Integrated Circuit Power Supply Redundancy Self-Healing Control Method and Device

Abstract

This specification provides an integrated circuit power supply analytical redundancy self-healing control method and apparatus. The integrated circuit power supply analytical redundancy self-healing control method includes: acquiring the current operating data and historical operating data at multiple historical moments collected by various sensors in an aging bench monitoring system; extracting fusion features from the historical operating data and the current operating data; processing the highly discrete data in the operating data and the prediction results of the operating data from the previous moment to obtain the prediction results of the operating data at the current moment; performing multi-task state evaluation based on the operating data, fusion features, and the prediction results of the operating data at the current moment to generate evaluation results of the output voltage, ripple voltage, and health status of the secondary power supply; updating the parameters of the PID controller based on the evaluation results; calculating and outputting control signals through the updated PID controller to perform self-healing control of the output power of the secondary power supply.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of aging test technology, and in particular to an integrated circuit power supply redundancy self-healing control method. Background Technology

[0002] Aging testing is a critical production step to ensure the long-term reliability of electronic products and is used for quality screening. A typical high-temperature aging test chamber for electronic products employs a tiered power supply architecture, including several primary aging test power supplies (referred to as "primary power supplies") and a programmable secondary aging test power supply module (referred to as "secondary power supplies"). When the high-temperature test chamber is operating normally, the secondary power supply converts the intermediate bus voltage provided by the primary power supply into the precise, stable, and adjustable DC operating voltage required by various integrated circuits and electronic components under test on the aging bench, thereby ensuring their long-term reliable operation under harsh testing environments of high temperature and full load.

[0003] The high-precision, low-ripple voltage output of the secondary power supply ensures stable power supply to the tested load. However, long-term operation can lead to component aging and poor heat dissipation, causing fluctuations in the secondary power supply's output voltage and failure to meet test requirements, thus reducing the reliability and consistency of aging tests. Due to insufficient sensor accuracy and untimely signal calibration, current methods for judging the true state of the secondary power supply's output power and controlling the signal suffer from inaccuracies in accuracy and response speed, increasing test risks. Therefore, an effective method is urgently needed to address these issues. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a method for self-healing control of redundancy in integrated circuit power supply analysis. One or more embodiments of this specification also relate to an integrated circuit power supply redundancy self-healing control device, 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 specification, an integrated circuit power supply redundancy self-healing control method is provided, comprising: Acquire the operating data of the secondary power supply at the current moment and the historical operating data at multiple historical moments collected by at least two sensors in the aging bench monitoring system; Extract fusion features from the historical operation data and the operation data; The highly discrete data in the running data and the prediction results of the running data at the previous time step are processed by an adaptive Kalman filter driven by a hypernetic network to obtain the prediction results of the running data at the current time step. Based on the operational data, the fusion features, and the operational data prediction results at the current moment, a multi-task state evaluation based on the attention mechanism is performed to generate evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply. Based on the evaluation results and reinforcement learning algorithm, the parameters of the PID controller are updated. The updated PID controller calculates and outputs a control signal to perform self-healing control of the output power of the secondary power supply.

[0006] Optionally, the integrated circuit power supply redundancy self-healing control method further includes: Based on the fusion features, a weighted average is performed on the low-dispersion data in the running data to generate the corresponding processing result; The processing results and the current time-based operational data prediction results are then subjected to gating fusion to generate a corresponding fusion result; Based on the operational data, the fusion features, and the fusion results, a multi-task state evaluation based on an attention mechanism is performed to generate evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply.

[0007] Optionally, the integrated circuit power supply redundancy self-healing control method further includes: The fused features are input into the supernetwork for mapping processing to generate an adaptive observation noise covariance matrix and an adaptive process noise covariance matrix, wherein the fused features include joint state features and discreteness features; The adaptive process noise covariance matrix and the posterior error covariance matrix of the previous time step are input into the Kalman filter for prediction processing to obtain the prior error covariance matrix of the current time step. Determine the error between the running data and the predicted running data at the current time, and calculate the Kalman gain based on the adaptive observation noise covariance matrix and the prior error covariance matrix at the current time; Based on the Kalman gain and the error, the prediction results of the current running data are corrected.

