Power device residual life prediction method and system, and computer equipment

Abstract

The invention provides a power device residual life prediction method and system and computer equipment, and belongs to the field of semiconductor device life prediction. The method comprises the following steps: collecting working state data of a power device in real time, and preprocessing the collected working state data to obtain time sequence data; inputting the preprocessed time sequence data into an LSTM model, performing feature extraction by using the LSTM model, capturing a health state and a degradation trend of the power device, and outputting a degradation feature vector; and inputting the degradation feature vector output by the LSTM model into a Bayesian network, and converting the degradation feature vector into probabilistic life distribution by using the Bayesian network to obtain a residual life prediction result. According to the method, the LSTM with strong time sequence prediction capability and the Bayesian network with strong probabilistic reasoning capability are combined to predict the residual life, so that the dynamic property and the accuracy of life prediction can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device lifetime prediction, and more specifically to a method and system for predicting the residual lifetime of power devices, and a computer device. Background Technology

[0002] Power devices (such as laterally diffused metal-oxide semiconductors, or LDMOS) are widely used in power management and RF power amplifiers due to their high voltage withstand capability, high current drive capability, and low power consumption. However, LDMOS devices are affected by thermal stress, electrical stress, radiation stress, and environmental factors (such as humidity and temperature fluctuations) during long-term operation, making their reliability and lifetime prediction critical issues. Degradation mechanisms of LDMOS devices include hot carrier effects, hot oxide breakdown, metal electromigration, and interface state generation. These degradation mechanisms lead to parameter drift in the device, such as charge trapping in the gate oxide layer, threshold voltage drift, and current decay.

[0003] Traditional LDMOS lifetime prediction methods typically rely on accelerated lifetime testing, which rapidly induces degradation by applying accelerated stress (such as high temperature and high voltage) under extreme conditions. Devices are usually exposed to extreme conditions such as high temperature, high voltage, and high current, which differ significantly from the actual operating environment of the device in real-world applications. The actual operating environment and conditions of electronic devices are diverse, including temperature fluctuations, load variations, and electromagnetic interference. These factors are often difficult to fully simulate in accelerated testing. Therefore, accelerated testing data may not adequately reflect the impact of the complex environment under actual operating conditions on device lifetime, leading to inconsistencies between predicted and actual lifetimes. Summary of the Invention

[0004] To address one of the aforementioned technical deficiencies, this invention provides a method and system for predicting the residual lifetime of power devices, as well as a computer device.

[0005] This invention provides a method for predicting the residual lifetime of power devices, comprising: Real-time acquisition of operating status data of power devices; preprocessing of the acquired operating status data to obtain time series data of power device operating status. The preprocessed power device operating state time series data is input into the LSTM model, and the LSTM model is used to extract features to capture the health status and degradation trend of the power device, and output the degradation feature vector characterizing the evolution of the health status of the power device. The degradation feature vector output by the LSTM model is input into a Bayesian network, which is then used to transform the degradation feature vector into a probabilistic lifetime distribution to obtain the residual lifetime prediction result.

[0006] In this embodiment of the invention, the power device residual lifetime prediction method further includes: Based on the real-time feedback of the remaining lifetime prediction results, the parameters of the LSTM model and the Bayesian network are dynamically updated.

[0007] In this embodiment of the invention, real-time acquisition of operating status data of power devices includes: By using various types of sensors integrated into the integrated circuit where the power device is located, the temperature, current or voltage of the power device during operation can be monitored and recorded in real time.

[0008] In this embodiment of the invention, the collected working status data is preprocessed to obtain time series data, including: The Kalman filter algorithm is used to denoise the collected working status data; The data after noise reduction is then standardized.

[0009] In this embodiment of the invention, the LSTM model includes: an input layer, multiple stacked LSTM units, and an output layer; The input layer is used to input preprocessed time-series data of the operating status of power devices; Multi-layer stacked LSTM units are used to extract degradation trend features by passing hidden states and memory states layer by layer; The output layer is used to output a degradation feature vector characterizing the evolution of the health state of the power device based on the degradation trend features extracted from the last LSTM unit.

[0010] In this embodiment of the invention, each LSTM unit in the multi-layer stacked LSTM unit includes: a memory unit, an input gate, a forget gate, and an output gate; The input gate is used to calculate the proportion of new information introduced at the current time step, generate candidate states, and multiply the candidate states by the output of the input gate and then add them to the memory unit to achieve the accumulation of long-term dependent features. The output gate is used to determine the hidden state of the output at the current time step, and combined with the transformation result of the memory unit, a degradation trend feature is generated. The forget gate is used to decide whether to discard or retain historical state information based on the current input and the hidden state of the previous time step by calculating the forgetting ratio.

