An electronic component real-time running state monitoring method and system

Abstract

This invention relates to a method and system for real-time monitoring of the operating status of electronic components, belonging to the field of electronic component condition monitoring technology. The method collects multi-dimensional operating data from multiple sources of sensors, evaluates the remaining lifetime prediction value and its prediction variance using a Bayesian neural network, assigns fusion weights based on the lifetime prediction value, and extracts deep features using a sparse autoencoder after attention fusion. The prediction variance serves as a regularization constraint to dynamically adjust feature extraction. The deep features are input into a lightweight neural network for state recognition, and an integral gradient method is used for fault attribution. The main cause channels with contribution exceeding a threshold are fed back to the data acquisition end to adjust the sampling frequency of the corresponding sensors. This invention achieves accurate perception and interpretable diagnosis of the operating status of electronic components, significantly improving the adaptability and reliability of the monitoring system.

Description

Technical Field

[0001] This invention relates to the field of electronic component status monitoring technology, specifically to a method and system for real-time monitoring of the operating status of electronic components. Background Technology

[0002] Electronic components are the core components of various electronic devices and systems, and their operational reliability directly affects the safety and stability of the entire system. Taking IGBT power modules as an example, their sudden failure can lead to serious consequences such as the shutdown of the motor drive system in new energy vehicles and the blocking of converter valves in smart grids. Currently, many technical solutions have been proposed for monitoring the operational status of electronic components, which can be mainly divided into the following categories: First, threshold monitoring methods based on single physical quantities, which set fixed thresholds for single parameters such as voltage, current, or temperature to trigger over-limit alarms; second, parameter identification methods based on physical models, which assess the status of components by establishing electrothermal models or failure physical models; and third, data-driven machine learning monitoring methods, which use algorithms such as neural networks to process sensor data to achieve fault identification. While these technologies have certain effectiveness under specific conditions, as electronic systems develop towards higher power density, higher integration, and more complex operating conditions, existing monitoring methods are gradually revealing many shortcomings in terms of sensing dimensions, response time, assessment reliability, and diagnostic interpretability.

[0003] Specifically, existing technologies suffer from the following drawbacks: First, they lack a single monitoring dimension. Traditional fault detection methods typically focus on only a single indicator for fault monitoring, making it difficult to comprehensively assess the health status of components under multi-field coupled failure modes. Furthermore, the linear threshold determination of a single physical quantity exhibits a high rate of missed detections and false positives when dealing with the nonlinear degradation process of deep submicron-level electronic components. Second, there is a significant contradiction between real-time performance and accuracy. While deep learning-based monitoring schemes offer high diagnostic accuracy, they require a large number of model parameters and high computational resources, making it difficult to achieve real-time inference on resource-constrained edge computing devices. On the other hand, traditional lightweight methods, although computationally simple, have limited accuracy. Third, existing life assessment methods are mostly deterministic, failing to quantify the reliability of the assessment results. When the input data contains noise or the components are in a highly nonlinear degradation stage, the uncertainty of the assessment results is high, yet lacks corresponding confidence characterization. Fourth, the monitoring strategy is statically fixed, with the system using a fixed sampling frequency and fixed fusion weights, making it impossible to adaptively adjust the monitoring strategy based on the actual degradation state of the components and real-time diagnostic results. Fifth, the inherent "black box" nature of deep learning models leads to a lack of interpretability in fault diagnosis results, making it difficult for engineers to understand the basis of the model's decisions and to make accurate fault location and maintenance decisions based on the diagnostic results. To address the above technical problems, this invention provides an artificial intelligence-based method and system for monitoring the operating status of electronic components. Summary of the Invention

[0004] In order to solve the problems of the prior art, the present invention provides a method and system for real-time monitoring of the operating status of electronic components.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Firstly, a method for real-time monitoring of the operating status of electronic components, comprising a data acquisition step, a data preprocessing step, and a monitoring result output step. The data acquisition step collects multi-source sensing data through multiple sensors deployed on the electronic components. The multi-source sensing data includes at least voltage data, current data, temperature data, and frequency data. The initial sampling frequency of each sensor is [missing information]. to The data preprocessing step involves cleaning, denoising, and temporal alignment of the multi-source sensor detection data to obtain preprocessed multi-source sensor data, and further includes: The operational life assessment step involves inputting the preprocessed multi-source sensor data into an operational life assessment model constructed based on a Bayesian neural network to obtain the predicted remaining life of the current electronic components and the prediction variance of the predicted remaining life. The prediction variance is used to quantify the uncertainty of the output result of the operational life assessment model. In the multi-source fusion step, based on the remaining lifetime prediction value, fusion weights are assigned to each sensor data source. The sensor data source with the smaller remaining lifetime prediction value has a higher weight in the fusion. A multi-source data fusion network based on an attention mechanism is used to fuse the weighted multi-source sensor data to obtain multi-source fusion features. In the deep feature extraction step, the prediction variance is used as a regularization constraint term for the sparse autoencoder feature extraction network to perform deep feature extraction on the multi-source fusion features. The larger the prediction variance, the stronger the regularization constraint, thus obtaining the deep features of the running state. The state recognition step involves inputting the deep features of the operating state into a pre-trained lightweight neural network state recognition model to obtain the real-time operating state recognition results of the electronic components and the confidence scores of the recognition results. The number of parameters in the lightweight neural network state recognition model is less than [a certain value]. ; In the fault attribution step, a gradient-based interpretability analysis method is used to perform backpropagation analysis on the decision-making process of the lightweight neural network state recognition model. The contribution of each channel data in the multi-source sensor detection data to the recognition result is calculated, fault attribution information is generated, and the sensor channels with a contribution exceeding a preset contribution threshold are marked as the main cause channels. The time domain intervals in the main cause channels with a contribution exceeding the preset contribution threshold are marked as sensitive time domain intervals. The main cause channels and the sensitive time domain intervals are fed back to the data acquisition step as dynamic feedback parameters to adjust the sampling frequency of the corresponding sensors in subsequent acquisitions to 2 to 10 times the original sampling frequency. The monitoring result output step outputs the real-time operating status identification result and the fault attribution information.

[0006] In one specific implementation of the first aspect, the operational life assessment model is a Bayesian neural network, and the prediction variance is obtained through Monte Carlo dropout during the inference process of the Bayesian neural network, specifically: performing Monte Carlo dropout on the same input data. Second forward propagation, The statistical variance of the T output results is calculated as the prediction variance.

[0007] In one specific implementation of the first aspect, in the multi-source fusion step, the fusion weights are allocated as follows: the sensor data sources are sorted in ascending order of their remaining lifetime prediction values, and the sensor data source with the smallest remaining lifetime prediction value receives the highest weight. The sensor data source with the largest remaining lifetime prediction value receives the lowest weight. The weights of the remaining sensor data sources are determined by linear interpolation based on the reciprocal of the predicted remaining lifetime, where... The range of the ratio is to .

[0008] In one specific implementation of the first aspect, in the deep feature extraction step, the objective function of the sparse autoencoder feature extraction network is:

[0009] in, To reconstruct the loss, For sparse penalty terms, Let Variance be the prediction variance. For network weights Norm, The sparsity penalty coefficient and The range of values ​​is to , Let be the uncertainty regularization coefficient, and let β take values ​​in the range of . to The and The prediction variance is determined on the validation set using either a grid search method or a Bayesian optimization method. The larger the regularization term, the more regularization terms are. The greater the constraint strength.

[0010] In one specific implementation of the first aspect, the number of parameters of the lightweight neural network state recognition model is less than 1M, and is obtained in the following way: A global neural network model is trained on a cloud server using a dataset of fault samples. The knowledge distillation method is employed, with the global neural network model serving as the teacher network, and the number of parameters being less than [a certain value]. The student network is the lightweight model to be trained. The distillation temperature is set to 3 to 8. The soft labels of the teacher network's output on unlabeled data are used as supervision signals to train the lightweight neural network state recognition model. The lightweight neural network state recognition model is deployed to an edge computing device.

[0011] In one specific implementation of the first aspect, the training of the lightweight neural network state recognition model includes a self-supervised pre-training phase and a supervised fine-tuning phase: In the self-supervised pre-training stage, a contrastive learning pre-training task is constructed using unlabeled electronic component operating status data to pre-train the encoder part of the lightweight neural network state recognition model and obtain pre-training weights. The unlabeled data includes normal operating status data and unlabeled operating status data, with a data volume of no less than 1000 sets. In the supervised fine-tuning stage, the pre-trained lightweight neural network state recognition model is fine-tuned with all parameters using labeled fault sample data. The number of labeled fault sample data is 5% to 30% of the unlabeled data.

[0012] In one specific implementation of the first aspect, a cloud-edge collaborative update step is also included: Each edge computing device uses locally collected operating status data to incrementally learn the lightweight neural network state recognition model deployed locally. The incremental learning adopts the elastic weight consolidation method to obtain local model update parameters. Calculate the local health index H and the generalization deviation index. The local health index H is determined based on the ratio of the predicted remaining lifespan of the current electronic components to their rated lifespan. , The range of values ​​is The generalization deviation index This represents the normalized Euclidean distance between the local model parameters and the global model parameters in the cloud. ; when Below the first preset threshold or When the value exceeds the second preset threshold, a model update upload operation is triggered to upload the local model update parameters to the cloud server. The first preset threshold is 0.3, and the second preset threshold is 0.15. The first preset threshold and the second preset threshold are determined according to the reliability requirements of the application scenario. The cloud server receives local model update parameters uploaded by each edge computing device, generates a new generation of global model parameters using a federated averaging algorithm, and distributes incremental update packages to each edge computing device. Each edge computing device receives the incremental update package and merges it with the local model parameters according to the fusion coefficient. Perform weighted fusion to complete the model update: ,in The value range is from 0.1 to 0.5.

