Integrated circuit burn-in station primary power self-healing control method and apparatus

Abstract

The embodiment of the application provides a kind of integrated circuit aging platform level power supply self-healing control method and device, its method includes: in the process of aging test, the monitoring signal of integrated circuit aging platform level power supply is acquired in real time, and the monitoring signal is preprocessed, to obtain the monitoring signal after preprocessing;Parallel multi-channel convolution long short time memory network PMCCNN-ALSTM model of fusion attention mechanism is constructed and trained, and the remaining useful life of the integrated circuit aging platform level power supply at the current time is obtained by inputting the monitoring signal after preprocessing into the PMCCNN-ALSTM model after training;Based on the remaining useful life of integrated circuit aging platform level power supply at the current time, the health score result at the current time is obtained;The health score result at the current time is used to carry out adaptive pulse width modulation PWM hierarchical self-healing control to the integrated circuit aging platform level power supply.

Description

Technical Field

[0001] This application relates to the field of integrated circuit aging bench technology, and in particular to a method and apparatus for self-healing control of the primary power supply of an integrated circuit aging bench. Background Technology

[0002] Aging test benches simulate various real-world usage environments and working conditions, subjecting products to changes in condition similar to those experienced after prolonged use within a short period. Aging tests typically require the test equipment to be powered on for extended periods to simulate the aging process of a product in actual use. The primary power supply, serving as the core of the aging test bench's power input, provides the basic electrical energy to the entire system (including the secondary power supply, signal generation module, etc.), and its stability directly determines whether the test can continue.

[0003] During the component aging test, the primary power supply is prone to hidden faults such as output voltage drift, increased ripple, and decreased conversion efficiency due to issues such as input grid fluctuations, aging of internal switching devices, and performance degradation of filter capacitors. If not intervened in time, it may lead to abnormal input of the secondary power supply, triggering overcurrent, undervoltage and other protection actions, or even causing test interruption. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method for self-healing control of the primary power supply of an integrated circuit aging bench. One or more embodiments of this application also relate to a self-healing control device for the primary power supply of an integrated circuit aging bench, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this application, a method for self-healing control of the primary power supply of an integrated circuit aging bench is provided, comprising: During the aging test, the monitoring signal of the primary power supply of the integrated circuit aging bench is acquired in real time, and the monitoring signal is preprocessed to obtain the preprocessed monitoring signal. A parallel multi-channel convolutional long short-term memory network PMCCNN-ALSTM model with an integrated attention mechanism is constructed and trained. The remaining service life of the primary power supply of the integrated circuit aging bench at the current moment is obtained by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model. Based on the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment, obtain the health score result at the current moment; The current health score result is used to perform adaptive pulse width modulation (PWM) hierarchical self-healing control on the primary power supply of the integrated circuit aging bench.

[0006] Preferably, the monitoring signal of the primary power supply of the integrated circuit aging bench includes at least the input voltage, output voltage, output current, forward voltage drop of the output rectifier diode, and on-state voltage drop of the main switch.

[0007] Preferably, the PMCCNN-ALSTM model includes a parallel multi-channel convolutional neural network (PMCCNN) unit composed of multiple parallel convolutional channels, an attention-based feature fusion unit, and a long short-term memory (LSTM) unit. The PMCCNN unit extracts frequency domain signals from the preprocessed monitoring signal at different times using each convolutional channel, and concatenates these signals in chronological order to form a temporal feature with a time dimension. The attention-based feature fusion unit performs deep spatiotemporal fusion processing on the temporal feature with a time dimension to obtain a context vector containing global key information at the current time. The LSTM unit uses the context vector containing global key information at the current time and the hidden state from the previous time step to obtain the hidden state at the current time, and uses this hidden state as the remaining lifespan of the primary power supply of the integrated circuit aging test bench at the current time.

[0008] Preferably, by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model, the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment is obtained, including: The frequency domain signal at each time step of the preprocessed monitoring signal is processed by one-dimensional convolutional feature mapping using the convolutional kernels in the PMCCNN unit of the trained PMCCNN-ALSTM model, so as to obtain the first output result of each feature map of each convolutional layer at each time step. By introducing the ReLU function into the PMCCNN unit, the first output result is subjected to non-linear activation processing to obtain the activation output result of each feature map of each convolutional layer at each time step; The activation output is downsampled and dimensionality reduced by the PMCCNN unit to obtain the low-dimensional deep features of each feature map at each time step. The low-dimensional deep features of all time steps are then concatenated in chronological order to form a temporal feature with a time dimension.

[0009] Preferably, obtaining the remaining lifespan of the primary power supply of the integrated circuit aging bench at the current moment by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model further includes: By utilizing the attention-based feature fusion unit in the trained PMCCNN-ALSTM model, the temporal features with the time dimension are sequentially subjected to spatial adaptive channel feature fusion and temporal attention mechanism processing to obtain the context vector containing global key information at the current moment.

[0010] Preferably, obtaining the remaining lifespan of the primary power supply of the integrated circuit aging bench at the current moment by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model further includes: The LSTM unit in the trained PMCCNN-ALSTM model is used to perform deep temporal modeling on the context vector containing global key information at the current moment to obtain the hidden state at the current moment, and the hidden state at the current moment is used as the remaining service life of the first-level power supply of the integrated circuit aging bench at the current moment.

[0011] Preferably, the adaptive pulse width modulation (PWM) hierarchical self-healing control of the primary power supply of the integrated circuit aging bench using the current health score result includes: Using the current health score and the remaining lifetime, determine the switching frequency for adaptive PWM. and duty cycle ; Using the switching frequency and duty cycle Adaptive PWM hierarchical self-healing control is applied to the primary power supply of the integrated circuit aging bench. Wherein, the switching frequency include:

[0012] Among them, the For the rated frequency, the To adjust the coefficient, the For health score results; The duty cycle include:

[0013]

[0014]

[0015]

[0016] Among them, the The deviation between the output voltage and the target voltage, the For the integration process The deviation between the output voltage and the target voltage at the corresponding moment; To initially set the duty cycle; the The proportionality coefficient; The integral coefficient; The differential coefficients; The ratio benchmark value under healthy conditions; The preset proportional gain adjustment coefficient; The ratio benchmark value under healthy conditions; The preset proportional gain adjustment coefficient; The differential benchmark value under healthy conditions; This is the preset proportional gain adjustment coefficient.

[0017] According to a second aspect of the embodiments of this application, a self-healing control device for the primary power supply of an integrated circuit aging bench is provided, comprising: The acquisition and preprocessing module is configured to acquire the monitoring signal of the primary power supply of the integrated circuit aging bench in real time during the aging test, and obtain the preprocessed monitoring signal by preprocessing the monitoring signal. The model building and training module is configured to build and train a parallel multi-channel convolutional long short-term memory network PMCCNN-ALSTM model with an attention mechanism, and obtain the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model. The acquisition module is configured to acquire the health score result at the current moment based on the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment. The self-healing control module is configured to perform adaptive pulse width modulation (PWM) hierarchical self-healing control on the primary power supply of the integrated circuit aging bench using the current health score result.

[0018] According to a third aspect of the embodiments of this application, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement any of the steps of the integrated circuit aging bench primary power supply self-healing control method.

[0019] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of any one of the integrated circuit aging bench primary power supply self-healing control methods.

[0020] According to a fifth aspect of the present application, a computer program is provided, wherein when the computer program is executed in a computer, the computer is instructed to perform the steps of the above-described integrated circuit aging bench primary power supply self-healing control method.