[0011] In this embodiment of the invention, the LSTM model is trained based on historical power device operating state data; the training process of the LSTM model includes forward propagation, error calculation, and parameter update.

[0012] In this embodiment of the invention, the Bayesian network includes nodes and boundaries; The nodes include: a health status node representing the current degree of degradation of the power device, an environmental stress node representing external environmental factors, a lifetime output node representing the remaining usable lifetime, and a historical degradation pattern node obtained by feature extraction from the LSTM network. Boundaries are used to represent the conditional dependencies between nodes.

[0013] In this embodiment of the invention, a Bayesian network is used to transform the degradation feature vector into a probabilistic lifetime distribution to obtain the residual lifetime prediction result, including: Extract time-series features from the degenerative feature vector, use them as observational evidence to input into a node related to health status, and use that node as the observation node. After the observed nodes are assigned values, the posterior probability distribution of the unobserved nodes is calculated using an inference algorithm based on the conditional probability table. By using the posterior probability distribution of unobserved nodes as the probability distribution of the lifetime interval, the expected value and confidence interval of this distribution can be obtained. The expected value is used as the predicted residual lifetime, and the confidence interval is used as the lifetime value interval. The predicted residual lifetime is then output.

[0014] Another aspect of the present invention provides a power device residual lifetime prediction system, comprising: The data acquisition module is used to collect the operating status data of the power devices in real time, and to preprocess the collected operating status data to obtain the time series data of the operating status of the power devices. The LSTM model module is used to take the preprocessed power device operating state time series data as input, use the LSTM model to extract features, capture the health status and degradation trend of the power device, and output the degradation feature vector characterizing the evolution of the power device's health status. The Bayesian network module is used to input the degradation feature vector and transform it into a probabilistic lifetime distribution to obtain the residual lifetime prediction result.

[0015] In this embodiment of the invention, the LSTM model includes: an input layer, multiple stacked LSTM units, and an output layer; The input layer is used to input preprocessed time-series data of the operating status of power devices; Multi-layer stacked LSTM units are used to extract degradation trend features by passing hidden states and memory states layer by layer; The output layer is used to output a degradation feature vector characterizing the evolution of the health state of the power device based on the degradation trend features extracted from the last LSTM unit.

[0016] In this embodiment of the invention, each LSTM unit in the multi-layer stacked LSTM unit includes: a memory unit, an input gate, a forget gate, and an output gate; The input gate is used to calculate the proportion of new information introduced at the current time step, generate candidate states, and multiply the candidate states by the output of the input gate and then add them to the memory unit to achieve the accumulation of long-term dependent features. The output gate is used to determine the hidden state of the output at the current time step, and combined with the transformation result of the memory unit, a degradation trend feature is generated. The forget gate is used to decide whether to discard or retain historical state information based on the current input and the hidden state of the previous time step by calculating the forgetting ratio.

[0017] In this embodiment of the invention, the Bayesian network includes nodes and boundaries; the nodes include: health state nodes representing the current degradation level of the power device, environmental stress nodes representing external environmental factors, lifetime output nodes representing the remaining usable lifetime, and historical degradation mode nodes obtained by feature extraction from the LSTM network; the boundaries are used to represent the conditional dependencies between the nodes.

[0018] In this embodiment of the invention, a Bayesian network is used to transform the degradation features of the degradation feature vector into a probabilistic lifetime distribution to obtain the residual lifetime prediction result, including: Extract time-series features from the degenerative feature vector, use them as observational evidence to input into a node related to health status, and use that node as the observation node. After the observed nodes are assigned values, the posterior probability distribution of the unobserved nodes is calculated using an inference algorithm based on the conditional probability table. By using the posterior probability distribution of unobserved nodes as the probability distribution of the lifetime interval, the expected value and confidence interval of this distribution can be obtained. The expected value is used as the predicted residual lifetime, and the confidence interval is used as the lifetime value interval. The predicted residual lifetime is then output.

[0019] The present invention also provides a computer device, including: a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method for predicting the residual lifetime of power devices.

[0020] The above technical solution dynamically combines the LSTM model with strong time-series prediction capabilities with the Bayesian network with strong probabilistic reasoning capabilities, breaking through the limitations of a single model. After capturing degradation features through LSTM, the Bayesian network is used to perform probabilistic inference by integrating multiple factors, which significantly improves the dynamism and accuracy of lifetime prediction.