[0013] In one specific implementation of the first aspect, an intelligent forgetting step is also included: Construct a forgetting mechanism based on a time decay function, for a collection time of... The effective weights of the historical operating status data used in the incremental learning are calculated as follows:

[0014] in, For data collection timestamps, This is the current timestamp. The attenuation coefficient is... The value range is from 0.001 to 0.1. The value is determined by cross-validation based on the time scale of the application scenario; When the effective weight When the value falls below a third preset threshold, the historical running status data is removed from the training set of the incremental learning. The third preset threshold is 0.05.

[0015] In one specific implementation of the first aspect, in the fault attribution step, the fault attribution information includes a heatmap of the contribution of each sensing channel and a mapping result of the physical location of the fault. The contribution heatmaps of each sensing channel are generated using the integral gradient method: the gradient of the output of the lightweight neural network state recognition model relative to the data of each input layer channel is calculated, and the path integral is performed along the time axis from the baseline to the current input to obtain the contribution score of each sensing channel at each time step. Sensing channels whose contribution scores exceed a preset contribution threshold are marked as primary cause channels, and time-domain intervals in the primary cause channels whose contribution scores exceed the preset contribution threshold are marked as sensitive time-domain intervals. The preset contribution threshold is 60% of the maximum contribution score. The fault physical location mapping result is obtained by matching the main cause channel and the sensitive time domain interval with a pre-established electronic component physical topology database, and outputs the physical location with the highest probability of fault occurrence and the corresponding failure mode. The physical topology database records the physical installation position of each sensor on the electronic component and the failure mode corresponding to each physical area. When the contribution score increases by more than a preset increase threshold in the next monitoring period after the sampling frequency is increased, an additional operation is triggered: the newly extracted high-frequency detail features are stored in the cloud sample library for subsequent special optimization training of the lightweight neural network state recognition model. The preset increase threshold is 20%.

[0016] Secondly, a real-time operating status monitoring system for electronic components includes: The multi-source sensing acquisition module includes voltage sensors, current sensors, temperature sensors, and frequency sensors deployed on electronic components. It is used to acquire multi-source sensing detection data, including at least voltage, current, temperature, and frequency data. The initial sampling frequency of each sensor is [missing information]. to ; The data preprocessing module is connected to the multi-source sensor acquisition module and is used to perform data cleaning, noise reduction and time alignment on the multi-source sensor detection data, and output the preprocessed multi-source sensor data. The operational life assessment module has a built-in operational life assessment model based on a Bayesian neural network, which is used to obtain the predicted remaining life of the current electronic components and the prediction variance of the predicted remaining life based on the preprocessed multi-source sensor data. The multi-source fusion module, connected to the lifetime assessment module, is used to assign fusion weights to each sensor data source based on the remaining lifetime prediction value, fuse the weighted multi-source sensor data through a multi-source data fusion network based on an attention mechanism, and output multi-source fusion features. The deep feature extraction module has a built-in sparse autoencoder feature extraction network, which is used to perform deep feature extraction on the multi-source fusion features with the prediction variance as a regularization constraint term, and output the running state deep features. The status recognition module, deployed on an edge computing device, has a built-in lightweight neural network status recognition model with less than 1M parameters. It is used to obtain the real-time operating status recognition result of electronic components and the confidence score of the recognition result based on the deep features of the operating status. The fault attribution module is used to perform backpropagation analysis on the decision-making process of the lightweight neural network state recognition model using the integral gradient method, calculate the contribution of each sensor data channel to the recognition result, generate a heatmap of the contribution of each sensor channel and a mapping result of the physical location of the fault, and send the sensor channels whose contribution exceeds the preset contribution threshold and their corresponding sensitive time domain intervals as dynamic feedback parameters to the multi-source sensor acquisition module so as to adjust the sampling frequency of the corresponding sensor to 2 to 10 times the original sampling frequency; The monitoring result output module is connected to the status identification module and the fault attribution module, and is used to output the real-time operating status identification result and the fault attribution information.

[0017] The beneficial effects of this invention are as follows: 1. This invention constructs a dynamic feedback closed loop driven by attribution information—breaking through the inherent paradigm of existing technologies where fault diagnosis results are only used for terminal display or offline analysis. It uses the main cause channel identifier and corresponding sensitive time domain interval marked by the contribution heatmap output from the integral gradient attribution step as dynamic feedback parameters, transmitting them back to the data acquisition layer in real time. This triggers an adaptive increase in the sampling frequency of the corresponding sensors within the sensitive time domain interval. This closed-loop mechanism realizes the reverse projection of diagnostic information to the perception layer, enabling dynamic matching between the resolution of data acquisition and the current fault information entropy. This ensures the effective capture of high-frequency detailed features within key time domain intervals while avoiding the waste of computational resources caused by continuous high sampling across all channels, maximizing the perception sensitivity of hidden fault features under limited resource constraints. When the increase in contribution obtained from secondary attribution after increasing the sampling frequency exceeds a preset threshold, the system automatically uploads the newly captured high-frequency detailed features to the cloud sample library for subsequent model-specific optimization iterations, forming a continuous optimization closed loop of "perception, diagnosis, feedback, and evolution." This dynamic feedback mechanism enables the monitoring system to leap from "passive acquisition" to "active perception," fundamentally solving the essential defect that key fault characteristics are easily limited by the Nyquist sampling theorem and cannot be captured under the traditional fixed sampling strategy; 2. This invention utilizes knowledge distillation technology to transfer the diagnostic capabilities of a large-capacity teacher network in the cloud to a student network with fewer than a certain threshold of parameters, forming a lightweight neural network model suitable for deployment on edge computing devices. This enables high-precision deep learning algorithms to achieve real-time inference in resource-constrained embedded environments, fundamentally solving the engineering challenge of traditional deep learning solutions being unable to be deployed at the edge due to their large model size. Simultaneously, this invention employs a self-supervised contrastive learning pre-training strategy, fully utilizing a large amount of unlabeled runtime data for encoder representation learning. Supervised fine-tuning requires only a small number of labeled fault samples, significantly reducing dependence on expensive fault labeling data and enhancing the method's engineering applicability in scenarios where labeled samples are scarce. Furthermore, this invention constructs a dynamic feedback closed loop driven by attribution information—the contribution heatmap output from the fault attribution step is not only used for result display, but its labeled main cause channel identifier and corresponding sensitive time domain interval are dynamically fed back to the data acquisition step, triggering an adaptive increase in the sampling frequency of the corresponding sensor within the sensitive time domain interval. This closed-loop mechanism enables dynamic matching between the resolution of data acquisition and the current fault information entropy. This ensures effective capture of high-frequency detailed features within key time domain intervals while avoiding the waste of computational resources caused by continuous high sampling across all channels. Under limited resource constraints, it maximizes the sensitivity to latent fault features. When the contribution increase from secondary attribution after increasing the sampling frequency exceeds a preset threshold, the system automatically uploads the newly captured high-frequency detailed features to the cloud sample library for subsequent model optimization iterations, forming a continuous optimization closed loop of "perception—diagnosis—feedback—evolution." Based on this, the invention adopts a cloud-edge collaborative update architecture. Each edge device achieves personalized model adaptation through local incremental learning. Simultaneously, a dual-trigger mechanism based on local health indicators and generalization deviation indicators enables selective updates from the edge to the cloud. The cloud server aggregates update parameters from each edge device using a federated averaging algorithm and distributes them to each edge device in incremental packets, ensuring continuous evolution of the global model while protecting the data privacy of each edge node. With the addition of an intelligent forgetting mechanism based on a time decay function, the system can proactively discard outdated data that exceeds the effective time window, effectively avoiding interference from historical data with current diagnoses and ensuring the continuous effectiveness of the monitoring model throughout its entire lifecycle. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0019] Figure 2 This is a schematic diagram of the multi-source fusion and deep feature extraction process of the present invention.

[0020] Figure 3 This is a schematic diagram of the fault attribution and dynamic feedback closed loop of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figures 1 to 3 The present invention relates to a method and system for real-time monitoring of the operating status of electronic components.

[0023] I. Overview of Technical Solution The present invention provides an artificial intelligence-based method for monitoring the operating status of electronic components, such as... Figure 1 As shown, the main steps include: The steps are as follows: data acquisition (S1), data preprocessing (S2), operational life assessment (S3), multi-source fusion (S4), deep feature extraction (S5), status identification (S6), fault attribution and dynamic feedback (S7), and monitoring result output (S8).

[0024] In this process, the fault attribution information output in step S7 is fed back to step S1 as a dynamic feedback parameter, forming a closed-loop adaptive monitoring architecture.

[0025] II. Specific implementation of each step.

[0026] Step S1: Data Acquisition.

[0027] (1) Sensor deployment.

[0028] Multiple types of sensors are deployed on the electronic components to be monitored. The specific selection and deployment methods of the sensors are as follows: Voltage sensor: Employs a high-impedance differential amplifier circuit to achieve non-invasive voltage detection. The voltage measurement range is determined based on the rated voltage of the component, typically ranging from 0.5 to 2 times the rated voltage, with a measurement accuracy of no less than [value missing]. (Full scale). For The module, with its voltage sensor deployed across the collector and emitter terminals, is used to collect the forward voltage drop. and turn-off voltage .

[0029] Current sensor: Utilizing a Hall effect current sensor or Rogowski coil for non-contact current detection, the current measurement range is 0.1 to 3 times the rated current, with a measurement accuracy of not less than [specified value]. .for The module, with its current sensor deployed in the main circuit, is used to collect the collector current. .

