A device residual life prediction method, device, medium and equipment

Abstract

The application discloses a kind of device residual life prediction method, device, medium and equipment, belong to life prediction field.Firstly, the real-time degradation data of target device in running process is collected by pre-set sensor;Then the real-time degradation data is input into the preset physical information neural network, the local time sequence mutation feature and long-term time sequence dependence feature in the real-time degradation data are extracted by the built-in double-channel encoder of the physical information neural network and are fused, and the residual service life prediction result of the target device is output based on the processing of the embedded physical constraint loss function to the fused feature.The application effectively solves the problem that the residual life of the device cannot be accurately and efficiently predicted by the prior art.

Description

Technical Field

[0001] This invention relates to the field of lifetime prediction, and more particularly to a method, apparatus, medium, and device for predicting the remaining lifetime of a device. Background Technology

[0002] Silicon carbide metal-oxide-semiconductor field-effect transistors (SiC MOSFETs), as a typical representative of wide-bandgap semiconductor devices, have been increasingly widely used in power electronic systems such as new energy power generation, electric vehicles, and smart grids due to their advantages such as high voltage resistance, high temperature resistance, and high switching frequency. Accurately predicting the remaining lifespan of SiC MOSFET devices during actual operation is of great significance for ensuring the safe and reliable operation of power electronic systems.

[0003] Currently, academia and industry primarily employ two main technical approaches for device lifetime prediction. One approach is model-driven, based on physical failure mechanisms. This involves establishing analytical equations describing the propagation of fatigue damage in materials (such as the Coffin-Manson equation and Paris crack law) and combining them with finite element analysis to calculate the stress-strain distribution within the device, thereby estimating the wear-out lifetime. However, this method suffers from extremely high computational complexity, and the material parameters within the model are fixed as constants, failing to adapt to the dynamic changes in physical parameters during long-term device aging. This results in prediction accuracy heavily reliant on the accuracy of the task profile. The other approach is purely data-driven deep learning methods, employing models such as Long Short-Term Memory (LSTM) networks and convolutional neural networks to establish a direct mapping between collected degradation parameters and remaining lifetime. While this method avoids complex physical modeling, it is essentially a "black box" model lacking transparency, severely lacking adherence to the objective physical laws of devices. Especially under harsh conditions in real-world engineering, such as sparse samples and a large amount of truncated data, traditional deep neural networks are prone to overfitting and outputting absurd predictions that violate the principle of thermodynamic irreversibility, such as unreasonable rebounds or oscillations in the predicted lifetime curve.

[0004] More importantly, the aging failure mechanism of SiC MOSFET devices is extremely complex and multidimensional. Compared to traditional silicon-based devices, SiC MOSFETs have a high density of intrinsic defects at the microscopic material level and introduce novel high-temperature interconnect processes at the packaging level. This causes their key aging characteristics (such as on-resistance and threshold voltage) to no longer follow the traditional linear or exponential monotonic decay pattern throughout their entire lifespan, but instead exhibit a non-monotonic evolution trajectory of "decline-recovery-rebound". These shortcomings prevent current technologies from accurately and efficiently predicting the remaining lifetime of the devices. Summary of the Invention

[0005] This invention provides a method, apparatus, medium, and device for predicting the remaining lifetime of a device, in order to solve the problem that the existing technology cannot accurately and efficiently predict the remaining lifetime of a device.

[0006] In a first aspect, this application provides a method for predicting the remaining lifetime of a device, comprising: Real-time degradation data of the target device during operation is collected through preset sensors; The real-time degradation data is input into a preset physical information neural network, so that the physical information neural network extracts local temporal mutation features and long-term temporal dependency features from the real-time degradation data through a built-in dual-channel encoder and fuses them. Then, the fused features are processed based on the embedded physical constraint loss function to output the remaining service life prediction result of the target device. The physical information neural network is trained based on historical degradation data and the physical constraint loss function, which incorporates a physical penalty term to constrain the non-monotonic evolution characteristics of the degradation process.

[0007] This application acquires real-time degradation data through sensors and utilizes a pre-defined physical information neural network for feature extraction and lifetime prediction. This network incorporates a dual-channel encoder to simultaneously capture local mutations and long-term dependent features. Furthermore, its training process embeds a physical penalty term constraining the non-monotonic evolution of the degradation process. This technique deeply integrates data-driven learning capabilities with the physical laws of device degradation, resulting in predictions that not only possess high nonlinear fitting accuracy but also strictly adhere to the principle of thermodynamic irreversibility. This fundamentally eliminates the "lifetime bounce" or non-physical oscillation phenomena that easily occur in purely data-driven models, significantly improving prediction reliability and physical interpretability under sparse sample and complex operating conditions. This application effectively solves the problem that existing technologies cannot accurately and efficiently predict the remaining lifetime of devices.

[0008] Furthermore, the step of extracting and fusing local temporal abrupt change features and long-term temporal dependency features from the real-time degradation data using a built-in dual-channel encoder specifically involves: The real-time degradation data is input into the spatial mutation channel and long-term dependency channel of the dual-channel encoder, respectively. Local temporal features are extracted by sliding filtering along the time dimension through the multi-scale one-dimensional convolutional layer in the spatial mutation channel, and the local temporal features are adaptively weighted by the channel attention mechanism to obtain the local mutation feature vector. The long-term time-dependent feature vector is obtained by performing forward and backward temporal evolution calculations on the real-time degraded data through the bidirectional long short-term memory network in the long-term dependent channel. The local mutation feature vector and the long-term temporal dependency feature vector are concatenated and fused to obtain the fused feature vector.

