confidence interval. .

[0102] Furthermore, the probability of failure can be determined based on the duration of future tasks: ;in, Indicates the probability of failure. Indicates the duration of a future mission. Indicates the indicator function, i.e., if , ,otherwise .

[0103] Since deep learning models may make confident but incorrect predictions on samples outside the training data distribution, this application can also continuously detect whether the currently acquired multi-source heterogeneous data deviates from the distribution range established based on the training samples; when a deviation from the distribution range is detected, an alarm is triggered to indicate that the reliability of the current prediction result has decreased.

[0104] Specifically, Mahalanobis distance is used to detect in real time whether the fused feature vector deviates from the training set distribution: ;in, This represents the Mahalanobis distance between the fused feature vector and the training set. The mean vector representing the features of the training set. The covariance matrix represents the features of the training set.

[0105] like This indicates that the currently acquired multi-source heterogeneous data deviates from the distribution range established based on the training samples, and the current sample is an out-of-distribution sample, triggering an alarm. Among these, This represents the preset threshold, and p represents the feature dimension. This represents the 95th percentile of the chi-square distribution.

[0106] As an example, feature dimension 95th percentile of chi-square distribution ,like If so, an alarm will be triggered.

[0107] In one embodiment of this application, when an alarm is triggered, the message "Warning: The current operating condition is outside the model training range, and the prediction results may be unreliable" is displayed. In some embodiments, alternative maintenance suggestions based on conservative strategies may also be provided, such as suggesting preventative replacement based on traditional maintenance cycles.

[0108] Furthermore, for each new failure sample, the model is fine-tuned: the bottom feature extraction layer is frozen, only the top layer and attention layer are fine-tuned, the learning rate is reduced to 0.0001, training is performed for 5-10 epochs, and the mean vector of the training set is updated. Covariance Matrix .

[0109] Furthermore, simulator training scenarios are diverse, ranging from smooth cruise to severe engine failure handling, and the load spectrum experienced by AMPs varies greatly. This makes static or simple life models unable to adapt to dynamically changing operating conditions, and prediction accuracy is difficult to guarantee. Moreover, existing technologies typically only output a remaining useful life (RUL) value, which is disconnected from actual operational decisions and fails to comprehensively consider practical constraints such as simulator training scheduling, spare parts inventory, and maintenance personnel arrangements. As a result, maintenance personnel find it difficult to formulate specific and executable maintenance plans based on the prediction results.

[0110] The embodiments of this application can also propose adaptive maintenance recommendations based on the remaining useful life prediction results.

[0111] As one possible implementation, the remaining life prediction results are input into a multi-constraint optimization model. By solving an integer programming problem that minimizes the overall maintenance cost, a maintenance suggestion window containing specific execution times is generated. The multi-constraint optimization model incorporates the simulator's operational constraints.

[0112] As an example, the objective function of an integer programming problem is a comprehensive maintenance cost function, which is a weighted average of replacement costs, downtime loss costs, and potential failure risk costs.

[0113] By establishing a mixed-integer programming model and combining the prediction results with operational constraints, the optimal replacement time window is solved: ;in, This represents minimizing the overall maintenance cost; decision variables. , indicating whether to switch items in the t-th time window (t=1,...,T), where T is the number of idle time slots within the planning period; This indicates the cost of a single replacement part, which can include spare parts costs and labor costs. This represents the cost of downtime per unit of time. This represents the duration of the t-th idle period; The cost of failure risk can be a monetized measure of factors such as training interruption and security risks. This indicates the probability of failure before the next task.

[0114] In this embodiment of the application, the operational constraints include safety risk constraints and task priority constraints. The safety risk constraints are determined by the maximum acceptable failure probability during future high-intensity training tasks, and the task priority constraints are used to restrict maintenance activities from avoiding preset high-priority training periods.

[0115] As an example, the constraint could be: Time window constraint: The replacement time must fall within the simulator's idle training period. ,in, This represents the k-th idle period without a training task.

[0116] Safety risk constraints: The probability of failure during future high-intensity training missions must not exceed the acceptable risk level threshold. ,in, Indicates the duration of future high-intensity training tasks. This represents the acceptable risk level threshold, typically set to a value of [value missing]. .