[0021] The self-healing control scheme for the primary power supply of an integrated circuit aging test bench provided in this application acquires monitoring signals of the primary power supply in real time during the aging test. These signals are preprocessed to obtain preprocessed monitoring signals. A parallel multi-channel convolutional long short-term memory (PMCCNN-ALSTM) model with an attention mechanism is constructed and trained. The preprocessed monitoring signals are input into the trained PMCCNN-ALSTM model to obtain the remaining lifespan of the primary power supply at the current moment. Based on the remaining lifespan of the primary power supply, a health score is obtained at the current moment. The health score is then used to perform adaptive pulse width modulation (PWM) hierarchical self-healing control on the primary power supply. This allows for the early detection of potential internal problems in the power supply, proactive compensation before performance degradation, prevention of test interruptions, and ensuring the smooth progress of the aging test as planned. Attached Figure Description

[0022] Figure 1 This is a flowchart of a self-healing control method for the primary power supply of an integrated circuit aging bench, provided in one embodiment of this application. Figure 2 This is an overall flowchart of a self-healing control method for the primary power supply of an integrated circuit aging bench, provided in one embodiment of this application. Figure 3 This is a structural diagram of a PMCNN unit provided in one embodiment of this application; Figure 4 This is a schematic diagram of a primary power supply self-healing control device for an integrated circuit aging bench, provided in one embodiment of this application. Detailed Implementation

[0023] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0024] This application provides a method for self-healing control of the primary power supply of an integrated circuit aging bench. This application also relates to a device for self-healing control of the primary power supply of an integrated circuit aging bench, a computing device, a computer-readable storage medium, and a computer program, which will be described in detail in the following embodiments.

[0025] Figure 1 This application provides a flowchart of a self-healing control method for the primary power supply of an integrated circuit aging bench, which includes the following steps: Step S101: During the aging test, the monitoring signal of the primary power supply of the integrated circuit aging bench is acquired in real time, and the monitoring signal is preprocessed to obtain the preprocessed monitoring signal. In one embodiment of this application, the monitoring signals of the primary power supply of the integrated circuit aging bench include at least the input voltage, output voltage, output current, forward voltage drop of the output rectifier diode, and on-state voltage drop of the main switch. During the aging test, the primary power supply of the high-temperature aging bench generates monitoring signals such as voltage and current. Since the directly measured monitoring signal data is large and contains a lot of redundant information, feature extraction is first performed on the directly measured data to obtain key operating parameters including input voltage, output voltage, output current, forward voltage drop of the output rectifier diode, on-state voltage drop of the main switch, and threshold voltage, resulting in a high-information-density feature sequence.

[0026] Step S102: Construct and train a parallel multi-channel convolutional long short-term memory network PMCCNN-ALSTM model with an attention mechanism, and obtain the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model. In one embodiment of this application, the PMCCNN-ALSTM model includes a parallel multi-channel convolutional neural network (PMCCNN) unit composed of multiple parallel convolutional channels, an attention-based feature fusion unit, and a long short-term memory (LSTM) unit. The PMCCNN unit is used to extract frequency domain signals at different times from the preprocessed monitoring signal through each convolutional channel, and to concatenate the frequency domain signals at different times into a temporal feature with a time dimension in chronological order. The attention-based feature fusion unit is used to perform deep spatiotemporal fusion processing on the temporal feature with a time dimension to obtain a context vector containing global key information at the current time. The LSTM unit is used to obtain the hidden state at the current time using the context vector containing global key information at the current time and the hidden state at the previous time, and to use the hidden state at the current time as the remaining service life of the primary power supply of the integrated circuit aging bench at the current time.

[0027] In one embodiment of this application, the remaining lifespan of the primary power supply of the integrated circuit aging bench at the current moment is obtained by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model, including: The frequency domain signal at each time step of the preprocessed monitoring signal is processed by one-dimensional convolutional feature mapping using the convolutional kernels in the PMCCNN unit of the trained PMCCNN-ALSTM model, so as to obtain the first output result of each feature map of each convolutional layer at each time step. By introducing the ReLU function into the PMCCNN unit, the first output result is subjected to non-linear activation processing to obtain the activation output result of each feature map of each convolutional layer at each time step; The activation output is downsampled and dimensionality reduced by the PMCCNN unit to obtain the low-dimensional deep features of each feature map at each time step after downsampling. The low-dimensional deep features of all time steps are then concatenated in chronological order to form a temporal feature with a time dimension. The attention-based feature fusion unit in the trained PMCCNN-ALSTM model is used to sequentially perform spatial adaptive channel feature fusion and temporal attention mechanism processing on the temporal features with the time dimension, so as to obtain the context vector containing global key information at the current time. The LSTM unit in the trained PMCCNN-ALSTM model is used to perform deep temporal modeling on the context vector containing global key information at the current moment to obtain the hidden state at the current moment, and the hidden state at the current moment is used as the remaining service life of the first-level power supply of the integrated circuit aging bench at the current moment.

[0028] This application introduces a parallel multi-channel convolutional long short-term memory network (PMCCNN-ALSTM) model with an integrated attention mechanism to model and predict the degradation trend of key operating parameters in step S101, thereby achieving accurate estimation of power supply performance degradation state. Based on extracting local features of heterogeneous parameters using a multi-channel convolutional structure, this model automatically focuses on and weights high-contribution degradation parameters and aging abrupt change moments through an adaptive feature fusion layer and a temporal attention mechanism, solving the problem of long sequence information forgetting. Simultaneously, it innovatively introduces the ArcFace loss function for joint optimization during the training phase, forcing the model to maximize the inter-class distance between different health states in the feature space. This significantly improves the robustness and discriminative power supply performance degradation state estimation while achieving high-precision remaining lifetime prediction.

[0029] Step S103: Based on the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment, obtain the health score result at the current moment; Step S104: Use the current health score result to perform adaptive pulse width modulation (PWM) hierarchical self-healing control on the primary power supply of the integrated circuit aging bench.

[0030] In one embodiment of this application, the adaptive pulse width modulation (PWM) hierarchical self-healing control of the primary power supply of the integrated circuit aging bench using the current health score result includes: Using the current health score and the remaining lifetime, determine the switching frequency for adaptive PWM. and duty cycle ; Using the switching frequency and duty cycle Adaptive PWM hierarchical self-healing control is applied to the primary power supply of the integrated circuit aging bench. Wherein, the switching frequency include:

[0031] Among them, the For the rated frequency, the To adjust the coefficient, the For health score results; The duty cycle include:

[0032]

[0033]

[0034]

[0035] Among them, the The deviation between the output voltage and the target voltage, the For the integration process The deviation between the output voltage and the target voltage at the corresponding moment; To initially set the duty cycle; the The proportionality coefficient; The integral coefficient; The differential coefficients; The ratio benchmark value under healthy conditions; The preset proportional gain adjustment coefficient; The ratio benchmark value under healthy conditions; The preset proportional gain adjustment coefficient; The differential benchmark value under healthy conditions; This is the preset proportional gain adjustment coefficient.

[0036] This application provides forward-looking data support for subsequent adaptive PWM regulation by accurately predicting parameter degradation trends. Based on the prediction results, an adaptive PWM regulation strategy is designed to dynamically adjust the duty cycle and switching frequency, enabling the power supply to actively distribute voltage and current stress at different degradation stages, thereby achieving protection of key components and self-healing control of system performance.

[0037] Figure 2 This is an overall flowchart of a self-healing control method for the primary power supply of an integrated circuit aging bench, provided in one embodiment of this application. Figure 2 As shown, it includes: Step 1: Monitor the primary power supply signals of the aging test bench (input voltage, output voltage, output current, forward voltage drop of the output rectifier diode, on-state voltage drop of the main switch, etc.). During the aging test, sensors are used to monitor the input voltage, output voltage, output current, forward voltage drop of the output rectifier diode, on-state voltage drop of the main switch, and threshold voltage of the primary power supply of the aging test bench. Large-scale monitoring data is obtained for the status assessment of the power supply system.

[0038] Step two involves data preprocessing operations such as denoising, normalization, and fast Fourier transform. Given the prevalence of environmental noise interference and significant differences in the dimensions of various physical quantities in the raw collected data, the following data preprocessing strategies were employed to construct high-quality model input: Firstly, denoising was performed to eliminate random interference in the data; secondly, normalization was implemented to ensure that parameters of different magnitudes were within a uniform comparison range. Simultaneously, methods such as Fast Fourier Transform were used to transform the time-domain data to the frequency domain, extracting frequency domain characteristics and providing richer information for subsequent analysis.