[0009] This application achieves decoupling and collaborative extraction of features at different scales in degradation signals by setting up spatial mutation channels and long-term dependency channels in parallel. Multi-scale one-dimensional convolutional layers can keenly capture local mutations caused by microscopic defects in the early stages of device failure, while bidirectional long short-term memory networks effectively mine long-term evolution patterns throughout the entire lifecycle. The introduction of a channel attention mechanism further enhances the model's ability to select key sensitive features. Compared to traditional serial single-channel networks, the dual-channel parallel fusion structure of this application can perform dual aggregation of spatiotemporal cross-features in a deeper latent space, significantly improving the model's response sensitivity to inflection points and mutation segments in the degradation trajectory.

[0010] Furthermore, the adaptive weighting of the local temporal features through the channel attention mechanism specifically involves: The local temporal features are subjected to global average pooling and global max pooling respectively to obtain the first channel descriptor and the second channel descriptor; The first channel descriptor and the second channel descriptor are input into a shared multilayer perceptron, and a channel weight matrix is ​​generated by nonlinear transformation through a dimension reduction weight matrix and a dimension increase weight matrix. The local temporal features are weighted channel by channel according to the channel weight matrix to obtain a weighted feature sequence. The weighted feature sequence is then weighted a second time in terms of spatial dimension through a spatial attention mechanism to obtain the local mutation feature vector.

[0011] This application further cascades a spatial attention mechanism on top of the channel attention mechanism, forming a dual attention enhancement structure. Channel attention is used to distinguish the importance of different feature channels, while spatial attention is used to locate the most critical degradation segments in the temporal dimension. The two work together to enable the network to adaptively focus on the core temporal intervals and feature channels that lead to rapid lifetime decay, much like an expert troubleshooting a problem, while effectively suppressing redundant background noise. This allows for the extraction of highly representative degradation features even from experimental data with extremely low signal-to-noise ratios.

[0012] Furthermore, the fused features are processed based on the embedded physical constraint loss function to output the remaining lifetime prediction result of the target device, specifically as follows: The fused features are input into the decoder built into the physical information neural network, so that the decoder generates a first latent variable increment characterizing the threshold voltage drift and a second latent variable increment characterizing the encapsulation degradation effect. The first latent variable increment and the second latent variable increment are nonlinearly mapped by the built-in soft linear unit activation function to constrain the numerical signs of the first latent variable increment and the second latent variable increment during the update process to conform to the preset monotonically increasing or decreasing physical direction. Obtain the preset first physical reconstruction coefficient and second physical reconstruction coefficient, and obtain the observation value of the real-time degraded data at the last sampling time as the anchor point; The first contribution component is obtained by multiplying the first latent variable increment after constraint by the first physical reconstruction coefficient. The second contribution component is obtained by multiplying the second latent variable increment after constraint by the second physical reconstruction coefficient. The first contribution component and the second contribution component are summed together, and the summation result is superimposed on the anchor point as a reference to reconstruct the macroscopic degradation index. The remaining useful life prediction result of the target device is output based on the macroscopic degradation index.

[0013] This application achieves the embedding of physical knowledge into the device degradation process and the interpretable reconstruction of macroscopic indicators. The decoder does not directly output macroscopic degradation values; instead, it generates two latent variable increments with clear physical meanings, representing threshold voltage drift and package degradation. These increments are then constrained by soft linear units to ensure the rationality and sign consistency of the microscopic physical processes. Subsequently, preset physical reconstruction coefficients are used to map the two latent variable increments back to the macroscopic on-state voltage drop index, and the latest measured value is used as an anchor point for summation. This decoupling-constraint-reconstruction approach avoids the difficulty of directly learning exponentially growing values ​​in neural networks, significantly improving the numerical stability and physical realism of long-term extrapolation predictions.

[0014] Furthermore, the physical information neural network is trained based on historical degradation data and the physical constraint loss function, specifically as follows: Obtain historical degradation data; The historical degradation data is input into the initial physical information neural network, the encoder in the initial physical information neural network extracts the state vector of the historical degradation data, and the state vector is passed to the long short-term memory network in the decoder; The initial physical information neural network is trained based on a pre-defined scheduling-based teacher-forced strategy; During the training process, the value of the physical constraint loss function is calculated based on the predicted output of the initial physical information neural network and the corresponding real label. The network parameters of the initial physical information neural network are updated through the backpropagation algorithm based on the value of the physical constraint loss function until the error meets the preset requirements and the training stops, thus obtaining the trained physical information neural network. The physical constraint loss function includes a data fidelity term, a non-monotonic evolution penalty term, a step size constraint term, a hedging constraint term, and a late-stage rise constraint term. The data fidelity term measures the deviation between the predicted value and the true value of the physical information neural network. The non-monotonic evolution penalty term positively penalizes the prediction results that violate the preset physical symbol constraints based on the temporal change direction of the intermediate latent variables in the predicted degradation process. The step size constraint term constrains the local slope of the predicted degradation trajectory to suppress non-physical oscillations. The hedging constraint term penalizes the prediction results when the changing trends of multiple latent variables cancel each other out. The late-stage rise constraint term ensures that the remaining service life prediction can enter the preset failure state in the later stage.