[0008] Optionally, the step of performing multi-task state evaluation based on the operating data, the fused features, and the current operating data prediction results, and generating evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply, includes: The fused features are mapped to query vectors, and the running data is mapped to keywords and key values; The query vector, the keyword, and the key value are subjected to context enhancement processing based on the cross-attention mechanism to generate context enhancement features; The context enhancement features are input into the physical parameter evaluation layer in the two-level cascaded network structure for processing, generating evaluation results corresponding to the output voltage and ripple voltage of the secondary power supply, respectively. The context enhancement features and the evaluation results corresponding to the output voltage and ripple voltage of the secondary power supply are input into the health state inference layer in the two-level cascaded network structure for processing, thereby generating the evaluation results corresponding to the health state of the secondary power supply.

[0009] Optionally, the integrated circuit power supply redundancy self-healing control method further includes: Obtain the parameter values ​​of at least two uncertainty parameters output by the physical parameter evaluation layer and the health status inference layer, wherein the at least two uncertainty parameters are used to evaluate the uncertainty of the evaluation result; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the evaluation result of the output voltage output by the physical parameter evaluation layer, calculate the first loss value of the normal-inverse gamma distribution negative log-likelihood loss function; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the evaluation result of the ripple voltage output by the physical parameter evaluation layer, calculate the second loss value of the normal-inverse gamma distribution negative log-likelihood loss function; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the health status assessment result output by the health status inference layer, calculate the third loss value of the normal-inverse gamma distribution negative log-likelihood loss function; The first loss value, the second loss value, and the third loss value are weighted and calculated to obtain the corresponding total loss value; The network parameters of the two-level cascaded network structure are optimized based on the total loss value.

[0010] Optionally, updating the parameters of the PID controller based on the evaluation results and the reinforcement learning algorithm includes: A state space vector is constructed based on the evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply, as well as the rated voltage of the secondary power supply. The state space vector is input into the policy network in the reinforcement learning framework for processing to generate an action space vector, wherein the action space vector is composed of the PID parameter adjustment amount of the PID controller. The parameters of the PID controller are updated based on the action space vector.

[0011] Optionally, the integrated circuit power supply redundancy self-healing control method further includes: The reward function value is calculated based on the state space vector and action space vector at the current moment; Determine the state space vector for the next moment, and store the current state space vector, action space vector, the next state space vector, and the function value into the experience replay buffer. Experience data is periodically collected in the experience replay buffer at a preset frequency, and the network parameters of the policy network and value network in the reinforcement learning framework are updated based on the experience data using the stochastic gradient descent algorithm.

[0012] According to a second aspect of the embodiments of this specification, an integrated circuit power supply redundancy self-healing control device is provided, comprising: The acquisition module is configured to acquire the operating data of the secondary power supply at the current moment and the historical operating data at multiple historical moments collected by at least two sensors in the aging bench monitoring system. The extraction module is configured to extract fusion features from the historical running data and the running data; The processing module is configured to process the highly discrete data in the running data and the prediction results of the running data at the previous time step through an adaptive Kalman filter driven by a hypernetic network to obtain the prediction results of the running data at the current time step. The generation module is configured to perform multi-task state evaluation based on the running data, the fusion features, and the running data prediction results at the current moment, and generate evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply. The control module is configured to update the parameters of the PID controller based on the evaluation results and reinforcement learning algorithm, and calculate and output a control signal through the updated PID controller to perform self-healing control of the output power of the secondary power supply.

[0013] According to a third aspect of the embodiments of this specification, 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 power supply redundancy self-healing control method.

[0014] According to a fourth aspect of the embodiments of this specification, 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 power supply resolution redundancy self-healing control methods.

[0015] According to a fifth aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described integrated circuit power supply redundancy self-healing control method.

[0016] This embodiment acquires the current operating data of the secondary power supply from various sensors in the aging bench monitoring system, as well as historical operating data from multiple historical time points. It extracts fusion features from the historical and current operating data. High-discreteness data in the operating data and the prediction results of the previous time point's operating data are processed to obtain the current time's operating data prediction result. Based on the operating data, fusion features, and the current time's operating data prediction result, a multi-task state assessment is performed, generating assessment results for the secondary power supply's output voltage, ripple voltage, and health status. Based on the assessment results, the PID controller's parameters are updated, and the updated PID controller calculates and outputs a control signal to perform self-healing control of the secondary power supply's output power. This embodiment effectively overcomes the errors and noise effects of a single sensor by fusing data from multiple sensors, thus improving the monitoring accuracy of the operating data. Furthermore, by processing the high-discreteness data through a Kalman filter data fusion network driven by a hypernetic network, and performing self-healing output power control based on reinforcement learning algorithms and a fuzzy PID controller, it not only facilitates accurate assessment of the secondary power supply system's state but also enables online adaptive adjustment and control of relevant secondary power supply parameters, thereby ensuring the stable operation of the secondary power supply system. Attached Figure Description