[0021] Other features and advantages of the technical solution of the present invention will be described in detail in the following detailed embodiments section. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of the power device residual lifetime prediction method provided in the embodiments of the present invention; Figure 2 This is a block diagram of the power device residual lifetime prediction system provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the technical solutions and advantages of the embodiments of the present invention clearer, the exemplary embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] Figure 1 This is a flowchart of the power device residual lifetime prediction method provided in an embodiment of the present invention. Figure 1 As shown, the power device residual lifetime prediction method provided in this embodiment includes the following steps: S110: Real-time acquisition of operating status data of power devices; preprocessing of the acquired operating status data to obtain time series data of operating status of power devices. S120: Input the preprocessed power device operating state time series data into the LSTM model, use the LSTM model to extract features, capture the health status and degradation trend of the power device, and output the degradation feature vector characterizing the evolution of the health status of the power device. S130: Input the degradation feature vector output by the LSTM model into the Bayesian network, and use the Bayesian network to transform the degradation feature vector into a probabilistic lifetime distribution to obtain the residual lifetime prediction result.

[0025] In step S110 above, virtual measurement technology is used to collect real-time operating status data of the power device. For example, various types of sensors (such as temperature sensors, current sensors, and voltage sensors) integrated into the integrated circuit where the power device is located are used to monitor and record data such as temperature, current, or voltage of the power device in real time during operation. A Kalman filter algorithm is used to denoise the collected operating status data. The denoised data is then standardized to convert it into a unified dimension, ensuring data quality and consistency.

[0026] In step S120 above, based on the application scenario of power device residual lifetime prediction, an LSTM (Long Short-Term Memory) model structure is designed, and the LSTM model is trained based on historical power device operating state data. The training process includes forward propagation, error calculation, and parameter update. The backpropagation algorithm is used to optimize the parameters of the LSTM model so that it can accurately predict the health status and degradation trend of power devices. The LSTM model is specifically designed for processing time series data and can capture the degradation patterns of power devices under different stress conditions. The preprocessed time series data is input into the LSTM model, and feature extraction and pattern recognition are performed through multi-layer LSTM units to capture the health status and degradation trend of power devices, outputting a degradation feature vector (i.e., degradation feature curve) to characterize the evolution of the health status of power devices.

[0027] The LSTM model consists of an input layer, multiple stacked LSTM cells, and an output layer. The input layer receives preprocessed time-series data of the power device's operating state. The input data dimension is (T, N), where T is the time step and N is the feature dimension, such as multi-sensor fusion signals like temperature, current, and voltage.

[0028] Each LSTM unit in a multi-layered stacked LSTM unit includes a memory unit (Cell State), an input gate, a forget gate, and an output gate. The input gate calculates the proportion of new information introduced at the current time step and generates candidate states using the Tanh function. These candidate states are multiplied by the input gate output and then added to the memory unit, accumulating long-term dependent features. The output gate determines the hidden state of the current time step output and, combined with the Tanh transform result of the memory unit, generates the final output vector. The forget gate calculates the forgetting proportion based on the current input and the hidden state of the previous time step using the Sigmoid function, deciding whether to discard or retain historical state information. By progressively passing hidden and memory states, multi-layered stacked LSTM units achieve hierarchical extraction of multi-scale temporal features, simultaneously capturing short-term fluctuation patterns and long-term degradation trends.

[0029] The output layer of the LSTM model performs dimensionality reduction or linear mapping on the hidden state sequence of the last LSTM unit across all time steps, outputting a multidimensional feature vector characterizing the evolution of the power device's health state.

[0030] In step S130 above, the Bayesian network includes nodes and boundaries. Nodes are used to characterize the health state of the LDMOS device, and boundaries are used to represent the conditional dependencies between nodes characterizing the health state. The degradation feature vector output by the LSTM model is input into the Bayesian network, which is then used to transform the degradation feature vector into a probabilistic lifetime distribution. The residual lifetime of the device is calculated based on the probabilistic lifetime distribution.

[0031] In this embodiment of the invention, the Bayesian network is a Directed Acyclic Graph (DAG), consisting of nodes and directed edges. Nodes include: health state nodes (representing the current degree of device degradation), environmental stress nodes (external environmental factors such as temperature, current, and voltage), historical degradation pattern nodes (obtained by feature extraction from an LSTM network), and lifetime output nodes (representing the remaining usable lifetime). Directed edges, or boundaries, represent the conditional dependencies between nodes; for example, environmental stress nodes influence health state nodes, and health state nodes influence lifetime output nodes. The Conditional Probability Table (CPT) stores the probability distribution P(X_i|Parents(X_i)) of each node under given conditions of its parent node.