[0015] This application constructs a multi-physical constraint composite loss function with five sub-losses during the training phase, overcoming the limitations of traditional single mean squared error loss which only pursues data fitting. The non-monotonic evolution penalty term directly constrains the direction of change of latent variables to conform to physical sign rules; the step size constraint term forces the predicted trajectory to be locally smooth and oscillating; the hedging constraint term prevents multiple physical mechanisms from canceling each other out and deceiving data fidelity; and the late-stage rise constraint term ensures that the model can correctly enter the failure zone at the end of its remaining lifetime. Simultaneously, the scheduling-based teacher-driven strategy gradually attenuates the guiding signal through cosine annealing, enabling the model to smoothly transition from supervised training to free autoregressive inference, significantly enhancing the model's generalization performance in practical deployments.

[0016] Furthermore, after outputting the predicted remaining lifetime of the target device, the method further includes: The degradation prediction trajectory corresponding to the remaining useful life prediction result is compared with the preset failure physical threshold boundary; When the predicted value on the degradation prediction trajectory first reaches or exceeds the failure physical threshold boundary, the target device is determined to have reached a failure state. The difference between the predicted time when the failure state is triggered and the current time is determined as the final remaining service life.

[0017] This application clarifies a quantitative output mechanism for remaining service life. By comparing the predicted degradation trajectory with a preset failure threshold, and using the first attainment or exceeding of the threshold as the failure criterion, the remaining number of cycles or remaining time from the current moment to the failure moment of the device can be objectively and accurately quantified. This judgment rule conforms to the failure definition based on the rate of change or absolute value threshold of key parameters in industry standards, providing a clear decision-making basis for condition-based maintenance and early warning of power electronic systems.

[0018] Secondly, this application provides a device for predicting the remaining lifetime of a device. The device for predicting the remaining lifetime of a device includes: The acquisition module is used to collect real-time degradation data of the target device during operation through preset sensors; The prediction module is used to input the real-time degradation data into a preset physical information neural network, so that the physical information neural network extracts local temporal mutation features and long-term temporal dependency features from the real-time degradation data through a built-in dual-channel encoder and fuses them. Then, based on the embedded physical constraint loss function, the fused features are processed to output the remaining service life prediction result of the target device. The physical information neural network is trained based on historical degradation data and the physical constraint loss function, which incorporates a physical penalty term to constrain the non-monotonic evolution characteristics of the degradation process.

[0019] The device remaining lifetime prediction apparatus provided in this application acquires real-time degradation data of the target device during operation using preset sensors through an acquisition module. This data is then processed by a prediction module that calls a physical information neural network with embedded physical constraints. Because this network incorporates a dual-channel encoder, it can extract and deeply fuse local temporal abrupt changes and long-term temporal dependencies from the degradation data in parallel, thereby simultaneously capturing sudden fluctuations caused by microscopic defects in the early stages of failure and macroscopic decay patterns throughout the entire lifespan. Furthermore, the network incorporates a physical constraint loss function with a non-monotonic evolution penalty term during the training phase. This ensures that the prediction module, when using the network for forward inference, forces the output remaining lifetime prediction results to strictly adhere to the irreversible physical laws of device degradation, effectively avoiding "lifetime bounce" and non-physical oscillations that easily occur in pure data-driven models under sparse sample conditions or noise interference. Based on this, the device can achieve high-confidence lifetime prediction without relying on complex finite element simulations or precise analytical equations. Therefore, it can significantly reduce computational resource consumption while providing physically interpretable, numerically stable, and engineering-adaptable technical support for condition-based maintenance and early warning decisions in power electronic systems.

[0020] Thirdly, this application provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform a device remaining lifetime prediction method as described above. Its beneficial effects are the same as those of the device remaining lifetime prediction method provided in the first aspect of this application.

[0021] Fourthly, this application provides a terminal device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement any of the device remaining lifetime prediction methods described in the first aspect. Attached Figure Description

[0022] Figure 1 : A schematic flowchart of an embodiment of the device remaining lifetime prediction method provided in this application; Figure 2 : A schematic diagram of an embodiment of the power cycle test principle provided in this application; Figure 3 : A schematic diagram of an embodiment of the target device provided in this application, showing the variation of on-state voltage drop with power cycle number under junction temperature fluctuation of 130°C; Figure 4 : A schematic diagram of an embodiment of the sequence-to-sequence prediction model of the physical information neural network provided in this application; Figure 5 : A schematic diagram of an embodiment of the multi-scale convolution and attention module in the spatial mutation channel provided in this application; Figure 6 : A schematic diagram of an embodiment of a single-cell structure of a long short-term memory network provided in this application; Figure 7 : A schematic diagram of an embodiment of the overall unfolded structure of the bidirectional long short-term memory network provided in this application; Figure 8 : A schematic diagram of an embodiment comparing the predicted remaining service life of the target device provided in this application under a junction temperature fluctuation of 130°C; Figure 9 : A schematic diagram of an embodiment of the device remaining lifetime prediction device provided in this application. Detailed Implementation

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

[0024] Example 1 Please refer to Figure 1 In order to solve the problem that existing technologies cannot accurately and efficiently predict the remaining lifetime of devices, this invention provides a method for predicting the remaining lifetime of devices, including steps S01-S02.