[0017] Figure 1 This is a flowchart of an integrated circuit power supply redundancy self-healing control method provided in one embodiment of this specification; Figure 2 This is a flowchart illustrating the processing procedure of an integrated circuit power supply redundancy self-healing control method according to one embodiment of this specification. Figure 3 This is a schematic diagram of a data validity analysis process provided in one embodiment of this specification; Figure 4 This is a schematic diagram of a data fusion process provided in one embodiment of this specification; Figure 5 This is a schematic diagram of a power self-healing control process provided in one embodiment of this specification; Figure 6 This is a schematic diagram of the structure of an integrated circuit power supply redundancy self-healing control device according to one embodiment of this specification; Figure 7 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0018] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0019] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0020] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0021] To address the shortcomings of current aging bench monitoring systems in terms of real-time performance and accuracy, and to improve the quality and efficiency of aging tests, this specification proposes an embodiment that enhances the monitoring dimensions of the secondary power supply system's operating status and deeply integrates multi-dimensional data to achieve precise control of the secondary power supply's output power. Furthermore, when the secondary power supply deviates from its output power due to component aging, instantaneous overload, or other reasons, affecting the effectiveness of the aging test, the proposed multi-dimensional data monitoring and multi-sensor fusion method enables rapid and automatic adjustment and recovery of the system's output operating voltage, achieving self-healing control of the secondary power supply system's output power.

[0022] This specification provides a method for self-healing control of redundancy in integrated circuit power supply analysis. It also relates to an integrated circuit power supply redundancy self-healing control device, a computing device, a computer-readable storage medium, and a computer program, which will be described in detail in the following embodiments.

[0023] Figure 1A flowchart of an integrated circuit power supply redundancy self-healing control method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0024] Step 102: Obtain the operating data of the secondary power supply at the current moment and the historical operating data at multiple historical moments from at least two sensors in the aging bench monitoring system.

[0025] The flowchart of the processing procedure of the integrated circuit power supply redundancy self-healing control method provided in the embodiments of this specification is as follows: Figure 2 As shown.

[0026] The specific process is as follows: The entire process is divided into two parts: the sensor data fusion stage and the output power self-healing control stage. The sensor data fusion stage begins with multi-sensor data acquisition, simultaneously acquiring the current operating data of the secondary power supply through multiple sensors. This operating data includes the output voltage and current signals of the secondary power supply. This operating data is then processed using quality labels to generate a labeled dataset. Subsequently, these labeled data undergo data validity analysis and data dispersion analysis. Combining the current operating data of the high-temperature test chamber with historical operating data from multiple historical moments, it is determined whether the amount of valid data is greater than or equal to an acceptable threshold. If so, the data is initially screened based on dispersion analysis, and then proceeds to the dispersion analysis stage. Dispersion analysis categorizes data based on its dispersion: for high-dispersion data, a weighted fusion algorithm based on Hyper-Network Kalman Filter (GHF-NET) is used for dynamic weight optimization, while for low-dispersion data, a more computationally efficient weighted average fusion is used. If the amount of valid data is less than the acceptable threshold, a prediction compensation mechanism is triggered, using a data generation model based on physical model constraints to complete the missing data. The fused / compensated data undergoes unified multi-task state evaluation based on an attention mechanism to generate the input control signals required by the PID controller. Subsequently, a reinforcement learning agent is constructed, and reinforcement learning is used to control PID parameter updates. A fuzzy PID controller then executes real-time output control to update the state and achieve self-healing control of the secondary power supply's output power. Based on this process, a complete closed loop of "acquisition-fusion-control-feedback" is formed. This process achieves high-reliability control through a triple fault-tolerance mechanism (data verification / fusion compensation / PID parameter tuning), making it suitable for industrial automation scenarios with stringent stability requirements.