[0030] Temperature sensor: Utilizes a surface-mount negative temperature coefficient (STC) sensor. Thermistor or infrared temperature sensor, attached to the surface of the component housing or near the heat sink, has a temperature measurement range of [missing information]. to (Adjusted according to application scenario), measurement accuracy not less than .

[0031] Frequency sensor: Employs a frequency meter or a frequency measurement circuit based on zero-crossing detection, with a frequency measurement range of 0.5 to 5 times the operating frequency of the component, and a measurement accuracy of not less than ± reading.

[0032] Alternatively, the following sensors may also be deployed: Vibration sensor: A piezoelectric accelerometer is used, with a frequency response range of [range missing]. Up to 10 kHz, used to monitor abnormal mechanical vibrations.

[0033] Voiceprint sensor: Employs a microphone array with a frequency response range of [range missing]. to It is used to monitor abnormal sounds.

[0034] On-resistance detection circuit: This circuit uses a four-terminal method to measure the on-resistance of components in real time. or .

[0035] (2) Sampling frequency setting.

[0036] The initial sampling frequency of each sensor is determined based on the bandwidth of the monitored signal and the operating frequency of the components, as follows: ; In high-frequency ripple monitoring or transient overvoltage monitoring scenarios, the sampling frequency of voltage and current sensors can reach 1MHz. The sampling frequency setting follows the Nyquist sampling theorem, which states that the sampling frequency should not be less than twice the highest frequency component of the monitored signal.

[0037] (3) Data acquisition and caching.

[0038] Each sensor continuously collects data at its own sampling frequency. The collected data is accompanied by a precise timestamp (accuracy not less than 1ms) and is temporarily stored in a circular buffer of the edge computing device. The capacity of the circular buffer is determined based on the sampling frequency and a preset time window, and it can typically buffer at least 10 seconds of continuous data.

[0039] Step S2: Data preprocessing.

[0040] The data preprocessing module performs the following processing on the acquired multi-source sensor detection data: (1) Data cleaning.

[0041] Outliers that clearly exceed the sensor's measurement range and completely missing data points are removed. The criteria for determining outliers are: data points that exceed the upper or lower limit of the sensor's measurement range are directly marked as invalid.

[0042] For missing values, linear interpolation is used to fill them. Specifically, for the first... missing data points Utilizing its adjacent valid data points and Perform linear interpolation:

[0043] in, For the first The timestamp of each data point.

[0044] Data segments with more than 5 consecutive missing sampling points are discarded entirely without interpolation filling.

[0045] (2) Noise reduction processing.

[0046] Wavelet transform is used for noise reduction. The specific steps are as follows: ① Select wavelet basis functions, preferably using the db4, db6, or db8 wavelet basis from the Daubechies wavelet family. The db4 wavelet basis has good time-frequency localization characteristics and is suitable for preserving transient features in electronic component signals.

[0047] ② Determine the number of decomposition layers. Based on the signal sampling frequency and noise characteristics, the number of decomposition layers should be set to 3 to 5. Too few decomposition layers will result in insufficient noise reduction, while too many decomposition layers may damage the high-frequency components of the useful signal.

[0048] ③ The wavelet coefficients at each level are processed using a soft thresholding method. The threshold is selected using either the unbiased risk estimation (SURE) threshold or a general threshold. ,in This is an estimate of the noise standard deviation. This is the signal length.

[0049] ④ The processed wavelet coefficients are used to reconstruct the signal, resulting in a denoised signal.

[0050] (3) Timing alignment.

[0051] Because different sensors may have different sampling frequencies and start-up times, it is necessary to perform timing alignment on the data from each channel. The specific method is as follows: ①The channel with the highest sampling frequency among all sensors is used as the reference time axis.

[0052] ② For data from other channels, cubic spline interpolation is used to resample the data to each time point on the reference time axis.

[0053] ③ The time alignment accuracy is not lower than the time resolution corresponding to the highest sampling frequency among the sensors (i.e., , (This is the highest sampling frequency).

[0054] After the above preprocessing, the preprocessed multi-source sensor data is obtained. Its dimensions are ,in Number of sensor channels This represents the number of sampling points within the time window.

[0055] Step S3: Service life assessment.

[0056] (1) Structure of Bayesian neural network.

[0057] The operational life assessment module uses a life assessment model built on a Bayesian Neural Network (BNN).

[0058] The difference between Bayesian neural networks and traditional deterministic neural networks lies in the fact that the weight parameters of the Bayesian network are probability distributions rather than deterministic values. Specifically, a Gaussian prior distribution is assigned to the weights of each layer of the network. The posterior distribution of the weights is learned through variational inference. .

[0059] The specific structure of a Bayesian neural network is as follows: Input layer: Input dimension is ,in Number of sensor channels The time window length is (preferably 128 time steps).

[0060] Hidden Layer 1: Fully connected layer with 256 neurons, activation function is... The weights are prior to a Gaussian distribution. .

[0061] Hidden layer 2: Fully connected layer, with 128 neurons, and the activation function is... The weights are prior to a Gaussian distribution. .

[0062] Hidden layer 3: Fully connected layer, with 64 neurons and an activation function of . The weights are prior to a Gaussian distribution. .

[0063] Output layer: Fully connected layer with 1 neuron, linear activation function, outputting the predicted remaining lifetime. (Units are hours or number of operating cycles).

[0064] Add a discard layer after each hidden layer. The discard rate is set to 0.1 to 0.3.

[0065] (2) Model training.

[0066] The Bayesian neural network is trained using variational inference. The objective function is the Evidence Lower Bound (ELBO):

[0067] The first term is the likelihood expectation, which measures how well the model fits the training data; the second term is... Divergence, a measure of approximate posterior distribution With prior distribution The differences between them serve a regularization function.

[0068] Training was performed using the Adam optimizer, with a learning rate set to [value missing]. to Set the batch size to 32 to 128, the number of training epochs to 100 to 500, and the patience value for early stopping to 20 epochs.

[0069] The training dataset contains operational data of electronic components at different degradation stages (from healthy state to end of life) and their corresponding remaining lifetime labels. The dataset can be constructed in two ways: Method 1: Accelerated Life Testing. Taking IGBT modules as an example, typical conditions for accelerated life testing are: Temperature cycling test: ambient temperature at to The cycle is repeated at least 1000 times to simulate bond line fatigue caused by thermal cycling stress. Power cycling test: Apply 130% to 150% of the rated current for no less than 50,000 cycles to simulate solder layer fatigue caused by power cycling; High-temperature storage test: The gate oxide layer is continuously stored at a constant temperature of 150°C for no less than 1000 hours to simulate the degradation of the gate oxide layer caused by high-temperature aging.

[0070] During accelerated life testing, data from each sensor is continuously collected, and key performance parameters of components (such as on-state voltage drop, threshold voltage, thermal resistance, etc.) are measured periodically (e.g., every 50 hours) to calibrate the remaining life label.

[0071] Method 2: On-site operational data collection. In actual operating environments, operational data from multiple electronic components of the same model are collected over a long period of time. The cumulative operating time and final failure time of each component are recorded to construct an operational dataset with remaining lifespan labels.

[0072] (3) Reasoning and uncertainty quantification During the inference phase, the Bayesian neural network outputs a predicted remaining lifetime value for the current electronic components. and its prediction variance .

[0073] Predictive variance The acquisition method is Monte Carlo Dropout: the same input data is processed using Monte Carlo Dropout. Second forward propagation ( Preferred During each forward propagation, a portion of neurons are randomly deactivated (with the same dropout rate as during training), resulting in... Predicted remaining life value Calculate this The statistical variance of each predicted value is used as the prediction variance. :

[0074] in, for The mean of the predictions.

[0075] The physical meaning is: when When smaller (e.g.) This indicates that the lifetime predictions output by the model are highly consistent under different discard paths, and the evaluation results are highly reliable. when When it is large (e.g.) This indicates that the model's predictions for the current input data are significantly inconsistent, and the reliability of the evaluation results is low—this low reliability may stem from data noise, data distribution shifts, or components being in a highly nonlinear degradation stage.

[0076] Input to step S3: Preprocessed multi-source sensor data (Dimension:) ).

[0077] Output of step S3: Predicted remaining lifetime (Scalar, unit: hours or number of operating cycles) and its prediction variance (scalar, unit and) same).

[0078] Step S4: Multi-source fusion.

[0079] (1) Integration weight allocation.

[0080] The remaining lifetime prediction value output in step S3 Based on this, fusion weights are assigned to each sensor data source.

[0081] The allocation logic for fusion weights is as follows: sensor data sources with smaller remaining lifetime prediction values—that is, the physical parameters reflected by the sensor exhibit a more severe degradation trend—have higher weights in the fusion process. This is because the changes in various physical parameters of components nearing the end of their lifespan are more drastic and diverse, containing more degradation information.

[0082] The specific weight allocation method is as follows: ① Sort each sensor data source according to its predicted remaining lifetime value from smallest to largest to obtain a sorting number. ( , The total number of sensor data sources, The data source with the smallest predicted remaining lifespan.

[0083] ②The sensor data source with the smallest predicted remaining lifetime ( ) to obtain the highest weight The sensor data source with the largest remaining lifetime prediction ( Obtain the lowest weight .

[0084] ③ The weights of the remaining sensor data sources are determined according to the following linear interpolation formula:

[0085] in, The ratio ranges from 3:1 to 10:1, preferably 5:1.

[0086] ④ Normalize each weight: This makes the sum of all weights equal to 1.

[0087] (2) Multi-source fusion based on attention mechanism The normalized weights are applied to each sensor data source (i.e., each channel data is multiplied by its corresponding weight) to obtain the weighted multi-source sensor data. .

[0088] Will Input a multi-source data fusion network based on an attention mechanism. This fusion network employs a multi-head self-attention structure: Number of heads to focus on: 4 to 8 (4 is preferred).