[0032] The residual lifetime distribution of the device is calculated using conditional probabilistic inference via Bayesian networks. The specific process is as follows: 1. Input and Observation Assignment: After the LSTM model generates the degradation feature vector (degradation feature curve), time-series features (e.g., degradation rate, health index trend slope, local fluctuation amplitude, etc.) are extracted from this curve and input as observational evidence into the node related to health status: Evidence={H(t),R(t),S(t)}; where H(t) is the health index, R(t) is the degradation rate, and S(t) is the short-term fluctuation feature. This node is the observation node.

[0033] 2. Conditional Probability Inference: After the observed nodes are assigned values, the Bayesian network uses the conditional probability formula: P(X|Evidence)=P(X,Evidence) / P(Evidence), combined with the network structure and conditional probability table, and employs inference algorithms (such as variable elimination, belief propagation, etc.) to calculate the posterior probability distribution of unobserved nodes. The posterior distribution of the lifetime output node L is: P(L|Evidence)=Σ_all unobservedvariables Π_iP(X_i|Parents(X_i)).

[0034] 3. Remaining life distribution generation: The posterior distribution P(L|Evidence) of the inferred life output node L is a probability distribution of the life interval, reflecting the likelihood distribution of the remaining available time of the device in the current state. This distribution can obtain the expected value (as the remaining life prediction value) and the confidence interval through numerical integration or discrete summation. The expected value is: E[L]=Σ_l l*P(L=l|Evidence); The confidence interval (e.g., 95% confidence interval) is to find the life value interval corresponding to the positions of the probability cumulative distribution function CDF at 2.5% and 97.5%.

[0035] 4. Output and application: Use E[L] as the remaining life prediction value, and combine it with the confidence interval or the failure probability vector to provide a basis for risk assessment for device maintenance. For example, when P(L<T_threshold) is greater than the preset value, maintenance or replacement strategies can be triggered to reduce the failure risk.

[0036] Through the above process, the Bayesian network realizes the calculation from the degradation feature vector to the remaining life probability distribution, can provide both a single prediction value and a measure of uncertainty, and significantly improves the reliability and practicality of life prediction.

[0037] In another embodiment, the above-mentioned remaining life prediction method for power devices further includes the following steps: According to the real-time feedback of the remaining life prediction results, dynamically update the parameters of the LSTM model and the parameters of the Bayesian network.

[0038] Feed the prediction results obtained by the Bayesian network back to the prediction system in real time, update the parameters of the LSTM model and the Bayesian network to adapt to new data and changing working conditions. The prediction system continuously optimizes and adjusts the prediction models (LSTM model and Bayesian network) through the feedback loop, which can improve the accuracy and real-time performance of life prediction.

[0039] Specifically, LSTM models and Bayesian networks are deployed in edge computing devices close to the power device operating environment, enabling real-time monitoring and immediate feedback. By monitoring the operating status of power devices in real time, the lifetime prediction system is adjusted based on the latest data and prediction results, improving the accuracy and reliability of predictions. The real-time feedback mechanism optimizes production processes, extends the lifespan of power devices, and improves production efficiency and product quality. These edge computing devices can be: industrial embedded controllers, smart gateway devices, field-programmable logic controllers (PLCs), or FPGA-SoC modules. Industrial embedded controllers are used to directly connect to sensors and control systems on the production line. Smart gateway devices have local data caching, computing, and communication capabilities, enabling the execution of deep learning inference engines and probabilistic inference modules at the edge. Field-programmable logic controllers (PLCs) or FPGA-SoC modules can deploy models in hardware-accelerated form within the FPGA's computing units when low-latency response is required.

[0040] Edge computing devices receive real-time operating status data from power devices via high-speed buses (such as SPI, Ethernet, CAN, or industrial fieldbus). After processing by the data acquisition and preprocessing modules, the data is sequentially input into an LSTM model and a Bayesian network to obtain the residual lifetime prediction results. The prediction results are immediately fed back through the following methods after generation: 1. Production Control System Interface: The prediction results are sent to the host computer or MES system via OPC UA or Modbus protocol to realize dynamic adjustment of process parameters such as production cycle time and load distribution; 2. Local alarms and maintenance prompts: Trigger visual / audible alarms on edge devices, or push events with remaining lifespan below a threshold to maintenance personnel's terminals via MQTT / HTTP.