[0025] S01: Collect real-time degradation data of the target device during operation through preset sensors; In a preferred embodiment of this invention, the step of collecting real-time degradation data of the target device during operation using a preset sensor specifically includes: In this embodiment, the real-time degradation data of the target device during operation is first collected using a preset sensor. The target device can specifically be a silicon carbide metal-oxide-semiconductor field-effect transistor (SiC MOSFET), such as the Wolfspeed C3M0021120K model, which uses a TO 247-4 package, has a rated voltage of 1200V, and a rated current of 100A. To realistically simulate the physical degradation process of the device under actual operating conditions, the SiC MOSFET is mounted on a dedicated semiconductor device power cycling test platform. This test platform applies a stress load with a specific junction temperature fluctuation amplitude to the device to reproduce fatigue damage caused by the mismatch of the thermal expansion coefficients of the packaging materials. In this embodiment, the power cycling conditions are set as follows: minimum junction temperature 35°C, maximum junction temperature 165°C, and junction temperature fluctuation amplitude 130°C. Figure 2 As shown, Figure 2 The schematic diagram of the power cycle test used in this embodiment is shown, wherein Figure 2 (a) is a graph showing the changes in the loop variable. Figure 2 (b) is a schematic diagram of the test circuit.

[0026] During the power cycling aging test, the system uses a high-frequency sampling circuit to collect degradation precursor parameters that gradually change with the aging process in real time. In this embodiment, the on-state voltage drop is preferably used as the raw degradation data. The on-state voltage drop is a global macroscopic indicator that can comprehensively reflect the internal physical changes of the device: on the one hand, it can map the threshold voltage drift caused by the accumulation of trap charge in the gate oxide layer; on the other hand, it can also intuitively reflect the change in on-resistance caused by mechanical damage to the metal contact surface. For this purpose, a preset voltage sensor is connected in parallel across the main electrode of the device. This voltage sensor continuously collects the on-state voltage drop value of the device in each switching cycle with a millisecond-level time resolution, forming a raw degradation sequence that evolves over time.

[0027] To ensure the accuracy of subsequent model training and the effectiveness of feature representation, the collected raw on-state voltage drop data is preprocessed. First, outliers in the original time series signal are identified and removed based on a three-standard-deviation criterion, and linear interpolation is used to fill in the missing data, thereby accurately restoring the signal basis. Then, a quadratic exponential smoothing algorithm is used to filter the cleaned sequence to effectively suppress high-frequency random noise and improve the signal-to-noise ratio of key features. Finally, while maintaining data fidelity, downsampling is performed on the filtered sequence to significantly reduce the sequence length by stripping away strongly coupled temporal redundancy, thus reducing the computational load on the subsequent neural network model. The real-time degraded data obtained after the above preprocessing is a one-dimensional time series vector, where each data point represents the on-state voltage drop value of the device at the corresponding sampling time. Figure 3 The figure shows a typical curve of the on-state voltage drop of the target device in this embodiment as a function of the number of power cycles after preprocessing, under a junction temperature fluctuation of 130°C.

[0028] Through the above methods, this embodiment can objectively and continuously acquire real-time degradation data of the target device during operation, providing a high-quality data foundation for the subsequent input of the physical information neural network.

[0029] S02: Input the real-time degradation data into a preset physical information neural network, so that the physical information neural network extracts local temporal mutation features and long-term temporal dependency features from the real-time degradation data through the built-in dual-channel encoder and fuses them. Then, based on the embedded physical constraint loss function, the fused features are processed to output the remaining service life prediction result of the target device. The physical information neural network is trained based on historical degradation data and the physical constraint loss function, which incorporates a physical penalty term to constrain the non-monotonic evolution characteristics of the degradation process.

[0030] In a preferred embodiment of this invention, the physical information neural network is trained based on historical degradation data and the physical constraint loss function, specifically as follows: In this embodiment, the physical information neural network needs to be trained in advance using historical degradation data and a composite loss function with embedded physical penalty terms. The purpose of this training process is to ensure that the network, during optimization, not only pursues the accuracy of fitting historical data, but is also forced to be guided to a solution space that conforms to the physical laws of device degradation.

[0031] First, the on-state voltage drop sequence of the target device in historical power cycling aging experiments is obtained as historical degradation data. Then, the preprocessed historical degradation data is standardized using Z-score zero-mean to eliminate the influence of different physical dimensions on gradient updates. To enable the network to effectively handle data with time-dependent characteristics, a sliding window technique is used to reconstruct the dataset from the standardized data. Specifically, the encoder's time window length is set to 100, the decoder's time window length to 20, and the sliding step size to 1. Through this technique, the original one-dimensional time series is converted into a two-dimensional tensor matrix containing rich historical state information, where each sample has a time window length of 100 and a feature dimension of 1. Next, according to a preset ratio (e.g., 70% training set, 30% validation set), the constructed tensor set is unbiasedly partitioned along the time dimension.

[0032] Then, an initial physical information neural network is constructed, consisting of a dual-channel encoder and a decoder. During training, historical degradation data from the training set is input into the initial network. The encoder extracts the state vector of the input sequence and passes it to the long short-term memory network in the decoder. During training, the context vector output by the encoder does not participate in the forward pass of the decoder; it is only combined with the output features of the decoder at the output. The input to the decoder is the anchor value (i.e., the observation at the end of the input sequence), which also participates in the computation of the dual-channel encoder. Furthermore, the state vectors of the input sequence extracted by the encoder that are unique to the long short-term memory network are separately passed to the long short-term memory network in the decoder to extract dependencies in the original long sequence.