[0089] Embedding dimension: determined based on the number of input channels, preferably 64 dimensions.

[0090] Feedforward network: Two fully connected layers, hidden layer dimension 256, activation function is ReLU.

[0091] The multi-head self-attention mechanism calculates the correlation between channels in the following way: Attention(Q,K,V)=softmax(Q·K^T / √d_k)·V Where Q, K, and V are the query, key, and value matrices, respectively, obtained by linearly transforming X_weighted; d_k is the dimension of the key vector.

[0092] The output of multi-head attention is a weighted sum of the features of each channel, which enables the fusion network to adaptively focus on the sensor data channels and time domain intervals that are most valuable for judging the current state.

[0093] The inputs for step S4 are: preprocessed multi-source sensor data X_preprocessed (dimension: N_channels×T) and remaining lifetime prediction value RUL_pred.

[0094] The output of step S4: multi-source fusion feature F_fusion (dimension: d_fusion×T, where d_fusion is the dimension of the fusion feature, preferably 64 dimensions).

[0095] Step S5: Deep feature extraction.

[0096] (1) Sparse autoencoder structure.

[0097] The deep feature extraction module employs a sparse autoencoder network, consisting of three parts: an encoder, a bottleneck layer, and a decoder. Encoder: Input layer (d_fusion dimension), fully connected layer (128-dimensional, ReLU), fully connected layer (64-dimensional, ReLU), bottleneck layer (32-dimensional, linear).

[0098] Bottleneck layer: 32-dimensional, outputs deep features of the running state. .

[0099] Decoder: Bottleneck layer (32-dimensional), fully connected layer (64-dimensional, ReLU), fully connected layer (128-dimensional, ReLU), output layer (d_fusion dimension, linear), reconstructing the original input.

[0100] (2) Objective function.

[0101] The training objective function of the sparse autoencoder is:

[0102] in: To measure the reconstruction loss, the mean squared error (MSE) of the original input is used. With decoder reconstruction output Differences between them:

[0103] As a sparsity penalty term, adopt Divergence measures the average activation rate of neurons in the bottleneck layer. Sparsity with Target Differences between them:

[0104] Among them, target sparsity The value range is from 0.01 to 0.1 (preferably 0.05). For the first The average activation rate of bottleneck neurons on the training set.

[0105] The sparsity penalty coefficient ranges from 0.01 to 0.5 (preferably 0.1) and is used to control the weight of the sparsity penalty term.

[0106] For network weights The Frobenius norm serves as a regularization mechanism for weight decay.

[0107] This is the uncertainty regularization coefficient, with a value ranging from 0.01 to 1.0 (preferably 0.1), used to control the weight of the uncertainty regularization term.

[0108] This is the prediction variance output in step S3.

[0109] (3) Uncertainty-driven dynamic regularization.

[0110] The core innovation of this objective function lies in: when the prediction variance... When the value is large, the regularization term The constraint strength automatically increases, forcing the encoder weights to shrink towards zero, thereby suppressing feature overfitting in the high uncertainty region; when When the value is small, the regularization constraint is weakened, allowing the encoder to fully learn the detailed features in the data.

[0111] and The specific value is determined on the validation set using either grid search or Bayesian optimization. The candidate value range for grid search is: , Choose the parameter combination that minimizes the reconstruction error on the validation set.

[0112] Input for step S5: Multi-source fusion features (Dimension:) ) and prediction variance .

[0113] Output of step S5: Running state depth features (Dimensions: 32)

[0114] Step S6: Status recognition.

[0115] (1) Lightweight neural network model structure.

[0116] The state recognition module uses a lightweight neural network model with fewer than 100 parameters. (1 million).

[0117] Preferably, the lightweight neural network adopts a depthwise separable convolution architecture, with the following specific structure: Input layer: Receives deep features of the running state. (32 dimensions).

[0118] Depthwise separable convolution block 1: Depthwise convolution (3×3, 32 channels), batch normalization, ReLU, pointwise convolution (1×1, 64 channels), batch normalization, ReLU.

[0119] Depthwise separable convolutional block 2: depthwise convolution (3×3, 64 channels), batch normalization, ReLU, pointwise convolution (1×1, 128 channels), batch normalization, ReLU.

[0120] Global average pooling: compresses the spatiotemporal dimension to 1 dimension.

[0121] Fully connected layer: 128-dimensional, 64-dimensional (ReLU), output layer (K-dimensional, Softmax), where K is the number of state categories (such as normal, warning, fault type 1, fault type 2, etc.).

[0122] (2) Control of model parameters.

[0123] The number of model parameters was kept below 1M through the following measures: By replacing standard convolution with depthwise separable convolution, the number of parameters is reduced by about 1 / 8 to 1 / 10. Control the number of convolutional kernels (the number of channels should not exceed 128); The number of neurons in a fully connected layer does not exceed 256.

[0124] Calculations show that the number of parameters for the above structure is approximately 0.5M to 0.9M, which meets the requirement that the number of parameters be less than 1M.

[0125] (3) Model training - knowledge distillation.

[0126] The lightweight neural network state recognition model is obtained through knowledge distillation: Step 1: Online training for teachers.

[0127] A high-precision global neural network model (teacher network) is trained on a cloud server using a dataset of fault samples. The teacher network can employ a ResNet50 or Transformer architecture with a large number of parameters, ranging from 10M to 100M, to ensure diagnostic accuracy.

[0128] The training dataset for the teacher network contains labeled fault samples of various fault modes, with a data volume of no less than 10,000 sets, covering normal operation and various fault states.

[0129] Step 2: Knowledge distillation.

[0130] Using a teacher network as the teacher and a student network with fewer than 1M parameters as the lightweight model to be trained, a knowledge distillation method is employed for training.

[0131] Distillation temperature The value is set to 3 to 8 (ideally 5). The role of the temperature parameter T is to soften the output probability distribution of the teacher network, enabling the student network to better learn the knowledge from the teacher network.

[0132] The distillation loss function is:

[0133] in: Use hard-label loss (cross-entropy loss) with real labels; For soft label loss (KL divergence loss), the teacher network is used at temperature. The output soft tags below; The weight for soft label loss is 0.3 to 0.7 (preferably 0.5). This is a temperature scaling factor used to compensate for the gradient shrinkage caused by temperature increases.

[0134] Step 3: Model Deployment.

[0135] The trained lightweight neural network state recognition model is quantized (FP32 to INT8 or FP16) and then deployed to edge computing devices. After quantization, the model size can be further compressed to 1 / 4 to 1 / 2 of the original size, and the inference speed is increased by 2 to 4 times.

[0136] (4) Self-supervised pre-training (preferred implementation method) To further reduce reliance on labeled fault data, the training of the lightweight neural network state recognition model preferably includes a self-supervised pre-training stage and a supervised fine-tuning stage: Self-supervised pre-training phase: Utilizing a large amount of unlabeled electronic component operating status data (no less than 1000 sets), a contrastive learning pre-training task (such as SimCLR or MoCo framework) is constructed to pre-train the encoder part of the lightweight model. Contrastive learning narrows the representation distance of positive sample pairs (different time segments of the same component) and widens the representation distance of negative sample pairs (different components or different operating conditions), enabling the encoder to learn good data representation capabilities without labels.

[0137] Supervised fine-tuning phase: Using labeled fault sample data (5% to 30% of the number of unlabeled data, preferably 10%), the pre-trained lightweight model is fine-tuned with all parameters.

[0138] Self-supervised pre-training can significantly reduce the reliance on labeled faulty data, reducing the amount of labeled data required by more than 70%.

[0139] (5) Reasoning and confidence scores.

[0140] During the inference phase, the deep features of the running state are... Input a lightweight neural network state recognition model, and the model outputs the probability distribution of each state category. .

[0141] The real-time running status identification result is the category with the highest probability: .

[0142] The confidence score is the maximum probability value: value range The confidence score reflects the model's certainty about the current recognition result—the higher the confidence score, the more reliable the recognition result.

[0143] Input for step S6: Running state deep features (Dimensions: 32)

[0144] Output of step S6: Real-time running status identification result (Category label) and confidence score (scalar, ).

[0145] Step S7: Fault attribution and dynamic feedback.

[0146] (1) Contribution calculation - integral gradient method.

[0147] The integrated gradient method is used to analyze the interpretability of the decision-making process of a lightweight neural network state recognition model.

[0148] The basic principle of the integral gradient method is to calculate the gradient of the model output relative to each input feature, integrate along the path from the baseline to the current input, and obtain the contribution of each feature to the model output.

[0149] The specific calculation steps are as follows: ① Determine the baseline input The baseline input is typically chosen as a "no information" input, such as all features being 0 or all features being the mean of the training set.

[0150] ②At the baseline and current input Perform linear interpolation between them to generate Interpolation points ( Preferred ):

[0151] ③ Calculate the gradient of the model output f relative to the input at each interpolation point. .

[0152] ④ Integrate the gradient at each interpolation point (numerically approximating by summation and averaging) to obtain the contribution score of each feature:

[0153] in, For the first The contribution score of each input feature. For the currently entered number 1 eigenvalue, The first input for the baseline Each feature value.

[0154] (2) Generation of contribution heatmap.

[0155] Score the contribution Arranged according to time sequence and channel dimension, a two-dimensional contribution heatmap is generated. Indicates the first Each sensor channel at time step The contribution score.

[0156] Sensor channels whose contribution scores exceed a preset contribution threshold are marked as primary cause channels. The preset contribution threshold is 60% of the maximum contribution score.

[0157] The time-domain intervals in the main factor channel whose contribution scores exceed the threshold are marked as sensitive time-domain intervals.