[0041] In this embodiment of the invention, the parameter update mechanism of the LSTM model and the Bayesian network is utilized. The LSTM model can adaptively adjust its feature extraction capability according to the latest running data, and the Bayesian network can correct the conditional probability distribution according to the latest state data, thereby ensuring the dynamism and accuracy of the prediction results and realizing the closed-loop operation of prediction and feedback.

[0042] Specifically, the latest prediction results and their deviation from subsequent actual lifetime detection data are received through a real-time feedback module, forming an online updated dataset. Model parameters are then updated through incremental learning or mini-batch training. The parameters updated for the LSTM model include the weight matrix from the input layer to the hidden layers, and the weights and biases of each gate (input gate, forget gate, output gate). The update method is Backpropagation Through Time (BPTT), with the learning rate adaptively adjusted based on the real-time data changes. The parameters updated for the Bayesian network include the probability values ​​in the conditional probability table (CPT) of each node. The update method is based on the parameter learning formula according to Bayes' theorem. P_new(X_i|Parents(X_i))=(count(X_i,Parents(X_i))+α) / (count(Parents(X_i))+α|X_i|); where α is a smoothing factor used to avoid probability distortion caused by low-frequency events.

[0043] By employing the above method, the LSTM model, with its strong temporal prediction capabilities, is dynamically combined with the Bayesian network, which boasts strong probabilistic inference capabilities. This overcomes the limitations of a single model. After capturing degradation features using LSTM, the Bayesian network integrates multiple factors for probabilistic inference, significantly improving the dynamism and accuracy of lifetime prediction. Furthermore, deploying the LSTM model and Bayesian network on edge computing devices enables dynamic updates and real-time feedback of model parameters. This overcomes the limitations of traditional offline-trained static models, allowing the prediction model to continuously adapt to changes in device state and ensuring the timeliness and reliability of prediction results.

[0044] Figure 2 This is a block diagram of the power device residual lifetime prediction system provided in an embodiment of the present invention. Figure 2 As shown, the power device residual lifetime prediction system provided in this embodiment includes: a data acquisition module, an LSTM model module, and a Bayesian network module. The data acquisition module is used to collect real-time operating state data of the power device, preprocess the collected operating state data to obtain power device operating state time series data. The LSTM model module is used to input the preprocessed power device operating state time series data, use the LSTM model for feature extraction, capture the health status and degradation trend of the power device, and output a degradation feature vector characterizing the evolution of the power device's health status. The Bayesian network module is used to input the degradation feature vector, use a Bayesian network to transform the degradation feature vector into a probabilistic lifetime distribution, and obtain the residual lifetime prediction result.

[0045] In other embodiments, the power device residual lifetime prediction system further includes a feedback update module, which is used to dynamically update the parameters of the LSTM model and the parameters of the Bayesian network based on the real-time feedback residual lifetime prediction results.

[0046] In a specific embodiment, the data acquisition module utilizes virtual measurement technology to collect real-time operating status data of the power devices. For example, it monitors and records data such as temperature, current, and voltage of the power devices in real time using various types of sensors (such as temperature sensors, current sensors, and voltage sensors) integrated into the integrated circuit where the power devices are located. A Kalman filter algorithm is used to denoise the collected operating status data, and the denoised data is then standardized to convert it into a unified dimension, ensuring data quality and consistency.

[0047] In a specific embodiment, the LSTM model module, based on the application scenario of power device residual lifetime prediction, designs an LSTM (Long Short-Term Memory) model structure and trains the LSTM model based on historical power device operating state data. The training process includes forward propagation, error calculation, and parameter update. The backpropagation algorithm is used to optimize the LSTM model parameters, enabling it to accurately predict the health status and degradation trend of power devices. The LSTM model is specifically designed for processing time series data and can capture the degradation patterns of power devices under different stress conditions. Preprocessed time series data is input into the LSTM model, and feature extraction and pattern recognition are performed through multi-layer LSTM units to capture the health status and degradation trend of power devices, outputting a degradation feature vector.