[0033] To enhance the model's generalization performance in real-world deployments, this embodiment employs a scheduled teacher-forced training strategy: the teacher-forced ratio is gradually reduced from 1.0 to 0.0 using cosine annealing, thus smoothly transitioning between the data distribution differences between the training and free-running inference phases. This embodiment uses the AdamW optimizer as the parameter update method, which can effectively and quickly find the model's global minimum while also avoiding the training results from getting trapped in local optima to some extent.

[0034] In each iteration, the composite loss function is calculated based on the predicted output of the neural network and the corresponding true label according to the initial physical information, and the network parameters are updated based on this value through the backpropagation algorithm. The composite loss function used in this embodiment is not the traditional single mean squared error loss, but embeds multiple physical penalty terms, specifically including a data fidelity term, a non-monotonic evolution penalty term, a step size constraint term, a hedging constraint term, and a late-rise constraint term.

[0035] Specifically, the physical information neural network uses a composite loss function L. total Its expression is: Among them, L data For the data fidelity term (mean squared error), L Vth and L Vpack L represents the non-monotonic evolution penalty terms for constraining threshold voltage drift and package degradation monotonicity, respectively. s To constrain the step size term of the local slope of the predicted curve, L c To prevent the two hidden variables from canceling each other out, L up To ensure the success of later-stage escalation constraints that prevent failure in later stages, the specific calculation methods for each term are as follows: Where V pre V true The predicted and actual values ​​are the macroscopic outputs of the model, where D1 and D2 are the first-order and second-order difference operators, respectively. , For the model's macroscopic output of predicted and actual increments, λ V1 , λ V2 , λ V3 , λ V4 For the corresponding constraint coefficients, mean(·) represents the calculation of the average value.

[0036] The aforementioned composite loss function is iterated continuously during the training process until its value converges below a preset threshold or reaches the maximum number of iterations. At this point, the network parameters are saved, resulting in a trained physical information neural network. In subsequent online predictions, even when faced with sparse and noisy real-time degradation data, this network can output remaining lifespan predictions that strictly adhere to physical laws and are numerically stable.

[0037] In a preferred embodiment of this invention, the real-time degradation data is input into a preset physical information neural network. The physical information neural network extracts local temporal abrupt change features and long-term temporal dependency features from the real-time degradation data using a built-in dual-channel encoder, fuses them, and then processes the fused features based on an embedded physical constraint loss function to output the remaining lifetime prediction result of the target device. Specifically: After completing the acquisition and preprocessing of real-time degradation data, this embodiment inputs the processed on-state voltage drop time series into a pre-trained physical information neural network. Figure 4The diagram illustrates the overall structure of the physical information neural network used in this embodiment. This network employs a sequence-to-sequence architecture and mainly consists of an encoder, a decoder, and a macroscopic state reconstruction module. The core processing flow of this network includes two stages: first, multi-scale features are extracted and fused using a built-in dual-channel encoder; then, the fused features are decoded and predicted based on a loss function with embedded physical constraints, ultimately outputting the remaining lifetime.

[0038] The physical information neural network's built-in dual-channel encoder consists of a parallel spatial mutation channel and a long-term dependency channel. Both channels simultaneously receive a preprocessed on-state voltage drop sequence as input.

[0039] In the spatial abrupt change channel, a multi-scale one-dimensional convolutional layer is first used to perform sliding convolution operations on the input sequence along the time dimension. Specifically, multiple convolutional kernels of different sizes are set (e.g., kernel lengths of 3, 5, and 7), and each kernel slides window by window along the time axis to extract local waveform features at different time spans, including micro-patterns such as abrupt peaks in the conduction voltage drop and changes in the inflection point slope. These multi-scale local features are then superimposed and input into the cascaded convolutional block attention module. Figure 5 The diagram illustrates the structure of the multi-scale convolution and attention module used in the spatial abrupt change channel in this embodiment. This module includes channel attention and spatial attention mechanisms. The calculation process for channel attention is as follows: Global average pooling and global max pooling are performed on the input feature map to obtain two channel descriptors. These two descriptors are then fed into a shared multilayer perceptron, where nonlinear transformations are performed using dimensionality reduction and dimensionality increase weights to generate a one-dimensional channel weight matrix. This process can be represented as: Where F is the input feature map, σ is the Sigmoid activation function, and MLP is a shared multilayer perceptron. The spatial attention calculation process is as follows: Average pooling and max pooling are performed on the input feature map along the channel dimension, the two resulting two-dimensional feature descriptors are concatenated along the channel axis, and then convolved using a convolution kernel with a large receptive field. Finally, the spatial weight matrix is ​​obtained by normalization using the Sigmoid function. In the formula, F represents the intermediate feature map tensor input to the network, and the network calculates the corresponding one-dimensional channel feature weight matrix M through the channel and spatial attention modules respectively. c (F) and the two-dimensional spatial feature weight matrix M sIn the calculation of channel attention, global average pooling (AvgPool(F)) and global max pooling (MaxPool(F)) are used to compress the input F along the spatial dimension, extracting the feature descriptor F that aggregates global spatial context information. c avg With F c max Subsequently, these two descriptors are fed into a shared multilayer perceptron (MLP) consisting of dimensionality-reduced weights W0 and dimensionality-increasing weights W1 to effectively learn the nonlinear interactions between channel features. In the spatial attention computation branch, pooling operations are performed along the channel dimension for compression. The resulting two two-dimensional spatial feature descriptors F... s avg With F s max The feature maps are concatenated along the channel axis using a concatenation operation [;]. The concatenated feature maps are then further processed by a convolutional kernel f with a large receptive field. 7×1 Convolutional operations are used to perform deep fusion of local spatiotemporal features. Finally, both channel and spatial modules are activated using the Sigmoid activation function at their ends. σ The final attention weights are smoothly normalized to the (0,1) interval. This joint attention mechanism achieves effective enhancement of key information and adaptive suppression of redundant features by accurately calculating the importance of features in different dimensions.