[0039] Step 3: Parameter degradation prediction based on PMCCNN-ALSTM; Considering that data from previous moments can influence predictions for later moments, and that typical CNNs can only extract features from data samples collected at a specific moment, resulting in a lack of temporal information for those features, this application employs a Parallel Multi-Channel Convolution Attention Long Short-Term Memory Network (PMCCNN-ALSTM) model that integrates an attention mechanism and the ArcFace loss function to effectively utilize the temporal information hidden in the signal. This model first uses a parallel feature extraction unit based on PMCCNN to capture the local temporal features of various heterogeneous parameters. Then, through adaptive fusion of multi-source features and a temporal attention mechanism, it enhances its ability to focus on key degradation moments. Subsequently, it performs LSTM-based temporal modeling to capture the long-term dependencies of parameter degradation. Finally, it implements joint training incorporating the ArcFace loss function to enhance the model's accuracy in distinguishing different degradation states.

[0040] 3.1 Parallel Feature Extraction Based on PMCCNN To address the challenge of traditional single-channel convolutional networks simultaneously handling the characteristics of heterogeneous parameters from multiple sources and the temporal information of long sequences, this embodiment constructs a parallel multi-channel convolutional neural network (PMCCNN) unit. For example... Figure 3 As shown, this unit consists of multiple parallel convolutional channels. Each convolutional channel extracts frequency domain signals at different times, and these signals are concatenated sequentially to form a temporal feature with a time dimension. This feature is then used as the input for the next stage of the model's computation.

[0041] Specifically, assuming the first The input frequency domain signal of each channel is This unit completes the mapping from raw high-dimensional data to low-dimensional deep features through the following three core computational steps: 1) One-dimensional convolutional feature mapping: Extracting local temporal patterns by sliding convolutional kernels along the time axis. The first convolutional layer Each feature map in Output at time Defined as:

[0042] in, This represents the convolution operation. For convolution kernel weights, This is a bias term.

[0043] 2) Nonlinear activation: The ReLU activation function is introduced to enhance the model's ability to express nonlinear aging trends by filtering positive signals and effectively suppressing background noise. The first convolutional layer Each feature map in Activation output at any moment Defined as:

[0044] in, This indicates a positive value operation.

[0045] 3) Downsampling dimensionality reduction: To reduce the feature dimensionality and thus decrease computational cost while preserving salient features within local regions, max pooling is employed. Let the pooling window size and stride be... Then the downsampled first Each feature map in Output at time Defined as:

[0046] in, This is the sliding offset within the pooling window, with a value ranging from 0 to... ; This indicates the corresponding index position of the input feature on the time axis.

[0047] 3.2 Feature Fusion Module Based on Attention Mechanism To overcome the shortcomings of traditional feature concatenation methods, such as difficulty in distinguishing the importance of heterogeneous parameters and inability to focus on key degradation moments, this embodiment constructs a feature fusion module based on an attention mechanism between the convolutional layer and the recurrent layer. This module does not directly input multi-channel features into the LSTM, but instead achieves accurate extraction of key degradation features and deep spatiotemporal fusion through two layers: adaptive channel fusion and temporal attention mechanism.

[0048] First, in the spatial dimension, adaptive channel feature fusion is implemented. Considering that different physical parameters (such as switching voltage drop and input voltage) have significantly different characterization capabilities for power supply aging, the model introduces learnable channel weight parameters. This parameter is used to analyze the deep features output by each convolutional channel. Adaptive weighted concatenation is performed, enabling the model to automatically amplify the weights of parameters with high aging correlation, suppress interference from insensitive parameters, and generate a unified fused feature sequence. :

[0049] in,

[0050] Building upon this, a temporal attention mechanism is introduced in the time dimension. Addressing the issue that degradation abrupt changes in long-sequence monitoring data are easily diluted by mundane historical information, this mechanism dynamically measures the contribution of features at different times to the current degradation state by calculating the attention score at each time step. Let... For fusion sequence exist The feature vectors at each time step are first used to calculate the nonlinear attention score using a fully connected layer and a hyperbolic tangent activation function. Subsequently, the scores are normalized into probability weights using the Softmax function. Finally, the feature sequences are weighted and summed using this weight to generate a context vector containing global key information. This vector highly condenses the core degradation information of the power source within the current window, providing high-quality input features for subsequent time-series modeling. Its mathematical expression is as follows:

[0051] in, , , These are all parameters that are automatically optimized during network training. This represents the length of the time window.

[0052] 3.3 LSTM-based timing modeling The context vector enhanced by the attention mechanism is then fed into a Long Short-Term Memory (LSTM) network for deep temporal modeling. Leveraging its unique gating structure, the LSTM unit effectively overcomes the vanishing gradient problem faced by traditional Recurrent Neural Networks (RNNs) when processing long-period aging data, accurately capturing long-term dependencies in parameter degradation. Inside the LSTM unit, the forget gate... The input gate is responsible for adaptively removing redundant historical state information based on the current input. Then, the new information at the current moment is incorporated into the cell state. The degree of this. This dynamic regulation mechanism of information flow enables the model to sensitively capture recent deterioration fluctuations while remembering long-term health trends. Latent state With cell state The update logic is described by the following formula:

[0053] in, and These are the weight matrix and bias term of the forget gate, respectively; and These are the weight matrix and bias term of the input gate, respectively; and These are the weight matrix and bias term for the candidate cell states, respectively; and These are the weight matrix and bias term of the output gate, respectively; This represents the hidden state at the current time t; The input features enhanced by the attention mechanism at the current time are the context vectors obtained in equation (7). , The hidden state of the previous time step and the hidden state of the previous time step are used together as input variables for the gating structure. and These refer to the Sigmoid and hyperbolic tangent activation functions, respectively, in conjunction with... (Element-wise multiplication) enables non-linear adjustment and precise filtering of information flow intensity. Under this mechanism, cell state... Through the fusion of the forgetting gate Selected historical memories With input gate The weighted new candidate information completes the state update, and finally generates the hidden state at the current moment through output gate regulation. .

[0054] 3.4 Joint Training Introducing the ArcFace Loss Function This embodiment uses Mean Squared Error (MSE) as the regression loss function, aiming to force the model to accurately fit the continuous trend of primary power supply performance degradation over time by minimizing the Euclidean distance between the model's predicted output and the true degradation label. The regression loss calculation formula is as follows:

[0055] in, This represents the total number of samples in the current training batch; The representation model is for the first The remaining lifetime prediction value is output after processing by an LSTM layer and a fully connected layer from a given input sample. This represents the true remaining life label for the sample, which is obtained by normalizing the difference between the actual failure time recorded in the aging test and the current operating time.

[0056] During model training, traditional mean squared error loss functions only focus on numerical regression accuracy, often failing to effectively distinguish subtle feature differences between a healthy power source and a slightly degraded one. Therefore, this embodiment innovatively introduces the ArcFace (Additive Angular Margin Loss) loss function as an auxiliary constraint, enhancing feature discriminativity by maximizing inter-class distance and minimizing intra-class distance.

[0057] The power supply lifecycle is divided into three discrete health status labels (healthy, moderately aged, and critically failing). ArcFace loss adds an angular margin to the angle between the feature vector and the weight vector. This forces the model to learn angular space features with higher discriminative power. The formula is defined as:

[0058] in, Represents the total number of samples. The angle between the feature vector corresponding to the true class and its weight vector. This represents the angle between the feature vectors of other categories and their corresponding weight vectors. Parameter and All of these are adjustable hyperparameters, among which It serves to scale the characteristic amplitude, and Used to control the angular spacing between categories.

[0059] Ultimately, to balance accurate quantitative prediction of remaining useful life with accurate identification of degradation status, the model employs a joint loss function. Perform end-to-end optimization training:

[0060] in, The hyperparameter is used to balance the weights of the two loss terms.

[0061] The trained model outputs a predicted Remaining Usable Life (RUL) of the power supply at the current moment. This prediction, expressed in time or cycle count, accurately quantifies the remaining usable life from the current moment until the expected failure time of the device, providing a scientific basis for subsequent health management and operational decisions. To enhance the stability and robustness of the results, the predicted sequence undergoes weighted average smoothing after output, thereby avoiding overreaction caused by instantaneous fluctuations and making the overall lifetime curve smoother and more consistent with the physical laws of degradation.