[0027] In practical applications, during high-temperature test chamber testing, each test channel (control and detection board) is equipped with multiple secondary power supply modules. The voltage and current outputs of each power supply module differ, and their output directions are also inconsistent. For the entire secondary power supply system, multiple redundant signal sensors are deployed to collect its output voltage and current signals; for example, each sensor has multiple redundant sensing points. Therefore, the raw data obtained from the secondary power supply can be represented as follows: in, This is the voltage sample value. This is the current sample value. It is also important to note that the voltage and current sampling frequency for monitoring the secondary power supply operation of the high-temperature test chamber should be set to ≥10kHz. All collected data will be stored in the monitoring system's historical database. Upon data entry, not only will the raw values ​​be stored, but also, through instantaneous anomaly detection, preliminary quality labels will be automatically assigned, providing metadata for subsequent analysis.

[0028] The moment data flows into the database, a lightweight real-time computing unit is used to quickly analyze each data point or data segment (such as data within a time window) and generate one or more quality tags. These tags will be stored as metadata along with the original data.

[0029] Specifically, the following quality labels are defined: Transient anomaly label (Outlier): 0 indicates normal and 1 indicates abnormal.

[0030] Short-term stability label (Stability): or .

[0031] Data Validation Labels (Validity): 0 indicates valid, and 1 indicates invalid (e.g., exceeding the sensor's physical range).

[0032] In the above, the superscript (i) represents the i-th sensor.

[0033] Let's take data from a single time window of a sensor as an example. Assume the window length is N, and the data sequence within the window is... .

[0034] For transient outlier detection, an improved Z-Score method is employed, which is sensitive to impulsive outliers. First, the median (MEDIAN) and median absolute difference (MAD) of the window sequence are calculated:

[0035] Next, calculate the robust Z-Score for each data point:

[0036] The constant 0.6745 is used to make MAD a consistent estimator of the standard deviation of the normal distribution.

[0037] Then, the following judgment rule is applied: in, It is a threshold, which can be obtained by calculating the quantiles from historical normal operation data or by directly taking 3.5 (corresponding to a 99.9% confidence interval).

[0038] For short-term stability, window statistics are used to define stability. Variance can be chosen as the evaluation metric. Next, the window variance is calculated: in, It is a window The mean.

[0039] Subsequently, the variance was normalized by calculating the coefficient of variation to eliminate the influence of different sensor dimensions.

[0040]

[0041] It is worth noting that when Special handling is required at times.

[0042] Then set a stability threshold and label it.

[0043]

[0044] For data validity, a hard judgment is made based on common sense physics and sensor specifications. In addition, if the secondary power module monitoring system is equipped with a corresponding sensor self-calibration module (such as a calibration unit based on standard test conditions), the data validity verification process for faulty sensors can be triggered.

[0045]

[0046] Completing the above process and labeling the data provides a data foundation for subsequent data stability analysis and fusion.

[0047] Next, online adaptive data validity analysis guided by quality labels can be performed.

[0048] This specification provides a schematic diagram of a data validity analysis process, as shown in the embodiments below. Figure 3 As shown.

[0049] First, regarding symbol definition and input: For the dataset of historical running data, each data point .

[0050] in, This represents the initial reading of sensor i at time t.

[0051] This represents the quality label vector (e.g., anomaly label) assigned to the data at this moment through validity analysis. Stability tag (The encoded value).

[0052] This represents the new observation data at the current moment (i.e., the operational data at the current moment, which does not yet have a quality label).

[0053] Before performing data validity analysis on the current operational data, the historical operational data is first cleaned and weighted to utilize the previously generated quality labels to build a cleaner and more reliable dataset for operational condition clustering.

[0054] First, invalid data is removed. Data marked as invalid is removed from the historical dataset X. or instantaneous anomaly Data points. Forming a preliminary cleanup set. .

[0055] Secondly, stability weighting is applied to the purification set: Each data point in the dataset is assigned a weight. The weight is determined by its stability label. Decide:

[0056] in, ,For example: .

[0057] Since highly stable data points have a greater impact on cluster centers in clustering, the clustering results are more representative of the stable operating conditions of the system.

[0058] Then, a weighted Gaussian Mixture Model (WGMM) was used for offline working condition clustering. The probability density function of WGMM is:

[0059] in, The mixing coefficients (weights) of the k-th Gaussian component satisfy the following condition: and . Let be the probability density function of the k-th Gaussian component.