[0040] In long-term dependent channels, a bidirectional long short-term memory network is used to perform time-series evolution calculations on the conduction voltage drop sequence in both forward and reverse directions. Figure 6 This diagram illustrates the internal structure of a single cell in a Long Short-Term Memory (LSTM) network. Figure 7 A schematic diagram of the overall unfolded structure of a Bidirectional Long Short-Term Memory (BiLSTM) network is shown. This network contains a forget gate, an input gate, and an output gate, and its update rule can be expressed as: In the formula, x k h is the input vector at the current time step. k-1 with h k Let c represent the hidden states at the previous time step and the current time step, respectively. k-1 With c k This represents the cellular state responsible for maintaining long-term dependencies. In terms of information filtering, the forgetting gate F... k The cell state c at the previous moment is determined by the Sigmoid activation function σ. k-1 The amount of information to be retained; Input gate I k The candidate cell state G generated by the tanh function kThey work together to control the proportion of new information written at the current moment. The combination of these two processes completes the current cell state c. k Linear updates. Finally, output gate O... k Combined with the updated c k This determines the final hidden state output h at the current moment. k In the specific calculation process, W F,h W F,x These represent the various gating units in processing the historical hidden state h. k-1 With the current input x k The learnable weight matrix corresponding to time; b F , b I , b G , b O This is the corresponding bias term.

[0041] The local mutation feature vector output by the spatial mutation channel and the long-term temporal dependency feature vector output by the long-term dependency channel are concatenated and fused along the feature dimension to form a fused feature vector, which is then fed into the decoder.

[0042] The decoder receives the fused feature vector and generates the prediction result using a physically decoupled approach. The decoder first outputs two latent variable increments: the first latent variable increment ΔV, representing the threshold voltage drift. th (t), and the second hidden variable increment ΔV characterizing the encapsulation degradation effect. pack (t). To ensure that both strictly adhere to the physical law of irreversible damage accumulation, polarity constraints are applied using a soft linear unit activation function: Where S is the step size scaling factor, used to limit the single-step increment magnitude and ensure the smoothness of the evolution of the degenerate trajectory; β soft The shape parameter of the soft linear unit controls the degree of nonlinearity of the activation function near the zero point. , This is the predicted value output by the network after decoupling.

[0043] Subsequently, the preset first physical reconstruction coefficient α and second physical reconstruction coefficient β are obtained, and the observation value of the real-time degradation data at the last sampling time is used as the anchor point V. anchor The first latent variable increment after constraints is multiplied by α to obtain the first contribution component, and the second latent variable increment after constraints is multiplied by β to obtain the second contribution component. These components are then summed and superimposed using the anchor point as a reference to reconstruct the macroscopic degradation index. In the formula, α and β are physical reconstruction coefficients, which quantify the weights of the impact of threshold voltage drift and package degradation on macroscopic indicators, respectively. This decoupled design avoids the problem of the neural network being forced to output exponentially increasing values ​​at the end of its lifespan, thereby effectively alleviating the saturation effect of the long short-term memory network in long-term prediction and significantly improving the numerical stability of degradation trajectory extrapolation. In the online prediction stage, the latest real-time data captured by the sliding window is iteratively input into the trained physical information neural network. The model uses a single-step autoregressive strategy to extrapolate the future degradation trajectory: at each moment, the data in the current sliding window is input into the network, the predicted value for the next moment is output, and then this predicted value is used as a new input to roll backward, gradually generating a complete prediction sequence. The reconstructed on-state voltage drop prediction sequence is compared with a preset failure physical threshold boundary (e.g., on-state voltage drop exceeding 5% of the initial nominal value). When the predicted value first reaches or exceeds the threshold, the device is determined to have failed, and the difference between the failure time and the current time is the remaining lifetime prediction result. Figure 8 The diagram shows the predicted remaining lifetime of the target device in this embodiment under a junction temperature fluctuation of 130°C. The solid line represents the predicted remaining lifetime trajectory output by the model, and the dashed line represents the actual remaining lifetime trajectory. From... Figure 8 It can be seen that the predicted trajectory maintains a very small deviation from the actual trajectory, and as the monitoring period increases, the prediction results continuously approach the actual physical expectations.

[0044] Furthermore, to verify the effectiveness of the proposed device remaining lifetime prediction method, an aging experiment was conducted on a silicon carbide metal-oxide-semiconductor field-effect transistor (SiC MOSFET) under power cycling conditions with junction temperature fluctuations of 130°C, and on-state voltage drop data throughout its entire lifecycle were collected. The proposed method was compared with existing classical deep learning methods (including Long Short-Term Memory network models and Transformer models). To simulate the harsh conditions of sparse samples in real industrial scenarios, when the monitoring period reached 25%, 50%, and 75% of the total lifetime, the three prediction methods extracted only the sequence data from the most recent time window for independent regression prediction, without introducing any feedback or correction from subsequent real aging data during the entire forward inference process. The prediction results of remaining lifetime (RUL) are shown in Table 1.