[0062] Step 4: Adaptive PWM Hierarchical Self-Healing Control Based on the obtained RUL prediction value, the system further constructs a health index (HI) to achieve a more intuitive state representation. The health index is usually normalized to the range of [0, 1], where 1 represents the power supply in the initial healthy state and 0 represents the power supply approaching failure. The calculation methods of HI are mainly divided into two categories. One is the normalized expression based on RUL, and the other is the failure threshold determined by combining experiments. In this way, the absolute time scale of RUL is converted into a unified relative index, which is convenient for comparison between different test conditions and different batches of power supplies. More importantly, the trend of the health index can directly reflect the dynamic evolution of the degradation process, enabling managers to quickly judge the health stage of the power supply and serving as a direct basis for system state evaluation and operation strategy selection.

[0063] On this basis, the system enters the self-healing control stage, and its core is the adaptive PWM algorithm. Traditional PWM control methods are mainly designed for efficiency and steady-state performance optimization, often without considering the impact of device aging on the operating state. Their parameters are set as fixed values based on experience and are difficult to adapt to performance changes during long-term operation. The adaptive PWM algorithm can determine the adjustment amplitude and timing of parameters such as duty cycle and switching frequency based on the dynamic changes of the health index and RUL.

[0064] First, clarify the three-level interval of the health index (HI ∈ [0, 1]). Combining the degradation characteristics of the power supply and engineering practicability, the division criteria are as follows: (1) Healthy interval: HI ∈ [0.8, 1.0], the power supply performance has no obvious attenuation and is in the best operating state.

[0065] (2) Moderate degradation interval: HI ∈ [0.3, 0.8), the core devices begin to decay, and slight voltage drift and increased ripple occur.

[0066] (3) Failure critical interval: HI ∈ [0, 0.3), approaching the failure threshold, and derating operation is required to ensure the basic output.

[0067] Specifically, when the health index of the power supply is relatively high (0.8 < HI ≤ 1.0) and the RUL prediction value is sufficient, the system maintains the standard switching frequency and executes the conventional PWM control strategy. At this time, the power supply is in the normal working state, the switching frequency is maintained within the preset range, and the duty cycle is maintained at the initial set value to ensure the efficient operation and stability of the power supply.

[0068] As the health index of the power supply gradually decreases (0.3 < HI ≤ 0.8), the system enters the moderate degradation interval. At this stage, the RUL prediction value is close to the warning interval, and the system will appropriately reduce the switching frequency to reduce the power loss and thermal stress of the device. At this time, the PID control algorithm is introduced to dynamically correct the PID coefficients according to the health index. It can adjust the duty cycle in real time to optimize power supply performance, reduce energy waste, and slow down the degradation process.

[0069] When the power supply's health score further declines (0≤HI≤0.3) and the predicted RUL value approaches or falls below the critical value, the system enters the critical failure range and enters derating operation mode. During this stage, the switching frequency is significantly reduced, and the system adopts a more conservative PWM control strategy. To reduce excessive thermal damage and power waste of components, PID control continues to delay power supply degradation by precisely adjusting the duty cycle, ensuring that the power supply can provide stable power output as much as possible until it completely fails, thus extending its lifespan. Through this mechanism, the power supply system possesses the ability to actively intervene and adaptively adjust, realizing closed-loop control from state perception and lifespan prediction to strategy execution, thereby embodying the core concept of self-healing.

[0070] In specific control strategies, switching frequency The adjustment rules are as follows:

[0071] in, For the rated frequency, To adjust the coefficient, Rate the health of the power supply.

[0072] Duty cycle The adjustment is based on a PID controller, according to the deviation between the output voltage and the target voltage. As shown below:

[0073] in, To initially set the duty cycle, This is the proportionality coefficient. The integral coefficient is... is the differential coefficient.

[0074] Considering the proportionality coefficient Too small a value will lead to a delayed response, failing to compensate for voltage fluctuations caused by degradation in a timely manner; too large a value will easily cause overshoot. Therefore, a suitable value should be established. The dynamic correction model is shown in the following equation, where This is the baseline value for the proportion under healthy conditions; This is the preset proportional gain adjustment coefficient.

[0075]

[0076] (1) In the healthy range, the performance of the internal switching devices and filter capacitors of the power supply does not show significant degradation, and the output voltage fluctuation is only caused by minor disturbances in the power grid. near It can both compensate for small deviations with a small percentage gain and avoid the influence of... Excessive power consumption can lead to output overshoot, thus ensuring efficient and stable operation of the power supply.

[0077] (2) After entering the moderate degradation range, the on-state voltage drop of the main switch increases, the capacitance of the filter capacitor decreases, the output voltage drift amplitude increases, and the ripple frequency increases. Synchronous improvement, by accelerating the proportional adjustment speed, promptly offsets voltage deviations caused by device degradation, while avoiding output oscillations caused by excessively rapid increases, thus balancing response speed and stability.

[0078] (3) When the power supply enters the critical failure range, the core components are close to the failure threshold, and the risk of output voltage drop increases significantly. Upgraded to The maximum value is used to maximize the proportional adjustment and quickly compensate for severe voltage drops; however, it needs to be set. The upper limit prevents the system from becoming unstable due to excessive gain, while the output voltage limit ensures that the power supply can maintain basic output stability even in derating mode.

[0079] Its core function is to eliminate steady-state errors in the power supply output. Its adjustment needs to closely match the error characteristics of different health stages and the anti-integral saturation requirements. The formula is shown below, where... This is the baseline value for the proportion under healthy conditions; This is the preset proportional gain adjustment coefficient.

[0080]

[0081] (1) In the healthy range, the performance of the core power supply components does not significantly degrade, and the output voltage has only a small steady-state error. Maintaining a baseline value under healthy conditions, errors are eliminated only through a weak integral action, avoiding output fluctuations caused by excessive integration.

[0082] (2) After entering the moderate degradation range, as the switching devices age and the capacitor performance deteriorates, the steady-state error of the power supply output gradually increases. If the original value is maintained, the error will continue to increase. A high value can lead to error accumulation that cannot be compensated for in a timely manner, therefore it needs to be increased slowly as HI decreases. And keep the increase within a certain range to ensure error compensation efficiency and prevent output overshoot caused by excessive integral action.

[0083] (3) When the power supply enters the critical failure range, the device degradation approaches its limit, and the steady-state error reaches its maximum. It needs to be stabilized at a relatively high value to enhance the error elimination capability, but to avoid control inaccuracy caused by integral saturation, it needs to be matched with ±0.2. The integral limiter ensures that Ki functions within its effective range.

[0084] Its main function is to suppress power output oscillations and improve response smoothness. Its adjustment logic needs to be inversely adapted to the stability requirements during power degradation, based on the following formula: This is the differential baseline value under healthy conditions; This is the preset proportional gain adjustment coefficient.

[0085]

[0086] (1) In the healthy range, the power supply operates stably, and the anti-interference requirement is high. Maintaining a baseline value under healthy conditions, the system responds quickly to fluctuations in the input power grid or changes in load through sufficient differential action, suppressing instantaneous oscillations and maintaining stable output.

[0087] (2) After entering the moderate degradation range, the drift of internal device parameters of the power supply leads to a decrease in the system's damping characteristics. If the system continues to maintain a high level of damping, the damping characteristics will be affected. It easily amplifies high-frequency interference signals, causing high-frequency oscillations in the output voltage. Therefore, it needs to be adjusted as HI decreases. The rate gradually decreases, preserving the basic differential suppression capability while avoiding the risk of oscillation.

[0088] (3) When the power supply is in the critical failure range, the system stability has been greatly reduced. At this time, it is necessary to prioritize ensuring a basically stable output rather than a rapid response. By stabilizing the lower value, the sensitivity to system instability factors is reduced by weakening the derivative action, balancing response speed and operational stability, preventing excessive output fluctuations caused by excessive derivative action, and ensuring that the power supply can still provide continuous and stable power in derating operation mode.