[0158] For example, if the temperature channel is within the time window If the contribution score exceeds the threshold, it indicates that the temperature anomaly during that time period is the main driving factor for the current fault determination.

[0159] (3) Fault physical location mapping.

[0160] The main channel and sensitive time domain interval are matched with a pre-established database of physical topology structures of electronic components.

[0161] The physical topology database is constructed as follows: ① Sensor physical location recording: Record the physical installation coordinates of each sensor on the electronic component. Taking the IGBT module as an example, record the chip number corresponding to each temperature sensor, the terminal position corresponding to the voltage sensor, and the loop position corresponding to the current sensor.

[0162] ② Failure Mode Library Establishment: Through fault injection experiments, the characteristic responses of each sensor channel under various failure modes are obtained. The specific method for fault injection experiments is as follows: Bond wire fatigue: The bond wires were gradually fatigued through a power cycling test (130% of the rated current was applied for 50,000 cycles), and the cycle-by-cycle variation trend of the on-state voltage drop V_CE was recorded. Solder layer cracking: Fatigue cracks were induced in the solder layer through temperature cycling tests (-40°C to +125°C, 1000 cycles), while the trends of thermal resistance and case temperature were recorded. Gate oxide degradation: The gate oxide layer was gradually degraded by a high-temperature gate bias test (150°C, 120% of the rated voltage applied to the gate, for 500 hours), while the changing trends of gate leakage current and threshold voltage were recorded.

[0163] Through the above fault injection experiments, the characteristic response vectors of each sensing channel under various failure modes are extracted to form a "failure mode-characteristic response" mapping table.

[0164] ③ Matching Algorithm: In real-time monitoring, the main cause channel and contribution distribution vector V_actual marked in step S7(2) are matched with the feature response vector V_template of each failure mode stored in the database. The matching distance uses weighted Euclidean distance:

[0165] in, The distance between the current data and the j-th failure mode is given. For the first The weights of each channel (obtained by normalizing the contribution score). For the first The actual contribution score of each channel For the first The first failure mode Template feature response of each channel.

[0166] The failure mode with the smallest matching distance is selected as the matching result, and the physical location with the highest probability of failure and the corresponding failure mode are output. When the minimum matching distance exceeds a preset threshold (preferably twice the minimum distance between all templates in the database), it is determined to be an "unknown failure mode," triggering the cloud-based expert system to intervene and analyze.

[0167] (4) Dynamic feedback adjustment.

[0168] The sensor channel identifier marked as the main cause channel and its corresponding sensitive time domain interval are used as dynamic feedback parameters and fed back to the data acquisition step S1.

[0169] The specific feedback mechanism is as follows: ① When a certain sensing channel is marked as the main channel, the sampling frequency of the sensor corresponding to that channel in subsequent acquisitions is increased to 2 to 10 times the original sampling frequency (determined according to the relative contribution score of the main channel; the higher the contribution, the greater the increase), in order to capture the higher frequency detailed features that the channel may have in the sensitive time domain interval.

[0170] For example, if the voltage channel is identified as the primary cause channel in this diagnosis and its contribution score is 90% of the maximum contribution score, the sampling frequency of the voltage sensor is increased from 10kHz to 50kHz (5 times).

[0171] ② Sensors not marked as the primary channel should maintain their original sampling frequency or be appropriately downsampled (to 50% to 100% of the original sampling frequency) to save storage and computing resources of edge computing devices.

[0172] ③ The dynamic feedback takes effect in the next monitoring cycle. Step S7 recalculates the contribution score in the next cycle based on the high-frequency data collected after increasing the sampling rate.

[0173] ④ If the contribution score obtained from the second calculation increases by more than the previous contribution score by more than the preset increase threshold (preferably 20%), it is determined that there are hidden fault features that are masked by the low sampling rate, and an additional operation is triggered: the high-frequency detail features that are re-extracted are stored in the cloud sample library for subsequent special optimization training of the lightweight state recognition model.

[0174] Input for step S7: Running state deep features Lightweight neural network state recognition model, current input data.

[0175] The output of step S7 is: fault attribution information (main cause channel identifier, sensitive time domain interval, fault physical location, failure mode) and dynamic feedback parameters (returned to step S1).

[0176] Step S8: Output monitoring results.

[0177] The monitoring results output steps output the following information: Real-time operating status identification results (normal / warning / fault type); Confidence score; Predicted remaining useful life and its prediction variance; Fault attribution information (primary cause channel, sensitive time domain interval, physical location of the fault, failure mode); Repair recommendations (generated based on fault type and physical location).

[0178] Output formats include, but are not limited to: A visual dashboard display on a local screen or a remote monitoring platform; Audible and visual alarm (triggered when the identification result indicates a fault); Structured fault report generation (including timestamp, component identification, fault description, attribution analysis, etc.); Repair suggestions are pushed to maintenance personnel via SMS, email, or APP.

[0179] III. Cloud-edge collaborative updates.

[0180] To achieve continuous evolution and personalized adaptation of the model, this invention preferably adopts a cloud-edge collaborative architecture for online model updates.

[0181] (1) Local incremental learning Each edge computing device uses locally collected real-time operational status data to incrementally learn the lightweight neural network state recognition model deployed locally.

[0182] Incremental learning employs the Elastic Weight Consolidation (EWC) method, which applies additional constraints to important parameters to prevent catastrophic forgetting while optimizing the loss of new data. The loss function is:

[0183] in, For the loss on new data, for The first information matrix diagonal elements (measured parameters) (importance) For the old model parameters, for Regularization coefficient (range 0.1 to 10).

[0184] (2) Dual-trigger upload mechanism.

[0185] Calculate the local health index H and the generalization deviation index. : Local health index The predicted remaining useful life based on the output of step S3. With rated life The ratio is determined as follows:

[0186] The range of values ​​is , The closer This indicates that the healthier the components, The closer This indicates that the more severe the degradation of the components, the better.

[0187] Generalization deviation index : for local model parameters With cloud-based global model parameters Normalized Euclidean distance between them:

[0188] This reflects the degree of difference between local data distribution and global data distribution. The larger the value, the greater the difference between local operating conditions and cloud-based training conditions.

[0189] The triggering condition is: when An update is triggered when the value falls below a first preset threshold (preferably 0.3) – indicating that the component has entered a severe degradation stage and requires cloud intervention for refined diagnosis; when An update is triggered when the value exceeds the second preset threshold (preferably 0.15) – indicating a significant change in local operating conditions that requires personalized adaptation in the cloud.

[0190] The first and second preset thresholds are determined based on the reliability requirements of the application scenario. For scenarios with high reliability requirements (such as aerospace), the first preset threshold can be appropriately increased to 0.4, and the second preset threshold can be appropriately decreased to 0.10.

[0191] When any trigger condition is met, the edge device uploads the local model update parameters obtained through incremental learning to the cloud server.

[0192] (3) Cloud aggregation.

[0193] The cloud server receives local model update parameters uploaded by each edge device and uses a federated averaging (FPA) algorithm to weight and aggregate the update parameters:

[0194] in, The number of edge devices participating in the aggregation. For the first Local data volume of an edge device The total amount of data across all edge devices. For the first Local model parameters uploaded by an edge device.

[0195] The federated aggregation process does not access the raw data of each edge device, thus protecting data privacy.

[0196] (4) Incremental issuance.

[0197] The cloud server calculates the incremental update package of the next-generation global model parameters relative to the previous generation global model parameters:

[0198] Incremental update package The data is distributed to each edge device. The size of the incremental packet is much smaller than that of the full model (typically 5% to 20% of the full model size), effectively saving communication bandwidth.

[0199] (5) Local integration.

[0200] After receiving the incremental update packet, each edge device performs a weighted fusion with the local model parameters:

[0201] in, The fusion coefficient has a value ranging from 0.1 to 0.5 (preferably 0.3). The smaller the size, the more local personalized features are retained. The larger the size, the more global knowledge is integrated.

[0202] IV. Intelligent forgetting.

[0203] A forgetting mechanism based on a time decay function is constructed to selectively forget old running state data that exceeds a preset time window.

[0204] Specifically, for historical operational status data collected at time t, the effective weights for incremental learning are calculated according to the following time decay function:

[0205] in: For data collection timestamps; This is the current timestamp; This is the attenuation coefficient, in seconds. The value ranges from 0.001 to 0.1.

[0206] The value of is determined using cross-validation based on the time scale of the application scenario: For rapidly changing scenarios (such as high-frequency switching power supplies). The value can be 0.01 to 0.1 (half-life is approximately 7 to 69 seconds); For slow degradation scenarios (such as long-running industrial equipment). It can take values ​​from 0.001 to 0.005 (half-life of approximately 139 to 693 seconds).

[0207] When effective weights When the value is below the third preset threshold (preferably 0.05), the historical operating status data is judged as "outdated data" and removed from the training set for incremental learning.

[0208] V. System Architecture.

[0209] This invention also provides an artificial intelligence-based electronic component operating status monitoring system, comprising the following modules, each corresponding to steps S1 to S8 of the above method: 1. Multi-source sensor acquisition module.

[0210] It includes voltage sensors, current sensors, temperature sensors, and frequency sensors deployed on electronic components, used to collect multi-source sensing and detection data including at least voltage data, current data, temperature data, and frequency data, with each sensor having an initial sampling frequency of 1Hz to 1MHz.

[0211] The module also has the function of receiving dynamic feedback parameters: when a feedback signal (including the main cause channel identifier and sensitive time domain interval) is received from the fault attribution module, the sampling frequency of the corresponding sensor is adjusted to 2 to 10 times the original sampling frequency.

[0212] 2. Data Preprocessing Module It connects to the multi-source sensor acquisition module to perform data cleaning, noise reduction, and time-series alignment on the multi-source sensor detection data, and outputs the pre-processed multi-source sensor data.