[0048] The LSTM model consists of an input layer, multiple stacked LSTM cells, and an output layer. The input layer receives preprocessed time-series data of the power device's operating state. The input data dimension is (T, N), where T is the time step size and N is the feature dimension, such as multi-sensor fusion signals like temperature, current, and voltage. Each LSTM cell in the multiple stacked cells includes a memory cell, an input gate, a forget gate, and an output gate. The input gate calculates the proportion of new information introduced at the current time step and generates candidate states using the Tanh function. These candidate states are multiplied by the input gate output and then added to the memory cell, accumulating long-term dependent features. The output gate determines the hidden state of the current time step output and combines it with the Tanh transform result of the memory cell to generate the final output vector. The forget gate calculates the forgetting proportion based on the current input and the hidden state of the previous time step using the Sigmoid function, deciding whether to discard or retain historical state information. Multi-layer stacked LSTM units achieve hierarchical extraction of multi-scale temporal features by passing hidden and memory states layer by layer, enabling the simultaneous capture of short-term fluctuation patterns and long-term degradation trends. The output layer of the LSTM model performs dimensionality reduction or linear mapping on the hidden state sequence of the last LSTM unit across all time steps, outputting a multi-dimensional feature vector characterizing the evolution of the power device's health state.

[0049] In a specific embodiment, the degradation feature vector output by the LSTM model is input into a Bayesian network. The Bayesian network transforms the degradation features into a probabilistic lifetime distribution, and the remaining lifetime of the device is calculated based on this distribution. The Bayesian network is a directed acyclic graph (DAG), consisting of nodes and directed edges. Nodes include: health state nodes (representing the current degree of device degradation), environmental stress nodes (external environmental factors such as temperature, current, and voltage), historical degradation pattern nodes (obtained by feature extraction from the LSTM network), and lifetime output nodes (representing the remaining usable lifetime). Directed edges, or boundaries, represent the conditional dependencies between nodes; for example, environmental stress nodes influence health state nodes, and health state nodes influence lifetime output nodes. The conditional probability table (CPT) stores the probability distribution of each node under given conditions of its parent node.

[0050] This paper utilizes Bayesian networks to transform degraded feature vectors into probabilistic lifetime distributions, yielding residual lifetime prediction results. The specific process is as follows: Time-series features are extracted from the degraded feature vectors and input as observational evidence into nodes related to health status, which are then designated as observation nodes. After the observation nodes are assigned values, the posterior probability distribution of unobserved nodes is calculated using an inference algorithm based on a conditional probability table. This posterior probability distribution of unobserved nodes is then used as the probability distribution of the lifetime interval, yielding the expected value and confidence interval of this distribution. The expected value is used as the predicted residual lifetime value, and the confidence interval is used as the lifetime value interval, outputting the residual lifetime prediction result. Bayesian networks enable the calculation of residual lifetime probability distributions from degraded feature vectors, providing both a single predicted value and an uncertainty measure, significantly improving the reliability and practicality of lifetime prediction.

[0051] In an optional embodiment, the aforementioned power device residual lifetime prediction system further includes a feedback update module. This module dynamically updates the parameters of the LSTM model and the Bayesian network based on the real-time feedback of the residual lifetime prediction results. The LSTM model and Bayesian network are deployed in an edge computing device close to the power device's operating site, enabling real-time monitoring and immediate feedback. The prediction results obtained by the Bayesian network are fed back to the edge computing device in real time, updating the parameters of the LSTM model and Bayesian network to adapt to new data and changing operating conditions. Continuously optimizing and adjusting the prediction model (LSTM model and Bayesian network) through the feedback loop improves the accuracy and real-time performance of lifetime prediction. Real-time monitoring of the power device's operating status via the edge computing device, and adjusting the power device residual lifetime prediction system based on the latest data and prediction results, improves the accuracy and reliability of the prediction. Utilizing the real-time feedback mechanism optimizes the production process, extends the lifespan of power devices, and improves production efficiency and product quality.

[0052] The specific details of the power device residual lifetime prediction system provided in this embodiment are the same as the specific details of the power device residual lifetime prediction method described above, and can be understood by referring to the specific details of the power device residual lifetime prediction method described above.

[0053] The power device residual lifetime prediction method and system provided by this invention have the following advantages compared with the prior art: 1. Significantly improves the dynamism and accuracy of lifetime prediction. Existing technologies typically rely on a single model, making it difficult to simultaneously address temporal feature extraction and uncertainty quantification. This invention, however, captures the nonlinear temporal patterns in device degradation through the gating mechanism of LSTM, while utilizing a Bayesian network to transform degradation features into a probabilistic lifetime distribution. Furthermore, by combining a dynamic parameter update mechanism implemented with edge computing devices, model parameters can be continuously optimized based on real-time monitoring data. This effectively addresses prediction needs under complex operating conditions such as device aging and sudden load changes, overcoming the lag problem of traditional static models.