[0089] Furthermore, the self-healing control method for the primary power supply of an integrated circuit aging test bench provided in this application embodiment includes, before acquiring the monitoring signal of the primary power supply of the integrated circuit aging test bench in real time during the aging test, an operation step of simulating the performance degradation of the primary power supply of the integrated circuit aging test bench, so as to provide a low-level verification platform for subsequent self-healing, which includes: Step S101: Based on the equivalent circuit model of the primary power supply (i.e., high-power power supply) of the integrated circuit aging bench, obtain the space state equation of the output voltage of the high-power power supply system. Step S102: Obtain the performance degradation models of multiple components with degradation characteristics in the high-power power supply system, and import the performance degradation models of the multiple components with degradation characteristics into the spatial state equation to obtain the spatial state model; Step S103: Using the spatial state model, obtain the performance degradation array data of the high-power power supply corresponding to each number of runs; Step S104: By concatenating the performance degradation array data of the high-power power supply corresponding to each number of runs, the performance degradation time sequence data of the high-power power supply is obtained.

[0090] In one embodiment of this application, the method further includes: calculating the maximum number of operations that the high-power power supply system needs to perform. Further, the step of calculating the maximum number of operations that the high-power power supply system needs to perform includes: calculating the maximum single-run time of the high-power power supply system; and calculating the maximum number of operations that the high-power power supply system needs to perform based on the simulation time length of the virtual simulation model and the maximum single-run time.

[0091] In one embodiment of this application, calculating the maximum single-run time of the high-power power supply system includes: obtaining the computer's physical memory capacity, the number of system space state variables, and the calculation step size; and using the computer's physical memory capacity, the number of system space state variables, and the calculation step size to calculate the maximum single-run time of the high-power power supply system.

[0092] In one embodiment of this application, obtaining the performance degradation array data of the high-power power supply corresponding to each number of runs using the spatial state model includes: Determine the current start time and the current operating state of the system parameters, and based on the current start time and the current operating state of the system parameters, calculate the intermediate stage state values ​​of the voltage flowing through the input capacitor, the voltage flowing through the output capacitor, and the current flowing through the input capacitor, respectively. By inputting the intermediate state values ​​of the voltage flowing through the input capacitor, the intermediate state values ​​of the voltage flowing through the output capacitor, and the intermediate state values ​​of the current flowing through the input capacitor into the spatial state equation, the derivative estimates of the intermediate state values ​​of the voltage flowing through the input capacitor, the intermediate state values ​​of the voltage flowing through the output capacitor, and the intermediate state values ​​of the current flowing through the input capacitor are obtained respectively. Update the current start time to the current end time, and based on the derivative estimates of the intermediate stage state values ​​of the voltage flowing through the input capacitor, the voltage flowing through the output capacitor, and the current flowing through the input capacitor, calculate the current end time state values ​​of the voltage flowing through the input capacitor, the voltage flowing through the output capacitor, and the current end time state values ​​of the current flowing through the input capacitor, respectively. Update the current operating status and operating time of each component in the system, and determine whether the updated operating time has reached the maximum operating time. If it is determined that the updated running time has reached the maximum running time, then the current running count is updated, and it is further determined whether the updated current running count has reached the maximum running count. If it is determined that the updated current running count has not reached the maximum running count, then the performance degradation array data of the high-power power supply corresponding to the current running count is saved.

[0093] In one embodiment of this application, obtaining the performance degradation array data of the high-power power supply corresponding to each number of runs using the spatial state model further includes: If it is determined that the updated runtime has not reached the maximum runtime, then the updated runtime will be used as the current starting time.

[0094] In one embodiment of this application, obtaining the performance degradation array data of the high-power power supply corresponding to each number of runs using the spatial state model further includes: If it is determined that the current number of runs has reached the maximum number of runs, then the performance degradation array data of the high-power power supply corresponding to the current number of runs is saved, and the performance degradation array data of the high-power power supply corresponding to all number of runs is calculated.

[0095] Furthermore, the high-power power supply system circuit includes: The input power supply voltage. This refers to the voltage across the two terminals (i.e., the power supply output voltage). The current flowing through the input capacitor, The current flowing through the inductor, The current flowing through the output capacitor. The internal resistance of the power supply For input capacitance, This is the equivalent series impedance of the input capacitor. For output capacitor, This is the equivalent series impedance of the output capacitor. For energy storage inductors, The equivalent series impedance of the energy storage inductor is... The impedance of the integrated circuit under test. This is the forward conduction voltage of the diode. This is the equivalent series impedance of the diode.

[0096] A simulation method for performance degradation of high-power power supplies includes the following steps: S0: Based on the high-power supply circuit model, the space state equation of the system output voltage can be obtained according to Kirchhoff's laws. The space state equation of the system output voltage is: (1.1) When the system is in the startup phase (i.e., diode forward biased and MOSFET turned on), The expression is: (1.2) in, , The expression is: (1.3) (1.4) in, The expression is: (1.5) When the system is in a stable operating phase (i.e., diode reverse biased and MOSFET turned on), The expression is: (1.6) in, The expression is the same as equation (1.4). The expression is: (1.7) When the MOSFET is in the off state, the system expression is the same during startup and stable operation. The expressions are all: (1.8) in, The expression is the same as equation (1.4). The expression is: (1.9) Switch to S1; S1: The user defines the types of components in the system that have degradation characteristics and those that do not. For example, if the inductor is set to be a component without degradation characteristics, then the inductance value during the entire simulation process... L and its equivalent series impedance If it is a constant, then it is a degenerate function model. After the definition is completed, go to S2; S2: User inputs the initial electrical parameters of all components, simulation time length. T The initial electrical physical quantities (voltage, current) of the internal components are: , , … , And the required calculation accuracy information for the numerical calculation part: absolute tolerance AbsTol, relative tolerance RelTol After completing the input, proceed to S3; S3: Extract parameter degradation models from the component performance degradation database for components with typical electrical signal degradation characteristics within the system. , …, And import it into the space state equation, where Define the number of components with degradation characteristics for the user, and after importing, transfer to S4; S4: Query computer's physical memory capacity And based on the calculation step size input by the user. Calculate the maximum running time per run. Based on T and Calculate the maximum number of runs required. ,in The calculation formula is: (1.10) in, The number of system space state variables, The number of bytes per data item is calculated and then transferred to S5. S5: Initialize the current iteration count of the system. The number of runs is Switch to S6; S6: Based on the space state equation of the power system output voltage, calculate the current iteration number. ， The number of runs is The output results and specific steps are as follows: S6.1: Based on the current time and the current operating status of the system parameters. The intermediate stage state is calculated as follows: (1.11) (1.12) (1.13) After the calculation is complete, proceed to S6.2; S6.2: Substitute the calculation results based on equations (1.11) to (1.13) into equations (1.1) to (1.9) to obtain the derivative estimates for the intermediate state. The details are as follows: (1.14) (1.15) (1.16) After the calculation is complete, proceed to S6.3; S6.3: Update the next moment The spatial state equation is: (1.17) (1.18) (1.19) After the calculation is complete, proceed to S7; It should be noted that this algorithm introduces three time points: the start time (current time), the intermediate time (used to assist the algorithm), and the end time (next time). The purpose is to calculate the value of the end time from the start time, while the intermediate time is mainly used by the algorithm.

[0097] S7: Determine the current running time Has the set maximum running time been reached? If so, proceed to S8 and record the current number of runs. Add 1 to go to S8; otherwise, return to S6. S8: Determine the current number of runs Has the maximum number of runs been reached? If the target is reached, proceed to S9; otherwise, save the current number of runs. The performance degradation array is Initialize the number of iterations Set the current status parameters of internal system components. And return to S6; S9: Organize the performance degradation arrays of each segment by... , , … , Connect and output as a performance degradation time series. The calculation ends.

[0098] It should be noted that the self-healing control method for the primary power supply of an integrated circuit aging bench provided in this application embodiment ensures that the primary power supply provides efficient and stable power for the subsequent fault-tolerant control operation steps of the high-temperature aging bench test chamber. The fault-tolerant control operation steps of the high-temperature aging bench test chamber include: Step S101: Determine the desired heating power of the heater in the integrated circuit high-temperature aging test chamber based on the temperature setpoint of the heater, including: The actual temperature of the heater in the high-temperature aging test chamber of the integrated circuit is collected, and the temperature difference of the heater is determined based on the actual temperature and the temperature set value. The desired heating power of the heater is obtained by inputting the temperature difference of the heater into an anti-integral saturation PI controller.