[0213] Data cleaning uses linear interpolation to fill missing values; noise reduction uses wavelet transform with the Daubechies wavelet family (db4 to db8) as the wavelet basis, 3 to 5 decomposition levels, and soft thresholding; time alignment uses cubic spline interpolation.

[0214] 3. Service life assessment module It incorporates a lifetime assessment model based on a Bayesian neural network, which is used to obtain the predicted remaining lifetime value RUL_pred of the current electronic components and the prediction variance of the predicted value based on preprocessed multi-source sensor data. .

[0215] The structure of a Bayesian neural network is as follows: input layer, fully connected layer (256-dimensional, ReLU, Dropout), fully connected layer (128-dimensional, ReLU, Dropout), fully connected layer (64-dimensional, ReLU, Dropout), and output layer (1-dimensional, linear). Prediction variance. Obtained through the Monte Carlo discarding method. .

[0216] 4. Multi-source fusion module.

[0217] Connected to the lifetime assessment module, it is used to assign fusion weights to each sensor data source based on the remaining lifetime prediction value RUL_pred. It then fuses the weighted multi-source sensor data through an attention-based multi-source data fusion network, outputting multi-source fusion features. .

[0218] The fusion weight allocation method is as follows: the sensor data source with the lowest remaining lifetime prediction value receives the highest weight. The sensor data source with the largest remaining lifetime prediction value receives the lowest weight. The weights of the remaining data sources are determined by linear interpolation. The ratio is 3:1 to 10:1. The attention mechanism is a multi-head self-attention structure with 4 to 8 attention heads.

[0219] 5. Deep feature extraction module.

[0220] Built-in sparse autoencoder feature extraction network for predicting variance As a regularization constraint term, it affects the multi-source fusion feature. Perform deep feature extraction and output the deep features of the running state. .

[0221] The bottleneck layer of the sparse autoencoder has a dimension of 32; the objective function is... ,in The value range is from 0.01 to 0.5. The value range is from 0.01 to 1.0.

[0222] 6. Status recognition module.

[0223] Deployed on edge computing devices, it incorporates a lightweight neural network state recognition model with fewer than 1M parameters, used to identify deep features of the operating state. Obtain real-time operating status identification results of electronic components And the confidence score of the identification result.

[0224] The lightweight neural network employs a deep separable convolutional architecture and is trained using a knowledge distillation method. The range is 3 to 8.

[0225] 7. Fault Attribution Module.

[0226] This method is used to perform backpropagation analysis on the decision-making process of a lightweight neural network state recognition model using the integral gradient method, calculate the contribution of each sensor data channel to the recognition result, and generate a heatmap of the contribution of each sensor channel and a mapping result of the physical location of the fault.

[0227] This module marks the sensor channels whose contribution exceeds the preset contribution threshold (60% of the maximum contribution score) as the main cause channels, and sends the main cause channels and their corresponding sensitive time domain intervals as dynamic feedback parameters to the multi-source sensor acquisition module, so as to adjust the sampling frequency of the corresponding sensors to 2 to 10 times the original sampling frequency.

[0228] When the contribution score increases by more than the preset increase threshold (20%) in the next monitoring period after the sampling frequency is increased, the module triggers an additional operation: the newly extracted high-frequency detail features are stored in the cloud sample library for subsequent special optimization training of the lightweight neural network state recognition model.

[0229] 8. Monitoring result output module.

[0230] It connects to the status recognition module and the fault attribution module to output real-time operating status recognition results and fault attribution information. Output formats include visual dashboard displays, audible and visual alarms, structured fault report generation, and maintenance suggestion push notifications.

[0231] The output of the multi-source sensor acquisition module is connected to the input of the data preprocessing module; the output of the data preprocessing module is connected to the input of the service life assessment module and the multi-source fusion module respectively; the output of the service life assessment module is connected to the input of the multi-source fusion module and the deep feature extraction module; the output of the multi-source fusion module is connected to the input of the deep feature extraction module; the output of the deep feature extraction module is connected to the input of the status recognition module; the output of the status recognition module is connected to the input of the fault attribution module and the monitoring result output module; the output of the fault attribution module is connected to the input of the monitoring result output module and sends the dynamic feedback parameters to the multi-source sensor acquisition module.

[0232] Example

[0233] Example 1: Status monitoring of IGBT power modules in new energy vehicles.

[0234] (1) Application scenarios This embodiment uses an IGBT power module (model: Infineon FS820R08A6P2B, rated voltage 750V, rated current 820A) in a new energy vehicle motor driver as the monitoring object. This IGBT module is used in the motor drive system of a certain brand of pure electric vehicle, with an operating ambient temperature range of -40°C to +85°C and a switching frequency of 10kHz.

[0235] (2) Sensor deployment Deploy the following sensors on the IGBT module: Voltage sensor: Deployed at the collector-emitter terminals of the IGBT module, using a high-impedance differential amplifier circuit, with a measurement range of 0V to 1000V, an accuracy of ±0.5%FS, and an initial sampling frequency of 10kHz.

[0236] Current sensor: A Hall effect current sensor (model: LEM HAS 800-S) is used, deployed in the main circuit, with a measurement range of -1200A to +1200A, an accuracy of ±1%FS, and an initial sampling frequency of 10kHz.

[0237] Temperature sensor: A surface-mount NTC thermistor (model: EPCOS B57891S) is attached to the surface of the IGBT module housing (6 temperature measurement points, corresponding to 6 IGBT chip positions), with a measurement range of -40°C to +150°C, an accuracy of ±1°C, and an initial sampling frequency of 5Hz.

[0238] Vibration sensor: A piezoelectric accelerometer (model: PCB 352C33) is used, which is mounted on the module heat sink. The frequency response range is 0.5Hz to 10kHz, and the initial sampling frequency is 10kHz.

[0239] (3) Construction of training dataset Accelerated life tests were conducted on IGBT modules of the same model under laboratory conditions to construct a training dataset: ① Temperature Cycling Test: The IGBT module is placed in a temperature cycling test chamber and cycled between -40°C and +125°C, with a heating rate of 10°C / min and a cooling rate of 5°C / min, a residence time of 30 minutes for each cycle, and 1000 cycles in total. After every 50 cycles, the module is removed, and key parameters such as on-state voltage drop V_CE, threshold voltage V_GE(th), and thermal resistance R_th are measured, and the current remaining lifetime is recorded.

[0240] ② Power Cycling Test: Apply 130% of the rated current (approximately 1066A) to the IGBT module, with a turn-on time of 2 seconds and a turn-off time of 3 seconds, for a total of 50,000 cycles. After every 5,000 cycles, measure key parameters and record the remaining lifespan label.

[0241] ③ High-temperature storage test: The IGBT module was placed in a 150°C constant temperature chamber and stored continuously for 1000 hours. Key parameters were measured every 100 hours.

[0242] Through the above accelerated life test, a total of about 20,000 sets of operational data with remaining life tags were obtained, including about 10,000 sets of normal operation data and about 10,000 sets of various fault status data.

[0243] (4) System deployment The edge computing device uses an ARM Cortex-M7 core microcontroller (STM32H743, 400MHz clock speed, 2MB Flash and 1MB SRAM) to deploy a lightweight neural network state recognition model (0.8M parameters, model size approximately 0.8MB after INT8 quantization). The cloud server uses a GPU server (NVIDIA A100) for teacher network training and global model aggregation.

[0244] (5) Monitoring process Initial stage (health status): In the initial stage of IGBT module operation (cumulative operating time 0 to 500 hours), each sensor collects data at the initial sampling frequency. After preprocessing in step S2, the Bayesian neural network in step S3 outputs the predicted remaining lifetime value. hourly, prediction variance (Lower) indicates high reliability of the evaluation results. In step S4, the predicted remaining lifetime values ​​of each sensor data source are similar, and the fusion weights are close to an equal weight allocation. In step S5, because The data is relatively small, with weak regularization constraints, allowing the feature extraction network to fully learn data details. Step S6 outputs a "normal" state with a confidence score of 0.96. In the fault attribution analysis in step S7, the contribution of each channel is evenly distributed, no primary cause channel is marked, and dynamic feedback does not trigger sampling frequency adjustment.

[0245] Degradation stage (bond wire fatigue): When the IGBT module had been running for about 3000 hours, due to long-term thermal cycling stress, the bonding wire of the No. 2 IGBT chip experienced progressive fatigue, and the on-resistance R_CE(on) gradually increased from the initial 1.2mΩ to 1.5mΩ.

[0246] Step S3 detected a significant decrease in the predicted remaining lifetime of the temperature channel (measuring point 2 on chip 2). (hours), prediction variance Rise to (Higher) indicates that the model has significant disagreement with the current evaluation results. Step S4 increases the fusion weight of the temperature channel from equal weight to the highest weight. This makes temperature information dominate the fusion features. In step S5, because... Larger, stronger regularization constraints ( (The term increases), and the encoder weights shrink towards zero, suppressing feature overfitting in high uncertainty regions.

[0247] Step S6 outputs the "bonding wire aging" status with a confidence score of 0.82.

[0248] In step S7, the integral gradient attribution analysis identifies the temperature channel (measurement point of chip 2) and the on-resistance channel as the main cause channels, with contribution scores of 85% and 72% of the maximum contribution scores, respectively. The sensitive time domain interval is the on-state interval of each switching cycle (approximately 50 μs). The fault physical location mapping result is: the emitter bond line of IGBT chip 2, with the failure mode being bond line peeling caused by thermal fatigue.