[0054] 2. Constructing a Real-Time Adaptive Prediction System. Addressing the problem that offline-trained models in existing technologies struggle to adapt to dynamic working environments, this invention embeds prediction functionality into the device terminal through an edge computing architecture, achieving closed-loop operation of data acquisition, model inference, and parameter updates. Specifically, the system continuously iterates the weight parameters of the LSTM network through a real-time feedback mechanism and dynamically adjusts the conditional probability relationships of state nodes based on Bayesian inference. This enables the prediction system to autonomously adapt to gradual or abrupt changes in the device's operating state, significantly improving response speed and prediction stability under sudden operating conditions.

[0055] 3. Multi-dimensional data quality optimization. Existing technologies often suffer from limited input information due to the single type of sensor or insufficient data preprocessing. However, this invention integrates multi-modal sensors into power devices to simultaneously collect multi-dimensional state parameters such as temperature, current, and voltage. It also combines Kalman filtering algorithm to denoise and align the raw data in time series, and standardizes the numerical range of data with different dimensions to provide highly consistent and low-noise input features for subsequent models, thereby improving the reliability of the entire prediction system from the data source.

[0056] 4. Deep Modeling of Device Degradation Mechanisms. Traditional methods, limited by model structure, struggle to capture long-term degradation features. This invention employs a multi-layer LSTM network that, through cascaded memory units and gating mechanisms, achieves hierarchical extraction of short-term fluctuations and long-term trends. For example, it can simultaneously identify the impact of instantaneous current peaks on devices and the material performance degradation caused by temperature accumulation. This hierarchical feature extraction mechanism significantly enhances the model's ability to characterize complex degradation modes, particularly suitable for identifying the coupling effect between progressive and sudden failures.

[0057] 5. The output format of probabilistic lifetime prediction is adopted. Unlike existing technologies that only provide a single lifetime estimate, this invention uses a probabilistic inference framework built with Bayesian networks to output the probability distribution of residual lifetime and its confidence interval. This technique combines observational evidence from healthy nodes with prior knowledge, using a probabilistic model to quantify the uncertainty of the prediction results. This provides a risk assessment basis for operational decisions; for example, when formulating maintenance strategies, the median lifetime prediction and the failure probability distribution can be comprehensively weighed, thereby avoiding overly conservative or risky decision-making biases.

[0058] The present invention also provides a computer device, including: a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method for predicting the residual lifetime of power devices.

[0059] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for predicting the residual lifetime of power devices, characterized in that, include: Real-time acquisition of operating status data of power devices; preprocessing of the acquired operating status data to obtain time series data of power device operating status. The preprocessed power device operating state time series data is input into the LSTM model, and the LSTM model is used to extract features to capture the health status and degradation trend of the power device, and output the degradation feature vector characterizing the evolution of the health status of the power device. The degradation feature vector output by the LSTM model is input into a Bayesian network, which is then used to transform the degradation feature vector into a probabilistic lifetime distribution to obtain the residual lifetime prediction result.

2. The method for predicting the residual lifetime of power devices according to claim 1, characterized in that, The method further includes: Based on the real-time feedback of the remaining lifetime prediction results, the parameters of the LSTM model and the Bayesian network are dynamically updated.

3. The method for predicting the residual lifetime of power devices according to claim 1, characterized in that, The real-time acquisition of operating status data of power devices includes: By using various types of sensors integrated into the integrated circuit where the power device is located, the temperature, current or voltage of the power device during operation can be monitored and recorded in real time.

4. The method for predicting the residual lifetime of power devices according to claim 1, characterized in that, The process of preprocessing the collected work status data to obtain time series data includes: The Kalman filter algorithm is used to denoise the collected working status data; The data after noise reduction is then standardized.

5. The method for predicting the residual lifetime of power devices according to claim 1, characterized in that, The LSTM model includes: an input layer, multiple stacked LSTM cells, and an output layer; The input layer is used to input preprocessed time-series data of the operating status of power devices; Multi-layer stacked LSTM units are used to extract degradation trend features by passing hidden states and memory states layer by layer; The output layer is used to output a degradation feature vector characterizing the evolution of the health state of the power device based on the degradation trend features extracted from the last LSTM unit.

6. The method for predicting the residual lifetime of power devices according to claim 5, characterized in that, Each LSTM cell in a multi-layer stacked LSTM cell includes: a memory cell, an input gate, a forget gate, and an output gate; The input gate is used to calculate the proportion of new information introduced at the current time step, generate candidate states, and multiply the candidate states by the output of the input gate and then add them to the memory unit to achieve the accumulation of long-term dependent features. The output gate is used to determine the hidden state of the output at the current time step, and combined with the transformation result of the memory unit, a degradation trend feature is generated. The forget gate is used to decide whether to discard or retain historical state information based on the current input and the hidden state of the previous time step by calculating the forgetting ratio.