[0060] These are all the unknown parameters of the model.

[0061] The likelihood function of a traditional GMM is WGMM becomes a weighted likelihood function. It incorporates the weights of data points. The goal of optimizing the entire dataset is:

[0062] The EM algorithm is used to train the WGMM model, which automatically determines K Gaussian distributions, each representing a load condition cluster (e.g., "light load stability", "heavy load fluctuation", "start-stop transition"). The final output is the parameters for each cluster k: the mean vector. Covariance matrix and the prior probability of the cluster .

[0063] Specifically, the training process based on the EM algorithm is as follows: ① Initialization: Initialize model parameters randomly or using K-means++. .

[0064] ② Given the current parameter estimate Under the condition of calculating data points The posterior probability (i.e., "responsibility") from the k-th component. The iteration proceeds for E steps (expected steps) until convergence (for the l-th iteration), calculated as follows:

[0065] This can be understood as data points. The probability generated by the k-th component.

[0066] ③ Utilizing the "responsibility" calculated using the E-step To update the model parameters Θ and maximize the weighted likelihood function, the mixing coefficients are first updated. :

[0067] Secondly, update the mean vector. : The new mean is a weighted average of all data points, where the weight of each data point is 1. High weight And high responsibility The data points have a greater impact on the mean.

[0068] Subsequently, the covariance matrix is ​​updated. :

[0069] ④ When the log-likelihood function The algorithm converges when the change in the magnitude of the change or the change in the parameter is less than a certain threshold.

[0070] Next, we construct the working condition clusters (offline segment). For each cluster k, we find the set of all data points belonging to that cluster. Calculate the statistics for the data within this cluster: Mean: .

[0071] Standard deviation: .

[0072] Then store all and This forms a set of statistical quantities for the operating condition cluster.

[0073] Once training is complete, online pattern matching and dynamic threshold generation (online segments) can be performed based on the trained WGMM model.

[0074] For new observation data, the operating condition it belongs to is determined in real time, and corresponding effective thresholds are generated. The method and steps are as follows: First, for new data points The posterior probability of belonging to each cluster k is calculated using the trained WGMM model:

[0075] The cluster with the highest posterior probability is identified as the current operating condition: .

[0076] Secondly, based on the matched cluster of working conditions The mean and standard deviation of the corresponding sensor i are read from the operating condition statistics library, and a dynamic threshold is generated:

[0077]

[0078] in, It is a constant factor (usually 2.5, 3, or 3.5) representing the width of the confidence interval.

[0079] Subsequently, an validity assessment is performed, determining the current sensor reading. Is it within the threshold range?

[0080] Output a validity flag, which can provide weights for subsequent fusion modules (e.g., the weight is reduced to 0 when invalid).

[0081] To adapt to slow system changes (such as equipment aging), the WGMM model and operating condition statistics library can be updated periodically (such as daily) with new incremental data with quality labels, so that the thresholds always remain consistent with the current system characteristics.

[0082] If the amount of valid data is less than an acceptable threshold, a predictive compensation mechanism is triggered, using a data generation model based on physical model constraints to complete the missing data. That is, when some sensor channels (such as a phase voltage sensor) fail or data is severely missing, some of the above methods will become ineffective. To address the issue of insufficient monitoring sensors, this specification's embodiments utilize a data generation and fusion evaluation method based on physical model constraints to generate physically reasonable alternative data, thereby achieving accurate assessment of the system state.

[0083] The specific implementation process is as follows: (1) As can be seen from the foregoing, the known quantities that can be used as model inputs include: Observational data from some sensors during normal operation: (Set up M available sensor data, where M is less than an acceptable threshold).

[0084] System input / control signals: (e.g., PWM duty cycle).

[0085] Unknown / to-be-generated quantity refers to data with missing channels. (Set up K missing sensor data, including voltage, current, temperature, etc.).

[0086] The parameters to be evaluated include: secondary power supply output voltage: Ripple voltage: .

[0087] (2) Constructing a Physically Constrained Generative Adversarial Network (PI-GAN) 1) Basic structure: Generator: Input: Random noise vector z, current control signal Historical observation data .

[0088] Output: Generated missing data .

[0089] The network structure can employ the Transformer module to capture the dynamic timing characteristics of the system.