[0249] The dynamic feedback parameters are fed back to step S1: the sampling frequency of the temperature sensor of chip No. 2 is increased from 5Hz to 50Hz (10 times), the sampling frequency of the voltage sensor is increased from 10kHz to 50kHz (5 times), and the sampling frequency of the current sensor is kept unchanged at 10kHz.

[0250] Results after feedback: In the next monitoring cycle after increasing the sampling rate, the temperature data sampled at 50Hz captured the transient temperature spikes during switching (approximately +8°C / switching cycle) that were indistinguishable at the original 5Hz sampling rate, and the voltage data sampled at 50kHz captured the minute increments of the on-state voltage drop (approximately 5mV / cycle) and its cycle-by-cycle fluctuation characteristics.

[0251] Secondary attribution analysis confirmed the specific location of bond line fatigue (emitter bond line of chip 2), significantly improving the attribution depth. The secondary calculation value of the contribution score increased by approximately 28% compared to the previous one, exceeding the 20% improvement threshold, triggering an additional operation: storing high-frequency detailed features in a cloud sample library for subsequent model-specific optimization training.

[0252] Meanwhile, sensors not marked as the primary channel (such as other temperature measurement points) maintain their original sampling frequency, and the computing resources of edge devices are dynamically reallocated to the data processing of the primary channel.

[0253] (6) Cloud-edge collaborative update During the six months of operation of this IGBT module, the local health index H gradually decreased from an initial 0.95 to 0.35. When H dropped below 0.3 (in the fifth month), a model update upload operation was triggered, and the edge device uploaded the model update parameters obtained from local incremental learning to the cloud server.

[0254] The cloud server receives updated parameters from IGBT monitoring edge devices across multiple vehicles, generates a new generation global model using a federated averaging algorithm, and distributes incremental update packets to each edge device. Each edge device then performs a weighted fusion of the incremental packets with its local model. ), and complete the model update.

[0255] The updated model improved the accuracy of bond line fatigue detection from 92% to 96%, and reduced the false alarm rate from 5% to 2%.

[0256] (7) Monitoring results The system outputs the following monitoring results in real time: Status identification: Bond wire aging (confidence levels 0.82, 0.91); Remaining lifetime prediction: approximately 4200 hours (prediction variance 0.06·RUL_nominal) 2 ); Fault attribution: Thermal fatigue of the emitter bond wire of IGBT chip No. 2; Maintenance Recommendation: It is recommended to schedule a shutdown for maintenance within 2000 hours to replace the IGBT module.

[0257] (8) Results data After six months of field operation and verification, the monitoring system of this embodiment achieved the following results: Fault detection rate: 98.5% (78.3% using the traditional threshold method); False alarm rate: 1.8% (12.5% ​​with the traditional threshold method); The false negative rate was 1.5% (compared to 21.7% using the traditional threshold method). Single inference time: 32ms (meets the 50ms real-time requirement); Edge device memory usage: 1.2MB (including model and data buffers).

[0258] Example 2: Application in status monitoring of industrial control chips.

[0259] (1) Application scenarios This embodiment uses the main control chip (model: NXP i.MX RT1060, Cortex-M7 core, operating frequency 600MHz) in an industrial PLC (Programmable Logic Controller) as the monitoring object. This chip is used in the PLC controller of an automated production line, and the operating environment temperature range is 0°C to +60°C.

[0260] (2) Sensor deployment and parameter settings A voltage sensor (measuring core voltage VDD_SOC and I / O voltage VDD_IO, initial sampling frequency 10kHz), a current sensor (measuring total current, initial sampling frequency 10kHz), a temperature sensor (attached to the chip surface, initial sampling frequency 5Hz), and a frequency sensor (measuring external clock frequency, initial sampling frequency 1kHz) are deployed around the chip.

[0261] The lightweight neural network state recognition model is trained using a self-supervised pre-training method: after pre-training with 10,000 sets of unlabeled running data, it is fine-tuned with only 200 sets of labeled fault data (i.e., labeled data is 2% of unlabeled data), which achieves the same recognition accuracy as fully supervised training (requiring 2,000 sets of labeled data) (97.2% vs 97.8%), reducing the need for labeled data by 90%.

[0262] (3) Monitoring results During the 12 months of system operation, a total of 3 abnormal states were successfully identified: Month 3: "Core voltage ripple anomaly" was identified (confidence level 0.89), which was attributed to the aging of the power supply filter capacitor. A warning was issued 2 weeks in advance, which prevented the program from crashing due to unstable power supply. Month 7: "Chip overheating warning" was identified (confidence level 0.92), attributed to dust accumulation on the cooling fan leading to decreased heat dissipation efficiency, triggering a repair work order for cleaning; Month 11: "Clock frequency drift" was identified (confidence level 0.85), attributed to external crystal oscillator aging, and a warning was issued 1 month in advance.

[0263] During the 12 months of cloud-edge collaborative operation, four model updates were triggered (twice for H and twice for D). Even after chip aging, the model maintained a diagnostic accuracy rate of over 95%. The intelligent forgetting mechanism gradually downweighted historical data older than 30 days to below 5%, ensuring that the model always adapts to the current chip degradation stage and workload conditions.

[0264] (4) Results data Early warning lead time for abnormal events: 14 days on average; Number of unplanned outages: reduced from an average of 4 times per year to 1 time; Maintenance costs: reduced by approximately 60%.

[0265] Example 3: Applied to status monitoring of radio frequency components in communication base stations.

[0266] (1) Application scenarios This embodiment focuses on a power amplifier (PA, model: Qorvo QPA9901, operating frequency band 3.3GHz to 4.2GHz, output power 40W) in a 5G communication base station as the monitoring object. This PA is used in the radio frequency front-end module of a certain operator's 5G base station.

[0267] (2) Sensor deployment and parameter settings A voltage sensor (measuring drain voltage V_DD, initial sampling frequency 100kHz), a current sensor (measuring drain current I_DD, initial sampling frequency 100kHz), a temperature sensor (attached near the PA transistor, initial sampling frequency 10Hz), and a frequency sensor (measuring the input RF signal frequency, initial sampling frequency 10kHz) are deployed on the PA module.

[0268] (3) Construction of physical topology database For the physical topology of this PA module, a physical topology database is constructed: ① Sensor physical location recording: The temperature sensor is mounted on the heat dissipation substrate surface of the PA transistor (Q1), about 2mm away from the chip; the voltage sensor is connected to the drain power supply terminal; the current sensor is connected in series in the drain power supply circuit.

[0269] ② Failure Mode Library Establishment: Obtain the characteristic responses of various failure modes through fault injection experiments: Output matching capacitor aging: The output matching capacitor was aged through a high-temperature storage test (85°C, 500 hours), and the changes in S-parameters and drain current were recorded. Transistor thermal degradation: Changes in gain and drain efficiency were recorded through a power cycling test (120% of rated power applied for 10,000 cycles). Bias circuit drift: The drift of the static operating point was recorded through a long-term aging test (2000 hours of operation under rated conditions).

[0270] (4) Monitoring results Under multi-carrier excitation, the nonlinear distortion of the PA intensifies with increasing temperature. After the system has been running for approximately 2000 hours, the fault attribution module in step S7 maps the contribution heatmap to the physical topology of the PA, successfully locating the capacitor aging failure in the output matching network with a location accuracy of over 90% of the physical area.

[0271] The dynamic feedback mechanism increased the temperature sensor sampling frequency from 10Hz to 100Hz, capturing transient temperature overshoots (peak temperatures about 12°C higher than average) that were indistinguishable at the original sampling rate, confirming that thermal stress is the main factor causing capacitor aging.

[0272] Repair time was reduced from an average of 4 hours to 1.5 hours (a reduction of 62.5%). Fault feature capture rate improved by approximately 35%, and the false negative rate dropped from 8% initially to below 2%.

[0273] (5) Results data Fault location accuracy: over 90% of physical areas; Repair time: reduced from 4 hours to 1.5 hours; False alarm rate: reduced from 8% to 2%.

[0274] This invention can be widely applied to various scenarios for monitoring the operational status of electronic components, including but not limited to: New energy vehicles: IGBT and MOSFET power modules in motor drive systems, power devices in on-board chargers, and monitoring chips in battery management systems; Industrial automation: power modules in frequency converters and servo drives, main control chips in PLCs, and controller chips in industrial robots; Smart grids: IGBT modules in converter valves, power devices in SVG (Static Var Generator), and power modules in energy storage converters; Aerospace: key electronic components in avionics systems, requiring redundant monitoring scenarios with high reliability; Communication equipment: radio frequency components (power amplifiers, low-noise amplifiers) in 5G base stations, and power devices in base station power supplies; Consumer electronics: CPUs and GPUs in high-end servers, and power modules in data center power supplies.

[0275] This invention is applicable to both online monitoring of newly built systems and intelligent upgrades of existing systems. The core algorithm can run offline on edge computing devices or be integrated with a cloud platform to form a complete monitoring solution. The modular design of the system allows for flexible configuration of sensor types and quantities, and adjustment of model parameters and thresholds according to specific application scenarios, exhibiting excellent scalability and adaptability.