[0045] Table 1 Comparison of prediction results with classic deep learning methods under different monitoring periods. As shown in Table 1, the purely data-driven Long Short-Term Memory (LSTM) network and Transformer model heavily rely on the accumulation of a large number of long-range time-series samples and lack the constraints of underlying physical laws. Therefore, they cannot output effective remaining lifetime predictions in the early and middle stages of device degradation (when the monitoring period is 25% and 50%). Even in the later stages where degradation characteristics are obvious (when the monitoring period is 75%), their predictions still show significant errors compared to the actual values. In contrast, the physical information neural network-driven method proposed in this application outputs prediction results of 69.31%, 46.13%, and 27.52% at monitoring periods of 25%, 50%, and 75%, respectively, closely approximating the actual remaining lifetime at the corresponding stages. This indicates that, by embedding physical consistency constraints, this application completely overcomes the prediction failure problem of traditional neural networks under conditions of limited observation windows and no subsequent data correction, achieving high-precision, smooth, and robust lifetime assessment at all stages of the device's entire lifecycle.

[0046] Furthermore, this application constructs a dual-channel parallel encoder, enabling a bidirectional long short-term memory network to deeply mine the sequence history dependence of macroscopic decay cycles. Simultaneously, multi-scale one-dimensional convolutional layers sensitively capture short-term fluctuations induced by local defects. By cascading channel attention and spatial attention mechanisms at the ends of the convolutional channels, adaptive enhancement of key failure features and effective suppression of redundant information are achieved. The synergistic effect of these techniques results in a significantly higher prediction accuracy than traditional cascaded single-channel networks.

[0047] In summary, this application acquires real-time degradation data through sensors and utilizes a pre-defined physical information neural network for feature extraction and lifetime prediction. This network incorporates a dual-channel encoder to simultaneously capture local mutations and long-term dependent features, and its training process embeds a physical penalty term to constrain the non-monotonic evolution characteristics of the degradation process. This technique deeply integrates data-driven learning capabilities with the physical laws of device degradation, resulting in predictions that not only possess high nonlinear fitting accuracy but also strictly adhere to the principle of thermodynamic irreversibility. This fundamentally eliminates the "lifetime bounce" or non-physical oscillation phenomena that easily occur in purely data-driven models, significantly improving prediction reliability and physical interpretability under sparse sample and complex operating conditions. This application effectively solves the problem that existing technologies cannot accurately and efficiently predict the remaining lifetime of devices.

[0048] Example 2 Please refer to Figure 9 This is a device for predicting the remaining lifetime of a device provided in the embodiments of this application.

[0049] In this embodiment, the device for predicting the remaining lifetime of the device includes an acquisition module 10 and a prediction module 20.

[0050] The acquisition module 10 is used to collect real-time degradation data of the target device during operation through a preset sensor; The prediction module 20 is used to input the real-time degradation data into a preset physical information neural network, so that the physical information neural network extracts local temporal mutation features and long-term temporal dependency features from the real-time degradation data through a built-in dual-channel encoder and fuses them. Then, it processes the fused features based on the embedded physical constraint loss function and outputs the remaining service life prediction result of the target device. The physical information neural network is trained based on historical degradation data and the physical constraint loss function, which incorporates a physical penalty term to constrain the non-monotonic evolution characteristics of the degradation process.

[0051] For ease of description and brevity, the embodiments of the device of the present invention include all the implementation methods in the above-described embodiments of the device remaining lifetime prediction method, and will not be repeated here.

[0052] Example 3 This application provides a computer-readable storage medium including a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to execute a method for predicting the remaining lifetime of a device. The method for predicting the remaining lifetime of a device, if implemented as a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0053] Example 4 This embodiment provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any of the device remaining lifetime prediction methods described in Embodiment 1.

[0054] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting the remaining lifetime of a device, characterized in that, include: Real-time degradation data of the target device during operation is collected through preset sensors; The real-time degradation data is input into a preset physical information neural network, so that the physical information neural network extracts local temporal mutation features and long-term temporal dependency features from the real-time degradation data through a built-in dual-channel encoder and fuses them. Then, the fused features are processed based on the embedded physical constraint loss function to output the remaining service life prediction result of the target device. The physical information neural network is trained based on historical degradation data and the physical constraint loss function, which incorporates a physical penalty term to constrain the non-monotonic evolution characteristics of the degradation process.

2. The method for predicting the remaining lifetime of a device according to claim 1, characterized in that, The process involves extracting and fusing local temporal abrupt change features and long-term temporal dependency features from the real-time degradation data using a built-in dual-channel encoder. The real-time degradation data is input into the spatial mutation channel and long-term dependency channel of the dual-channel encoder, respectively. Local temporal features are extracted by sliding filtering along the time dimension through the multi-scale one-dimensional convolutional layer in the spatial mutation channel, and the local temporal features are adaptively weighted by the channel attention mechanism to obtain the local mutation feature vector. The long-term time-dependent feature vector is obtained by performing forward and backward temporal evolution calculations on the real-time degraded data through the bidirectional long short-term memory network in the long-term dependent channel. The local mutation feature vector and the long-term temporal dependency feature vector are concatenated and fused to obtain the fused feature vector.