[0090] Discriminator: The discriminator is used to distinguish between real data and generated data, prompting the generator to produce statistically realistic data.

[0091] Input: Actual missing data (Trained from the complete dataset) or generated data .

[0092] Output: The probability that the data is true.

[0093] 2) General equations of physical constraints.

[0094] For MOSFETs or other components, it's often impossible to determine their specific circuit connection, but all electrical components must adhere to the laws of conservation of energy and thermodynamics. The losses of any electrical component (MOSFET, IGBT, resistor, inductor) are converted into heat. Therefore, the embodiments in this specification establish the general physical constraint equation for a two-stage power supply system: Input Power - Output Power = Power Loss This leads to a rise in temperature.

[0095] Therefore, a new physical loss function is established based on the constraints of the component's own characteristics:

[0096] in, The physical constraint loss function is one of the objectives to be minimized when training a neural network. This is the L2 norm (squared). It is used to measure the distance (error) between the predicted value and the target value. Current / voltage data generated for the GAN network. The power loss calculation function (for MOSFETs, it can be...) ). As a pre-trained thermodynamic surrogate model, based on the current and voltage generated by the generator, using physical formulas (or pre-trained models) The theoretical temperature rise was "calculated". The "temperature rise" data generated by the generator is a value directly output by the generation network, representing what it "thinks" is the temperature change of the device at this time.

[0097] Regardless of whether the component is in a Buck, Boost, or inverter system, its electro-thermal conversion physical characteristics remain unchanged. Thus, the electro-thermal coupling consistency constraint becomes a topology-independent physical constraint mechanism. It is implanted into the generative network.

[0098] 3) Pre-trained thermodynamic surrogate model Since in most cases, the specifics are unclear. Parameters. Or, if the system contains complex heat sinks and air-cooling conditions, traditional RC networks are generally insufficient to describe it. In such cases, training a small neural network to replace the physical equations becomes more general. Utilizing the electrical-thermal analogy—current analogy to power loss, voltage analogy to temperature, resistance analogy to thermal resistance, capacitance analogy to heat capacity, etc.—a thermodynamic surrogate model can be designed to simulate the physical laws governing the temperature rise of power devices. The goal of pre-training the thermodynamic surrogate model is to learn the "power..." The mapping relationship of "temperature".

[0099] First, the proxy model architecture needs to be designed. Since heat conduction has a significant time lag, RNN-like networks are more in line with the physical nature than fully connected layers, so the base model can be selected as LSTM or GRU.

[0100] Input: Current power (Calculated from current and voltage), temperature at the previous moment (Reflects thermal inertia).

[0101] Output: Current temperature .

[0102] Secondly, data preparation and pre-training are necessary. Before training the backbone PI-GAN, the surrogate model needs to be trained first. Data can be obtained by performing several simulations of typical operating conditions using COMSOL / Icepak to obtain accurate data. Alternatively, you can use the "transient thermal impedance curve" from the original system manual to generate a batch of composite data.

[0103] First, construct the dataset, with the input sequence. tag sequence Train an LSTM model and minimize MSELoss.

[0104]

[0105] Once the surrogate model training converges, all its weight parameters are frozen. In subsequent PI-GAN training, it will be embedded into the computation graph as a non-trainable, differentiable "physical operator".

[0106] 4) Implantation of PIGAN The pre-trained proxy model (denoted as...) This is embedded as a non-trainable computational layer into the discriminator or loss computation flow of PI-GAN. The specific workflow is as follows: In the main training loop of PI-GAN, the generator G is responsible for reconstructing the missing electrical parameters based on partial observation data.

[0107] Input: The input vector consists of random noise z and the control signal at the current time. (such as driving pulses) and historical observation windows composition.

[0108] Output: The generator outputs the missing electrical data sequence to be completed (e.g., MOSFET drain current). or drain-source voltage ):

[0109] To assess the physical rationality of the generated data, instead of directly using circuit topology equations, the generated data is mapped to the thermal domain using the electro-thermal energy conversion law, and verified through a surrogate model, thereby achieving physical consistency constraint mapping.

[0110] In electro-thermal conversion, the corresponding instantaneous power loss is first calculated. For example, for a MOSFET, the conduction loss is calculated according to Joule's law:

[0111] The calculated power sequence Input into the proxy model frozen in step 1 Since the surrogate model is differentiable, it will output a predicted temperature sequence that implicitly contains the law of thermal inertia: .in, The value depends entirely on the current generated by the generator. The two are related through complex nonlinear physical laws.