[0117] Spare parts constraint: The number of replacement parts must not exceed the existing inventory. , where S represents the current AMP spare parts inventory.

[0118] Task priority constraint: No component replacement will be scheduled during high-priority training tasks. .

[0119] Solve the above mixed integer programming model using the branch and bound method or dynamic programming, and output the optimal replacement time. .

[0120] As an example, read the work schedule for the next 7 days and identify the free periods. Assume the free periods are: Period 1: March 5th, 02:00-06:00 (4 hours), Period 2: March 6th, 23:00-03:00 (4 hours), Period 3: March 7th, 01:00-05:00 (4 hours). Construct an integer programming model: Decision variables: Assuming replacement cost is 5000 yuan, downtime loss is 2000 yuan / hour, failure risk cost is 10000 yuan, and there are 2 spare parts in stock, then the objective function is: ,constraint: Safety constraints: If ,but The optimal solution is obtained by solving the problem. Therefore, it is recommended to replace the part between 02:00 and 06:00 on March 5th.

[0121] In one embodiment of this application, if the fused feature vector does not deviate from the distribution range established based on the training samples, an executable maintenance work order containing specific suggested replacement windows, reasons, required spare parts, and other information is output according to the solution result of integer programming.

[0122] As an example, the output executable maintenance work order is as follows: Predicted remaining lifespan: 125 hours (90% confidence interval: [98,152] hours), probability of failure within the next 7 days: 32%, suggested replacement window: 2024-03-05 02:00-06:00, reason for replacement: the probability of failure is expected to reach 15% before the next high-intensity go-around training, exceeding the safety threshold of 10%, required spare part: left seat AMP module (model: AMP-XXXX, stock: 2 pieces). The maintenance work order has been automatically created, please confirm.

[0123] Please refer to Figure 2. A remaining lifetime prediction system for a full-motion simulator control load loop amplifier according to a second embodiment of this application is used to perform the above-described method for predicting the remaining lifetime of a full-motion simulator control load loop amplifier. The system includes a data acquisition module 100, a health status index generation module 200, and a remaining lifetime prediction module 300.

[0124] The data acquisition module 100 is used to acquire multi-source heterogeneous data related to the operating status of the control load loop amplifier (AMP). The multi-source heterogeneous data includes at least pilot control behavior data, AMP operating data, and simulated training scenario data. The health status index generation module 200 is used to process the multi-source heterogeneous data through a dual-path degradation feature extraction model combined with a junction temperature inversion network constrained by a physical model to estimate the internal junction temperature of the power device and decouple and evaluate various degradation features to generate a health status index. The remaining lifetime prediction module 300 is used to predict the remaining lifetime of the AMP based on the evolution trajectory of the health status index using a dynamic degradation model. The degradation rate of the dynamic degradation model is adaptively adjusted according to real-time junction temperature fluctuations and operating load.

[0125] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the system described above can be found in the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0126] It should be noted that the remaining lifetime prediction method and system for a fully-motion simulator control load loop amplifier provided in the above embodiments are only illustrative examples of the above functional module division. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of this application can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of this application are only for distinguishing each module or step and are not considered as an improper limitation of this application.

[0127] A device according to a third embodiment of this application includes: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to implement the above-described method for predicting the remaining lifetime of a full-motion simulator control load loop amplifier.

[0128] A computer-readable storage medium according to a fourth embodiment of this application stores computer instructions for execution by a computer to implement the above-described method for predicting the remaining lifetime of a fully dynamic simulator control load loop amplifier.

[0129] A computer program product according to the fifth embodiment of this application, when run on an electronic device, causes the electronic device to execute the above-described method for predicting the remaining lifetime of a fully dynamic simulator-controlled load loop amplifier.

[0130] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes and related descriptions of the electronic devices, computer-readable storage media, and computer program products described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0131] Referring now to Figure 3, a schematic diagram of a computer system for implementing the systems, methods, and electronic devices of this application is shown. The server shown in Figure 3 is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this application.