3. The method for predicting the remaining lifetime of a device according to claim 2, characterized in that, The adaptive weighting of the local temporal features through the channel attention mechanism is specifically as follows: The local temporal features are subjected to global average pooling and global max pooling respectively to obtain the first channel descriptor and the second channel descriptor; The first channel descriptor and the second channel descriptor are input into a shared multilayer perceptron, and a channel weight matrix is ​​generated by nonlinear transformation through a dimension reduction weight matrix and a dimension increase weight matrix. The local temporal features are weighted channel by channel according to the channel weight matrix to obtain a weighted feature sequence. The weighted feature sequence is then weighted a second time in terms of spatial dimension through a spatial attention mechanism to obtain the local mutation feature vector.

4. The method for predicting the remaining lifetime of a device according to claim 1, characterized in that, The embedded physical constraint loss function processes the fused features and outputs the remaining lifetime prediction result of the target device, specifically: The fused features are input into the decoder built into the physical information neural network, so that the decoder generates a first latent variable increment characterizing the threshold voltage drift and a second latent variable increment characterizing the encapsulation degradation effect. The first latent variable increment and the second latent variable increment are nonlinearly mapped by the built-in soft linear unit activation function to constrain the numerical signs of the first latent variable increment and the second latent variable increment during the update process to conform to the preset monotonically increasing or decreasing physical direction. Obtain the preset first physical reconstruction coefficient and second physical reconstruction coefficient, and obtain the observation value of the real-time degraded data at the last sampling time as the anchor point; The first contribution component is obtained by multiplying the first latent variable increment after constraint by the first physical reconstruction coefficient. The second contribution component is obtained by multiplying the second latent variable increment after constraint by the second physical reconstruction coefficient. The first contribution component and the second contribution component are summed together, and the summation result is superimposed on the anchor point as a reference to reconstruct the macroscopic degradation index. The remaining useful life prediction result of the target device is output based on the macroscopic degradation index.

5. The method for predicting the remaining lifetime of a device according to claim 1, characterized in that, The physical information neural network is trained based on historical degradation data and the physical constraint loss function, specifically as follows: Obtain historical degradation data; The historical degradation data is input into the initial physical information neural network, the encoder in the initial physical information neural network extracts the state vector of the historical degradation data, and the state vector is passed to the long short-term memory network in the decoder; The initial physical information neural network is trained based on a pre-defined scheduling-based teacher-forced strategy; During the training process, the value of the physical constraint loss function is calculated based on the predicted output of the initial physical information neural network and the corresponding real label. The network parameters of the initial physical information neural network are updated through the backpropagation algorithm based on the value of the physical constraint loss function until the error meets the preset requirements and the training stops, thus obtaining the trained physical information neural network. The physical constraint loss function includes a data fidelity term, a non-monotonic evolution penalty term, a step size constraint term, a hedging constraint term, and a late-stage rise constraint term. The data fidelity term measures the deviation between the predicted value and the true value of the physical information neural network. The non-monotonic evolution penalty term positively penalizes the prediction results that violate the preset physical symbol constraints based on the temporal change direction of the intermediate latent variables in the predicted degradation process. The step size constraint term constrains the local slope of the predicted degradation trajectory to suppress non-physical oscillations. The hedging constraint term penalizes the prediction results when the changing trends of multiple latent variables cancel each other out. The late-stage rise constraint term ensures that the remaining service life prediction can enter the preset failure state in the later stage.

6. The method for predicting the remaining lifetime of a device according to claim 1, characterized in that, After outputting the predicted remaining lifetime of the target device, the method further includes: The degradation prediction trajectory corresponding to the remaining useful life prediction result is compared with the preset failure physical threshold boundary; When the predicted value on the degradation prediction trajectory first reaches or exceeds the failure physical threshold boundary, the target device is determined to have reached a failure state. The difference between the predicted time when the failure state is triggered and the current time is determined as the final remaining service life.

7. A device for predicting the remaining lifetime of a device, characterized in that, include: The acquisition module is used to collect real-time degradation data of the target device during operation through preset sensors; The prediction module is used to input the real-time degradation data into a preset physical information neural network, so that the physical information neural network extracts local temporal mutation features and long-term temporal dependency features from the real-time degradation data through a built-in dual-channel encoder and fuses them. Then, based on the embedded physical constraint loss function, the fused features are processed to output the remaining service life prediction result of the target device. The physical information neural network is trained based on historical degradation data and the physical constraint loss function, which incorporates a physical penalty term to constrain the non-monotonic evolution characteristics of the degradation process.

8. The device for predicting the remaining lifetime of a device according to claim 7, characterized in that, The process involves extracting and fusing local temporal abrupt change features and long-term temporal dependency features from the real-time degradation data using a built-in dual-channel encoder. The real-time degradation data is input into the spatial mutation channel and long-term dependency channel of the dual-channel encoder, respectively. Local temporal features are extracted by sliding filtering along the time dimension through the multi-scale one-dimensional convolutional layer in the spatial mutation channel, and the local temporal features are adaptively weighted by the channel attention mechanism to obtain the local mutation feature vector. The long-term time-dependent feature vector is obtained by performing forward and backward temporal evolution calculations on the real-time degraded data through the bidirectional long short-term memory network in the long-term dependent channel. The local mutation feature vector and the long-term temporal dependency feature vector are concatenated and fused to obtain the fused feature vector.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the device remaining lifetime prediction method as described in any one of claims 1 to 6.

10. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the method for predicting the remaining lifetime of a device as described in any one of claims 1 to 6.

Images