7. The method for predicting the residual lifetime of power devices according to claim 1, characterized in that, The LSTM model is trained based on historical power device operating state data; The training process of the LSTM model includes: forward propagation, error calculation, and parameter update.

8. The method for predicting the residual lifetime of power devices according to claim 1, characterized in that, The Bayesian network includes nodes and boundaries; The nodes include: a health status node representing the current degree of degradation of the power device, an environmental stress node representing external environmental factors, a lifetime output node representing the remaining usable lifetime, and a historical degradation pattern node obtained by feature extraction from the LSTM network. The boundary is used to represent the conditional dependencies between nodes.

9. The method for predicting the residual lifetime of power devices according to claim 8, characterized in that, The process of transforming the degradation feature vector into a probabilistic lifetime distribution using a Bayesian network to obtain the residual lifetime prediction result includes: Extract time-series features from the degenerative feature vector, use them as observational evidence to input into a node related to health status, and use that node as the observation node. After the observed nodes are assigned values, the posterior probability distribution of the unobserved nodes is calculated using an inference algorithm based on the conditional probability table. By using the posterior probability distribution of unobserved nodes as the probability distribution of the lifetime interval, the expected value and confidence interval of this distribution can be obtained. The expected value is used as the predicted residual lifetime, and the confidence interval is used as the lifetime value interval. The predicted residual lifetime is then output.

10. A power device residual lifetime prediction system, characterized in that, Includes: The data acquisition module is used to collect the operating status data of the power devices in real time, and to preprocess the collected operating status data to obtain the time series data of the operating status of the power devices. The LSTM model module is used to take the preprocessed power device operating state time series data as input, use the LSTM model to extract features, capture the health status and degradation trend of the power device, and output the degradation feature vector characterizing the evolution of the power device's health status. The Bayesian network module is used to input the degradation feature vector and transform it into a probabilistic lifetime distribution to obtain the residual lifetime prediction result.

11. The power device residual lifetime prediction system according to claim 10, characterized in that, The LSTM model includes: an input layer, multiple stacked LSTM cells, and an output layer; The input layer is used to input preprocessed time-series data of the operating status of power devices; Multi-layer stacked LSTM units are used to extract degradation trend features by passing hidden states and memory states layer by layer; The output layer is used to output a degradation feature vector characterizing the evolution of the health state of the power device based on the degradation trend features extracted from the last LSTM unit.

12. The power device residual lifetime prediction system according to claim 11, characterized in that, Each LSTM cell in a multi-layer stacked LSTM cell includes: a memory cell, an input gate, a forget gate, and an output gate; The input gate is used to calculate the proportion of new information introduced at the current time step, generate candidate states, and multiply the candidate states by the output of the input gate and then add them to the memory unit to achieve the accumulation of long-term dependent features. The output gate is used to determine the hidden state of the output at the current time step, and combined with the transformation result of the memory unit, a degradation trend feature is generated. The forget gate is used to decide whether to discard or retain historical state information based on the current input and the hidden state of the previous time step by calculating the forgetting ratio.

13. The power device residual lifetime prediction system according to claim 10, characterized in that, The Bayesian network includes nodes and boundaries; The nodes include: a health status node representing the current degree of degradation of the power device, an environmental stress node representing external environmental factors, a lifetime output node representing the remaining usable lifetime, and a historical degradation pattern node obtained by feature extraction from the LSTM network. The boundary is used to represent the conditional dependencies between nodes.

14. The power device residual lifetime prediction system according to claim 13, characterized in that, The process of using a Bayesian network to transform the degradation features of the degradation feature vector into a probabilistic lifetime distribution to obtain the residual lifetime prediction result includes: Extract time-series features from the degenerative feature vector, use them as observational evidence to input into a node related to health status, and use that node as the observation node. After the observed nodes are assigned values, the posterior probability distribution of the unobserved nodes is calculated using an inference algorithm based on the conditional probability table. By using the posterior probability distribution of unobserved nodes as the probability distribution of the lifetime interval, the expected value and confidence interval of this distribution can be obtained. The expected value is used as the predicted residual lifetime, and the confidence interval is used as the lifetime value interval. The predicted residual lifetime is then output.

15. A computer device, characterized in that, include: Memory, which stores computer programs; A processor for executing the computer program to implement the power device residual lifetime prediction method according to any one of claims 1-9.

Images