[0099] Step S102: By inputting the obtained operating data of the heater in the high-temperature aging test chamber of the integrated circuit into the pre-built and trained fusion prediction model, the degradation parameter prediction results of the heater are obtained, and the degradation parameter prediction results are corrected by introducing the Arrhenius model to obtain the corrected degradation parameter prediction results. In one embodiment of this application, the degradation parameter prediction results are corrected by introducing an Arrhenius model, resulting in the following corrected degradation parameter prediction results: By introducing the Arrhenius model, the first degradation rate of the heater at the current actual temperature and the second degradation rate at the reference temperature are calculated respectively. Based on the first degradation rate and the second degradation rate, a temperature correction coefficient is calculated, and the degradation parameter prediction result is corrected using the temperature correction coefficient to obtain the initial corrected degradation parameter prediction result. The initial modified degradation parameter prediction results are corrected in real time using the gradient descent algorithm to obtain the corrected degradation parameter prediction results.

[0100] In one embodiment of this application, the corrected degradation parameter prediction result includes: It includes real-time degradation parameters that estimate resistance and power degradation rate, degradation trend curves that include estimated resistance and power degradation rate over a future period, and dynamic fault warning thresholds that include mild and severe warnings.

[0101] In one embodiment of this application, after obtaining the corrected degradation parameter prediction result, the method further includes: When the estimated resistance exceeds the mild or severe warning level, a pre-set warning mechanism is activated to ensure the normal operation of the system. When the deviation between the estimated resistance and the actual measured value exceeds a set threshold, the prediction deviation correction mechanism is activated. The actual measured value is used to update the operating data of the heater input to the trained fusion prediction model. At the same time, the parameters introduced into the Arrhenius model are fine-tuned so that the deviation between the corrected estimated resistance and the actual measured value does not exceed the set threshold.

[0102] Step S103: Based on the expected heating power of the heater and the predicted result of the corrected degradation parameters, the current error of the heater is obtained, and the current error of the heater and the predicted result of the corrected degradation parameters are used to perform fault-tolerant control on the integrated circuit high-temperature aging test chamber.

[0103] In one embodiment of this application, fault-tolerant control of the integrated circuit high-temperature aging test chamber is performed using the current error of the heater and the predicted result of the corrected degradation parameters, including: Using the current error of the heater and the predicted result of the corrected degradation parameter, a current control law for heating the heater is calculated so that the heater is heated according to the current control law, thereby obtaining the actual current of the heater; Using the actual current of the heater, it is determined whether fault-tolerant control is required for the integrated circuit high-temperature aging test chamber. If it is determined that fault-tolerant control is required for the integrated circuit high-temperature aging test chamber, the fault warning information is further used to determine the type of fault-tolerant control. Based on the determined fault-tolerant control type, fault-tolerant control is applied to the high-temperature aging test chamber for the integrated circuit.

[0104] In one embodiment of this application, the current control law for heating the heater is calculated using the current error of the heater and the predicted result of the corrected degradation parameters, including: Using the current error and the rate of change of the current error of the heater, the sliding surface is calculated, and using the sliding surface, the basic control gain term and the exponential decay term are calculated respectively. Using the predicted results of the corrected degradation parameters, the degradation compensation term is calculated; By accumulating the basic control gain term, the exponential decay term, and the degradation compensation term, a current control law for heating the heater is obtained.

[0105] Furthermore, the specific method of fault-tolerant control of the high-temperature aging test chamber includes two parts: outer loop temperature control and inner loop power / current control.

[0106] Part 1: Outer Loop Temperature Control: As the top-level control module of the system, the core task of the outer loop temperature control is to calculate the expected power output of the heater based on the temperature target required by the experiment, and to provide a clear control direction for the inner loop. Its design details directly determine the final accuracy of temperature control.

[0107] Temperature error calculation: Input signals include "temperature setpoint" (Preset by the test program, supporting stepped heating / constant temperature / cooling modes, such as heating at 5℃ / min to 100℃ and holding at that temperature for 2 hours) and "actual temperature" (Data is collected by distributed thermocouples inside the chamber, positioned at key locations around the specimen, and the average value is taken to eliminate the influence of uneven temperature distribution). To avoid sensor noise (such as instantaneous jumps), ... A first-order low-pass filter (cutoff frequency 0.1Hz) is used, and then the formula is applied. Calculate the temperature error to ensure the error signal is smooth and reliable (fluctuation ≤ ±0.2℃).

[0108] Anti-integral saturation PI controller: Since the high-temperature test chamber has no active cooling device and relies on natural heat dissipation for cooling, the control input (heating power) can only be "increased" and not "decreased." The derivative element is prone to control mismatch during the cooling phase (e.g., the derivative output triggers heating when the temperature drops, which actually delays cooling). Therefore, a PI controller is adopted, and an "integral separation" anti-saturation strategy is introduced: When | |>5℃ (large error, such as during the heating stage): Only the proportional element (P gain) is activated to quickly increase the heating power and shorten the heating time; When | |≤5℃ (small error, such as during the heat preservation stage): Activate the proportional-integral circuit (I gain) to eliminate steady-state error through integration (such as temperature deviation ≤±0.5℃ during the heat preservation stage) and avoid temperature overshoot caused by integral saturation (overshoot ≤2℃).

[0109] When the heating power reaches the upper limit, the controller automatically pauses the integral calculation (freezes the current integral value) and only retains the proportional adjustment: to avoid the continuous accumulation of integral terms (i.e., "integral saturation") caused by "the temperature has not reached the target but the power has been saturated"; when the temperature rises and the power demand is lower than the upper limit, the integral action is resumed to ensure that it can quickly converge to the target value after exiting the saturation state.

[0110] Expected power output The output of the anti-integral saturation PI controller is the "desired heating power". Its value range is dynamically adjusted based on the heater's rated power (1000W) and the testing phase: Warming phase: Rapidly increase the temperature inside the chamber and ensure a stable heating rate (e.g., 5℃ / min, with a deviation ≤ ±0.5℃ / min). Insulation stage: It only compensates for heat loss from the cabinet (heat dissipation power of about 200W) and maintains stable temperature; Cooling phase: Turn off the heating and rely on natural heat dissipation to cool down, ensuring that the cooling rate meets the test requirements (e.g., 2℃ / min, deviation ≤ ±0.3℃ / min).

[0111] Part Two: Inner Loop Power / Current Control: The power / current inner loop, as the system's underlying execution module, deploys the core algorithm of this application (adaptive sliding mode control). Its task is to quickly offset local disturbances such as heater degradation and power supply fluctuations, ensuring that the actual power accurately tracks the desired power, thus guaranteeing temperature stability from the source. This is based on the heater's "pure resistive load" characteristic (power and current satisfy P=I). 2 R (where R is the heater resistance) This application uses closed-loop control to stabilize the current flowing through the heater, ensuring that the power accurately tracks the desired target from the source, while also offsetting local disturbances such as degradation and power fluctuations.

[0112] (1) Logic for converting power error to current target The control objective of the inner loop is to "track the desired power". "Transformed into "tracking the desired current" The core principle is to establish a dynamic relationship between power and current based on the state of resistor degradation, thereby avoiding control deviations caused by heater resistor degradation. a. Calculation of expected current: According to the theoretical relationship between power and current, P=I 2 R, Derivation of the formula for the desired current:

[0113] in, The "current estimated resistance of the heater" output by the parameter degradation prediction layer (dynamically updated as degradation occurs). This refers to the "desired heating power" output from the outer temperature loop.

[0114] Example: if =500W, after degradation =1.2 ( (where the initial resistance is), then Through dynamic adjustment To offset the power loss caused by increased resistance.

[0115] b. Current error calculation: Collect the actual current of the heater Calculate the current error:

[0116] Compared to "power error", current error more directly reflects the energy input status at the load end, which can reduce "power calculation lag" (such as power error calculation deviation caused by voltage fluctuations) and improve control response speed.