[0276] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for real-time monitoring of the operating status of electronic components, comprising a data acquisition step, a data preprocessing step, and a monitoring result output step, wherein the data acquisition step acquires multi-source sensing data through multiple sensors deployed on the electronic components, the multi-source sensing data including at least voltage data, current data, temperature data, and frequency data, and the initial sampling frequency of each sensor is [missing information]. to The data preprocessing step involves cleaning, denoising, and temporal alignment of the multi-source sensor detection data to obtain preprocessed multi-source sensor data. The characteristic of this step is that... Also includes: The operational life assessment step involves inputting the preprocessed multi-source sensor data into an operational life assessment model constructed based on a Bayesian neural network to obtain the predicted remaining life of the current electronic components and the prediction variance of the predicted remaining life. The prediction variance is used to quantify the uncertainty of the output result of the operational life assessment model. In the multi-source fusion step, based on the remaining lifetime prediction value, fusion weights are assigned to each sensor data source. The sensor data source with the smaller remaining lifetime prediction value has a higher weight in the fusion. A multi-source data fusion network based on an attention mechanism is used to fuse the weighted multi-source sensor data to obtain multi-source fusion features. In the deep feature extraction step, the prediction variance is used as a regularization constraint term for the sparse autoencoder feature extraction network to perform deep feature extraction on the multi-source fusion features. The larger the prediction variance, the stronger the regularization constraint, thus obtaining the deep features of the running state. The state recognition step involves inputting the deep features of the operating state into a pre-trained lightweight neural network state recognition model to obtain the real-time operating state recognition results of the electronic components and the confidence scores of the recognition results. The number of parameters in the lightweight neural network state recognition model is less than [a certain value]. ; In the fault attribution step, a gradient-based interpretability analysis method is used to perform backpropagation analysis on the decision-making process of the lightweight neural network state recognition model. The contribution of each channel data in the multi-source sensor detection data to the recognition result is calculated, fault attribution information is generated, and the sensor channels with a contribution exceeding a preset contribution threshold are marked as the main cause channels. The time domain intervals in the main cause channels with a contribution exceeding the preset contribution threshold are marked as sensitive time domain intervals. The main cause channels and the sensitive time domain intervals are fed back to the data acquisition step as dynamic feedback parameters to adjust the sampling frequency of the corresponding sensors in subsequent acquisitions to 2 to 10 times the original sampling frequency. The monitoring result output step outputs the real-time operating status identification result and the fault attribution information.

2. The method for real-time monitoring of the operating status of electronic components according to claim 1, characterized in that, The operational life assessment model is a Bayesian neural network, and the prediction variance is obtained through Monte Carlo dropout during the inference process of the Bayesian neural network. Specifically, this involves performing Monte Carlo dropout on the same input data. Second forward propagation, The statistical variance of the T output results is calculated as the prediction variance.

3. The method for real-time monitoring of the operating status of electronic components according to claim 1, characterized in that, In the multi-source fusion step, the fusion weights are allocated as follows: the sensor data sources are sorted from smallest to largest according to their remaining lifetime prediction values, and the sensor data source with the smallest remaining lifetime prediction value receives the highest weight. The sensor data source with the largest remaining lifetime prediction value receives the lowest weight. The weights of the remaining sensor data sources are determined by linear interpolation based on the reciprocal of the predicted remaining lifetime, where... The range of the ratio is to .

4. The method for real-time monitoring of the operating status of electronic components according to claim 1, characterized in that, In the deep feature extraction step, the objective function of the sparse autoencoder feature extraction network is: in, To reconstruct the loss, For sparse penalty terms, Let Variance be the prediction variance. For network weights Norm, The sparsity penalty coefficient and The range of values ​​is to , Let be the uncertainty regularization coefficient, and let β take values ​​in the range of . to The and The prediction variance is determined on the validation set using either a grid search method or a Bayesian optimization method. The larger the regularization term, the more regularization terms are. The greater the constraint strength.

5. The method for real-time monitoring of the operating status of electronic components according to claim 1, characterized in that, The lightweight neural network state recognition model has fewer than 1M parameters, which are obtained through the following methods: A global neural network model is trained on a cloud server using a dataset of fault samples. The knowledge distillation method is employed, with the global neural network model serving as the teacher network, and the number of parameters being less than [a certain value]. The student network is the lightweight model to be trained. The distillation temperature is set to 3 to 8. The soft labels of the teacher network's output on unlabeled data are used as supervision signals to train the lightweight neural network state recognition model. The lightweight neural network state recognition model is deployed to an edge computing device.

6. The method for real-time monitoring of the operating status of electronic components according to claim 5, characterized in that, The training of the lightweight neural network state recognition model includes a self-supervised pre-training stage and a supervised fine-tuning stage: In the self-supervised pre-training stage, a contrastive learning pre-training task is constructed using unlabeled electronic component operating status data to pre-train the encoder part of the lightweight neural network state recognition model and obtain pre-training weights. The unlabeled data includes normal operating status data and unlabeled operating status data, with a data volume of no less than 1000 sets. In the supervised fine-tuning stage, the pre-trained lightweight neural network state recognition model is fine-tuned with all parameters using labeled fault sample data. The number of labeled fault sample data is 5% to 30% of the unlabeled data.

7. The method for real-time monitoring of the operating status of electronic components according to claim 5, characterized in that, It also includes cloud-edge collaborative update steps: Each edge computing device uses locally collected operating status data to incrementally learn the lightweight neural network state recognition model deployed locally. The incremental learning adopts the elastic weight consolidation method to obtain local model update parameters. Calculate the local health index H and the generalization deviation index. The local health index H is determined based on the ratio of the predicted remaining lifespan of the current electronic components to their rated lifespan. , The range of values ​​is The generalization deviation index This represents the normalized Euclidean distance between the local model parameters and the global model parameters in the cloud. ; when Below the first preset threshold or When the value exceeds the second preset threshold, a model update upload operation is triggered to upload the local model update parameters to the cloud server. The first preset threshold is 0.3, and the second preset threshold is 0.

15. The first preset threshold and the second preset threshold are determined according to the reliability requirements of the application scenario. The cloud server receives local model update parameters uploaded by each edge computing device, generates a new generation of global model parameters using a federated averaging algorithm, and distributes incremental update packages to each edge computing device. Each edge computing device receives the incremental update package and merges it with the local model parameters according to the fusion coefficient. Perform weighted fusion to complete the model update: ,in The value range is from 0.1 to 0.

5.

8. The method for real-time monitoring of the operating status of electronic components according to claim 7, characterized in that, It also includes the intelligent forgetting step: Construct a forgetting mechanism based on a time decay function, for a collection time of... The effective weights of the historical operating status data used in the incremental learning are calculated as follows: in, For data collection timestamps, This is the current timestamp. The attenuation coefficient is... The value range is from 0.001 to 0.

1. The value is determined by cross-validation based on the time scale of the application scenario; When the effective weight When the value falls below a third preset threshold, the historical running status data is removed from the training set of the incremental learning. The third preset threshold is 0.

05.

9. The method for real-time monitoring of the operating status of electronic components according to claim 1, characterized in that, In the fault attribution step, the fault attribution information includes a heatmap of the contribution of each sensing channel and a mapping result of the physical location of the fault. The contribution heatmaps of each sensing channel are generated using the integral gradient method: the gradient of the output of the lightweight neural network state recognition model relative to the data of each input layer channel is calculated, and the path integral is performed along the time axis from the baseline to the current input to obtain the contribution score of each sensing channel at each time step. Sensing channels whose contribution scores exceed a preset contribution threshold are marked as primary cause channels, and time-domain intervals in the primary cause channels whose contribution scores exceed the preset contribution threshold are marked as sensitive time-domain intervals. The preset contribution threshold is 60% of the maximum contribution score. The fault physical location mapping result is obtained by matching the main cause channel and the sensitive time domain interval with a pre-established electronic component physical topology database, and outputs the physical location with the highest probability of fault occurrence and the corresponding failure mode. The physical topology database records the physical installation position of each sensor on the electronic component and the failure mode corresponding to each physical area. When the contribution score increases by more than a preset increase threshold in the next monitoring period after the sampling frequency is increased, an additional operation is triggered: the newly extracted high-frequency detail features are stored in the cloud sample library for subsequent special optimization training of the lightweight neural network state recognition model. The preset increase threshold is 20%.

10. A real-time operating status monitoring system for electronic components, characterized in that, include: The multi-source sensing acquisition module includes voltage sensors, current sensors, temperature sensors, and frequency sensors deployed on electronic components. It is used to acquire multi-source sensing detection data, including at least voltage, current, temperature, and frequency data. The initial sampling frequency of each sensor is [missing information]. to ; The data preprocessing module is connected to the multi-source sensor acquisition module and is used to perform data cleaning, noise reduction and time alignment on the multi-source sensor detection data, and output the preprocessed multi-source sensor data. The operational life assessment module has a built-in operational life assessment model based on a Bayesian neural network, which is used to obtain the predicted remaining life of the current electronic components and the prediction variance of the predicted remaining life based on the preprocessed multi-source sensor data. The multi-source fusion module, connected to the lifetime assessment module, is used to assign fusion weights to each sensor data source based on the remaining lifetime prediction value, fuse the weighted multi-source sensor data through a multi-source data fusion network based on an attention mechanism, and output multi-source fusion features. The deep feature extraction module has a built-in sparse autoencoder feature extraction network, which is used to perform deep feature extraction on the multi-source fusion features with the prediction variance as a regularization constraint term, and output the running state deep features. The status recognition module, deployed on an edge computing device, has a built-in lightweight neural network status recognition model with less than 1M parameters. It is used to obtain the real-time operating status recognition result of electronic components and the confidence score of the recognition result based on the deep features of the operating status. The fault attribution module is used to perform backpropagation analysis on the decision-making process of the lightweight neural network state recognition model using the integral gradient method, calculate the contribution of each sensor data channel to the recognition result, generate a heatmap of the contribution of each sensor channel and a mapping result of the physical location of the fault, and send the sensor channels whose contribution exceeds the preset contribution threshold and their corresponding sensitive time domain intervals as dynamic feedback parameters to the multi-source sensor acquisition module so as to adjust the sampling frequency of the corresponding sensor to 2 to 10 times the original sampling frequency; The monitoring result output module is connected to the status identification module and the fault attribution module, and is used to output the real-time operating status identification result and the fault attribution information.

Images