[0132] As shown in Figure 3, the computer system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in Read Only Memory (ROM) 302 or programs loaded from storage section 308 into Random Access Memory (RAM) 303. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0133] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a cathode ray tube (CRT), liquid crystal display (LCD), and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0134] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0135] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0136] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0137] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0138] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0139] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

Claims

1. A method for predicting the remaining lifetime of a load loop amplifier in a fully-motion simulator, characterized in that, include: Acquire multi-source heterogeneous data related to the operating status of the control load loop amplifier (AMP), wherein the multi-source heterogeneous data includes at least pilot control behavior data, AMP operating data, and simulated training scenario data; process the multi-source heterogeneous data through a dual-path degradation feature extraction model combined with a junction temperature inversion network constrained by a physical model to estimate the internal junction temperature of the power device and decouple and evaluate various degradation features to generate a health status index; based on the evolution trajectory of the health status index, use a dynamic degradation model to predict the remaining lifetime of the AMP, wherein the degradation rate of the dynamic degradation model is adaptively adjusted according to real-time junction temperature fluctuations and operating load.

2. The method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to claim 1, characterized in that, The decoupling assessment of multiple degradation characteristics includes: analyzing the high-frequency transient components of voltage or current signals in the AMP operating data through a first processing path to obtain a bond wire health feature vector, wherein the bond wire health feature vector is used to characterize degradation characteristics related to bond wire fatigue; analyzing the time evolution characteristics of the thermal impedance spectrum derived from the AMP operating data through a second processing path to obtain a solder layer health feature vector, wherein the solder layer health feature vector is used to characterize degradation characteristics related to solder layer fatigue; and using an attention mechanism to dynamically weight and fuse the bond wire health feature vector and the solder layer health feature vector to generate a health status index.

3. The method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to claim 1, characterized in that, The estimation of the internal junction temperature of the power device includes: constructing a physical information neural network with the measured case temperature and dissipated power as inputs and the residuals of the heat transfer physical equations as constraints; and retrieving the internal junction temperature based on the physical information neural network.

4. The method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to claim 3, characterized in that, The estimation of the internal junction temperature of the power device further includes: predicting the change in the thermal resistance parameter in the heat transfer physical equation based on the degradation characteristics of the decoupling assessment; and dynamically correcting the heat transfer physical equation using the change in the thermal resistance parameter.

5. The method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to claim 1, characterized in that, The dynamic degradation model is a nonlinear Wiener process model. The adaptive adjustment process of the degradation rate includes: calculating a dynamic acceleration factor based on the fluctuation amplitude of the internal junction temperature and the conduction duration of the power device; accumulating the dynamic acceleration factor into an equivalent damage time, and using the equivalent damage time as a time variable of the nonlinear Wiener process model to adaptively adjust the degradation rate.

6. The method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to claim 1, characterized in that, The method further includes: performing multiple remaining lifetime predictions on the same multi-source heterogeneous data to obtain multiple remaining lifetime prediction values; sorting the multiple remaining lifetime prediction values, using a preset low quantile as the lower bound and a preset high quantile as the upper bound to generate a confidence interval.

7. The method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to claim 1, characterized in that, The method further includes: inputting the remaining life prediction results into a multi-constraint optimization model, and generating a maintenance suggestion window containing specific execution times by solving an integer programming problem that minimizes the overall maintenance cost, wherein the multi-constraint optimization model incorporates the simulator's operational constraints.

8. The method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to claim 7, characterized in that, The objective function of the integer programming problem is the comprehensive maintenance cost function, which is composed of a weighted average of replacement cost, downtime loss cost, and potential failure risk cost.

9. The method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to claim 7, characterized in that, The operational constraints include safety risk constraints and task priority constraints. The safety risk constraints are determined by the maximum acceptable failure probability during future high-intensity training tasks, and the task priority constraints are used to restrict maintenance activities from avoiding preset high-priority training periods.

10. The method for predicting the remaining lifetime of a fully-motion simulator control load loop amplifier according to claim 1, characterized in that, The method further includes: continuously detecting whether the currently acquired multi-source heterogeneous data deviates from the distribution range established based on training samples; when a deviation from the distribution range is detected, an alarm is triggered to indicate that the reliability of the current prediction result has decreased.

Images