[0117] (2) Adaptive sliding mode current controller Leveraging the robustness of adaptive sliding mode, the sliding surface design and control law are optimized for current control objectives. The core principle is to dynamically adapt the current error to the heater degradation state in real time, ensuring that the current is tracked without overshoot or chattering.

[0118] a. Sliding surface design: Abandoning the traditional power sliding surface, a dual-state sliding surface of "current error + current change rate" is adopted to enhance the tracking capability of current dynamics:

[0119] in, The current error change rate (calculated in real time using a differential algorithm). The slope of the sliding surface (as...) Dynamic adjustment: When it increases, Synchronous increase improves current response speed; When stable, Reduce current to avoid overshoot.

[0120] b. Current control law (suppresses chattering + resists degradation): A composite control law combining an exponential reaching law and a degradation parameter correction is used to control the actuator's actual output current. Tracking target value This eliminates high-frequency chattering in sliding mode and also counteracts changes in load characteristics caused by degradation.

[0121] Meaning and function of each parameter: , Basic control gain (dynamically adjusted according to current level, especially in high-current scenarios) , (Increase synchronously to ensure matching control intensity) : Exponential decay term (when At that time, this term rapidly approaches 0, eliminating current chattering. Degradation compensation item (based on) With initial resistance Deviation calculation: This compensates for voltage loss caused by increased resistance in advance, preventing the current from failing to track the target due to "insufficient voltage".

[0122] Online parameter estimation (real-time correction) ): Combined current sampling value Voltage sampling value across the heater Real-time correction via gradient descent algorithm :

[0123] in, To estimate the step size (a small value ensures stable estimation), this formula is used to... Real-time approximation of the true resistance value ensures Calculation accuracy is improved to avoid current control failure caused by resistance estimation errors.

[0124] (3) Collaboration with other modules Degradation prediction layer output With power degradation Real-time input adaptive sliding mode controller: a. Used for correction Calculation (to avoid increased resistance leading to...) Low power (insufficient power). b. Used to adjust degradation compensation items ( When it increases, Synchronous increase to ensure that the current control force matches the degree of degradation.

[0125] When the fault-tolerant module detects an "inner loop anomaly" (judgment criterion: current error | |>10% Or the current fluctuation amplitude is >15%. ), triggering the following fault-tolerant actions: a. If the fault is "current sampling failure", switch to the backup current sampling circuit, and simultaneously perform degradation prediction. With voltage sample value Estimate To ensure uninterrupted current feedback; b. If the error is "current regulation module failure", switch to the backup current regulation module (same specifications as the main module, pre-calibrated) and reset the adaptive sliding mode controller parameters.

[0126] c. If the error is "heater failure", switch to the backup heater and reset the degradation prediction layer.

[0127] Among them, the parameter degradation prediction model in the inner loop power / current control: the degradation parameter prediction module is the core of "early detection of degradation, guidance of control and fault tolerance". Through the closed loop process of "data acquisition → preprocessing → fusion modeling → result output → feedback correction", it can accurately predict the error of heater resistance and temperature sensor.

[0128] (1) Data acquisition and preprocessing ① Data Acquisition Parameters and Hardware Configuration To assess the degradation characteristics of key components in the high-temperature test chamber, the core parameters to be collected and the hardware requirements were determined. These parameters include: the voltage across the heater terminals. Used to reflect the operating voltage status of the heater; heater circuit current Used to calculate actual resistance and power; actual temperature inside the chamber. Used to provide stress parameters in high-temperature environments; cumulative operating time of components. It is used to reflect the cumulative effect of degradation time.

[0129] ② Data preprocessing process The raw data collected contains noise (such as voltage fluctuations caused by electromagnetic interference) and missing data (such as momentary sensor disconnection), requiring three levels of preprocessing to ensure data validity. a. Outlier removal: The “3σ criterion” is used to identify outlier data (such as voltage instantaneously exceeding the rated value by 120%), and the moving average of the five adjacent data points is used to replace them; b. Noise filtering: Kalman filtering (to suppress high-frequency electromagnetic noise) is used for voltage and current signals, and first-order low-pass filtering (cutoff frequency 0.1Hz, to eliminate interference from ambient temperature fluctuations) is used for temperature signals. c. Data normalization: Map all parameters to the [0,1] interval (e.g., ... , , (This is ±20% of the heater's rated voltage) to avoid affecting the model training accuracy due to differences in dimensions (such as voltage V, time min).

[0130] (2) “LSTM+Arrhenius” fusion prediction model The architecture adopts a "serialized fusion" approach: first, the nonlinear temporal law of degradation is learned through LSTM, and then the degradation rate under high temperature environment is corrected through the Arrhenius model, which takes into account both the generalization of data-driven approaches and the physical interpretability of model-driven approaches.

[0131] ① First stage: LSTM neural network (data-driven layer) Function: Based on historical operating data, fit the relationship between heater resistance R and time. The nonlinear degradation curve.

[0132] Input layer: Dimensions are [number of samples, time step, number of features], where "time step" takes historical data from the past 60 minutes (i.e., 60 time points, corresponding to...). (1-minute interval); Number of features: 4 (normalized) , , , ); Hidden layers: 2 layers of LSTM units, 32 neurons per layer, using the "ReLU" activation function to avoid gradient vanishing; Output layer: 1 neuron, outputting the predicted heater resistance value for the next 10 minutes. ”; Training strategy: Dataset partitioning: 70% of historical data is used for training, 20% for validation, and 10% for testing; Loss function: Mean Squared Error (MSE), i.e. ,in To predict the resistance value, This is the measured resistance value; Optimizer: Adam optimizer, with an initial learning rate of 0.001 that decays by 10% every 100 rounds to ensure training convergence and avoid overfitting.

[0133] ② Second stage: Arrhenius model (physical correction layer) Function: Based on the physical characteristics of high-temperature environments, correct the prediction results of LSTM and solve the problem of "increased prediction bias of LSTM under extreme high temperatures".

[0134] Arrhenius model formula: Degradation rate k and absolute temperature ( The relationship is:

[0135] in: Pre-exponential factor (obtained from heater material handbook); Activation energy (determined through high-temperature accelerated aging tests); Gas constant: 8.314 J / (mol·K); Correction logic: Calculate the "temperature correction factor" ",in The degradation rate at the current temperature. The degradation rate at a reference temperature (e.g., 100°C); use Correcting the LSTM prediction results:

[0136] (3) Prediction result output and early warning mechanism ① The core output parameter module updates the prediction results every 10 minutes, outputting three types of key information: Real-time degradation parameters: (Estimated resistance of heater) (Power degradation, (Rated power of the heater). Degradation trend curve: Outputs the degradation trend curve for the next hour. , The change curve provides the control module with "long-term degradation prediction"; Fault warning threshold: Based on component failure criteria (e.g., heater resistance exceeding 120% of the initial value is considered a near-fault), a dynamic warning threshold is generated by combining prediction results—when... ( When the initial resistance is 0, a "mild warning" is triggered; when At that time, a "severe alert" was triggered.

[0137] ②Prediction bias correction mechanism When the deviation between the predicted value and the actual measured value exceeds 5% (e.g. , When measuring the actual resistance, activate the correction to ensure prediction accuracy: Use actual measured values Update the historical input data of the LSTM; Fine-tuning the activation energy of the Arrhenius model (Adjustment range ≤ 5%), so that the corrected and The deviation should be within 5% to avoid long-term forecast drift.

[0138] (4) Collaboration with other modules ①Supports inner-loop adaptive sliding mode control: Will Real-time input to the inner loop for calculation At the same time, adjust the slope of the sliding surface. ( (the initial slope); Optimize outer loop PI parameters: when When power degradation is severe, a "degradation signal" is sent to the outer loop, and the outer loop automatically increases the proportional gain of the PI controller. (Increase by 10%-20%) to improve temperature response speed and offset the effects of insufficient power.

[0139] ② Coordination with fault-tolerant control module Provides a basis for preventative fault tolerance: When a "severe warning" is triggered, a "heater is about to fail" signal is sent to the fault tolerance module; the fault tolerance module preheats the backup heater and loads the initial resistance parameters of the backup heater, shortening the original "fault switching time ≤ 0.5 seconds" to ≤ 0.2 seconds, thus avoiding temperature runaway.

[0140] Assisted fault location: When the fault-tolerant module detects an abnormal current (e.g., current is 0), it combines the degradation prediction with... Historical trends (e.g., the past 1 hour) (Continued increase) can be quickly identified as "gradual heater failure caused by degradation", avoiding the problem of "misjudging the fault type based solely on real-time signals" in the original solution, and improving the accuracy of fault-tolerant decision-making.

[0141] Figure 4 This application provides a schematic diagram of the structure of a primary power supply self-healing control device for an integrated circuit aging bench according to an embodiment of the present application. Figure 4 As shown, the device includes: The acquisition and preprocessing module is configured to acquire the monitoring signal of the primary power supply of the integrated circuit aging bench in real time during the aging test, and obtain the preprocessed monitoring signal by preprocessing the monitoring signal. The model building and training module is configured to build and train a parallel multi-channel convolutional long short-term memory network PMCCNN-ALSTM model with an attention mechanism, and obtain the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model. The acquisition module is configured to acquire the health score result at the current moment based on the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment. The self-healing control module is configured to perform adaptive pulse width modulation (PWM) hierarchical self-healing control on the primary power supply of the integrated circuit aging bench using the current health score result.

[0142] The above is a schematic scheme of a primary power supply self-healing control device for an integrated circuit aging bench according to this embodiment. It should be noted that the technical solution of this primary power supply self-healing control device for an integrated circuit aging bench belongs to the same concept as the technical solution of the aforementioned primary power supply self-healing control method for an integrated circuit aging bench. For details not described in detail in the technical solution of the primary power supply self-healing control device for an integrated circuit aging bench, please refer to the description of the technical solution of the aforementioned primary power supply self-healing control method for an integrated circuit aging bench.

[0143] A computing device is provided according to one embodiment of this application. The components of the computing device include, but are not limited to, a memory and a processor. The processor and the memory are connected via a bus, and a database is used to store data.

[0144] The computing device also includes access devices that enable the computing device to communicate via one or more networks. Examples of such networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. Access devices may include one or more of any type of wired or wireless network interface (e.g., a Network Interface Card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0145] The computing device can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device can also be a mobile or stationary server.

[0146] The processor is used to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-mentioned integrated circuit aging bench primary power supply self-healing control method.

[0147] An embodiment of this application also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described integrated circuit aging bench primary power supply self-healing control method.

[0148] An embodiment of this application also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described integrated circuit aging bench primary power supply self-healing control method.

[0149] The above are illustrative schemes of a computing device, a computer-readable storage medium, and a computer program according to this embodiment. It should be noted that each technical solution belongs to the same concept as the above-described battery swapping method based on battery appearance recognition and battery swapping authorization. Details not described in detail for each technical solution can be found in the description of the above-described integrated circuit aging bench primary power supply self-healing control method.

[0150] The preferred embodiments of this application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of this application shall be within the scope of the claims.

Claims

1. A self-healing control method for the primary power supply of an integrated circuit aging test bench, characterized in that, include: During the aging test, the monitoring signal of the primary power supply of the integrated circuit aging bench is acquired in real time, and the monitoring signal is preprocessed to obtain the preprocessed monitoring signal. A parallel multi-channel convolutional long short-term memory network PMCCNN-ALSTM model with an integrated attention mechanism is constructed and trained. The remaining service life of the primary power supply of the integrated circuit aging bench at the current moment is obtained by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model. Based on the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment, obtain the health score result at the current moment; The adaptive pulse width modulation (PWM) hierarchical self-healing control of the primary power supply of the integrated circuit aging bench is performed using the current health score result, which includes: Using the current health score and remaining lifetime, determine the switching frequency for adaptive PWM. and duty cycle ; Utilizing the switching frequency and duty cycle Adaptive PWM hierarchical self-healing control is applied to the primary power supply of the integrated circuit aging bench. Wherein, the switching frequency include: Among them, the For the rated frequency, the To adjust the coefficient, the For health score results; The duty cycle include: Among them, the The deviation between the output voltage and the target voltage, the For the integration process The deviation between the output voltage and the target voltage at the corresponding moment; To initially set the duty cycle; the The proportionality coefficient; The integral coefficient; The differential coefficients; The ratio benchmark value under healthy conditions; The preset proportional gain adjustment coefficient; The ratio benchmark value under healthy conditions; The preset proportional gain adjustment coefficient; The differential benchmark value under healthy conditions; This is the preset proportional gain adjustment coefficient.

2. The method according to claim 1, characterized in that, The monitoring signals of the primary power supply of the integrated circuit aging bench include at least the input voltage, output voltage, output current, forward voltage drop of the output rectifier diode, and on-state voltage drop of the main switch.

3. The method according to claim 1, characterized in that, The PMCCNN-ALSTM model comprises a parallel multi-channel convolutional neural network (PMCCNN) unit consisting of multiple parallel convolutional channels, an attention-based feature fusion unit, and a long short-term memory (LSTM) unit. The PMCCNN unit extracts frequency domain signals from the preprocessed monitoring signal at different times using each convolutional channel, and concatenates these signals in chronological order to form a temporal feature with a time dimension. The attention-based feature fusion unit performs deep spatiotemporal fusion processing on the temporal feature with a time dimension to obtain a context vector containing global key information at the current time. The LSTM unit uses the context vector containing global key information at the current time and the hidden state from the previous time step to obtain the hidden state at the current time, and uses this hidden state as the remaining lifespan of the primary power supply of the integrated circuit aging test bench at the current time.

4. The method according to claim 3, characterized in that, By inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model, the remaining lifespan of the primary power supply of the integrated circuit aging bench at the current moment is obtained, including: The frequency domain signal at each time step of the preprocessed monitoring signal is processed by one-dimensional convolutional feature mapping using the convolutional kernels in the PMCCNN unit of the trained PMCCNN-ALSTM model, so as to obtain the first output result of each feature map of each convolutional layer at each time step. By introducing the ReLU function into the PMCCNN unit, the first output result is subjected to non-linear activation processing to obtain the activation output result of each feature map of each convolutional layer at each time step; The activation output is downsampled and dimensionality reduced by the PMCCNN unit to obtain the low-dimensional deep features of each feature map at each time step. The low-dimensional deep features of all time steps are then concatenated in chronological order to form a temporal feature with a time dimension.

5. The method according to claim 4, characterized in that, The remaining lifespan of the primary power supply of the integrated circuit aging bench at the current moment is obtained by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model. By utilizing the attention-based feature fusion unit in the trained PMCCNN-ALSTM model, the temporal features with the time dimension are sequentially subjected to spatial adaptive channel feature fusion and temporal attention mechanism processing to obtain the context vector containing global key information at the current moment.

6. The method according to claim 5, characterized in that, The remaining lifespan of the primary power supply of the integrated circuit aging bench at the current moment is obtained by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model. The LSTM unit in the trained PMCCNN-ALSTM model is used to perform deep temporal modeling on the context vector containing global key information at the current moment to obtain the hidden state at the current moment, and the hidden state at the current moment is used as the remaining service life of the first-level power supply of the integrated circuit aging bench at the current moment.

7. A self-healing control device for the primary power supply of an integrated circuit aging bench, the device being used to implement the method as described in any one of claims 1-6, characterized in that, include: The acquisition and preprocessing module is configured to acquire the monitoring signal of the primary power supply of the integrated circuit aging bench in real time during the aging test, and obtain the preprocessed monitoring signal by preprocessing the monitoring signal. The model building and training module is configured to build and train a parallel multi-channel convolutional long short-term memory network PMCCNN-ALSTM model with an attention mechanism, and obtain the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment by inputting the preprocessed monitoring signal into the trained PMCCNN-ALSTM model. The acquisition module is configured to acquire the health score result at the current moment based on the remaining service life of the primary power supply of the integrated circuit aging bench at the current moment. The self-healing control module is configured to perform adaptive pulse width modulation (PWM) hierarchical self-healing control on the primary power supply of the integrated circuit aging bench using the current health score result.

8. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the self-healing control method for the primary power supply of the integrated circuit aging bench as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the integrated circuit aging bench primary power supply self-healing control method according to any one of claims 1 to 6.

Images