[0112] The system's overall optimization objective consists of adversarial loss and physical consistency loss, and the generator parameters are updated using the backpropagation algorithm. Physical consistency loss Used to calculate the temperature predicted by the surrogate model Compared with the actual observed temperature of the system The difference between (or the generated temperature data) forces the generator to produce a temperature rise that conforms to physical facts:

[0113] The total loss function is defined as:

[0114] Gradient flow during backpropagation It will pass through the frozen surrogate model (using the chain rule), calculate the partial derivative of temperature with respect to power, and then pass it back to the current:

[0115] Through this mechanism, the generator can "learn" to generate high-quality data that conforms to electro-thermal physical constraints, without knowing the specific differential equations, simply by being guided by gradients.

[0116] If the amount of valid data is greater than or equal to an acceptable threshold, then data dispersion analysis is performed. Traditional ANOVA methods can measure the average deviation of data points from the mean, but are highly sensitive to outliers. Information entropy measures the uncertainty or disorder of data distribution; the more uniform and unpredictable a distribution, the higher its entropy value. Introducing information entropy allows us to assess the disorder or uncertainty of data by calculating its "information entropy" within a sliding window, thus providing a more robust assessment of data dispersion.

[0117] First, define the following symbols: It is a sequence of N observations from a certain sensor within a time window.

[0118] The information entropy (dispersion index) of sequence X.

[0119] High dispersion threshold (adaptively adjustable).

[0120] Low dispersion threshold (adaptively adjustable).

[0121] B is the number of intervals used to calculate entropy (usually taken as...). or (i.e., Sturges' formula).

[0122] For the data X within the window, its range is divided into B equal-width intervals (histogram binning), and the frequency of data points within each interval is counted. The probability estimate for each interval (the b-th interval) is calculated:

[0123] in, It is the number of data points that fall within the b-th interval.

[0124] Calculate the information entropy of the data distribution within this window according to the definition of Shannon entropy:

[0125] Among them, if This indicates that the data is almost entirely concentrated within a certain interval, with extremely low dispersion and is very stable. If A large value indicates that the data is evenly distributed across all intervals, with high dispersion and is very chaotic.

[0126] The calculated entropy value H(X) is compared with the dynamic threshold, and the discrete state S is output:

[0127] To enable the system to adapt to different sensors and operating conditions, the threshold and It should not be a fixed value. The embodiments in this specification are designed to use it as a statistic based on historical window entropy values:

[0128]

[0129] in: It is the moving average of the entropy values ​​over the past M time windows. It is the moving standard deviation of the entropy values ​​over the past M time windows.

[0130] Step 104: Extract fusion features from the historical running data and the running data.

[0131] Specifically, once the data dispersion analysis is completed, subsequent data processing can be carried out based on the analysis results.

[0132] First, define the input. Let be a sliding window data matrix containing N sensors, each with its most recent M time-step readings. This is the original input to the network.

[0133] Let t be the current observation vector of all N sensors (i.e., (The last column of the matrix).

[0134] By analyzing input data Generate a gate value This is used to determine the weights of the fusion strategy.

[0135] Specifically, a shared neural network (such as a one-dimensional CNN) can be used to... Extract fusion features from them.

[0136]

[0137] in, It is the extracted fusion feature vector, which encodes the joint state and discreteness of the current multi-sensor data. , These represent the parameter weights and biases of a one-dimensional CNN neural network.

[0138] Features Gating values ​​are generated using a fully connected layer and a sigmoid activation function.

[0139] here It is the Sigmoid function, which compresses the output to the (0,1) interval.

[0140] After the gate value is generated, for low-dispersion data, it can be used as the current observation value. Generate a set of adaptive weights and perform a weighted average. Specific steps: (1) Weight generation Use a lightweight quantum network (such as a single-layer fully connected + Softmax) to extract features Generate a weight vector.

[0141]

[0142] in, ,and The Softmax function ensures that all weights are positive and sum to 1.

[0143] (2) Fusion computing Calculate the weighted average:

[0144] It is the output of the weighted average branch.

[0145] For highly discrete data, an improved Kalman filter driven by a hypernet is used for processing. In addition, missing data generated by a data generation model based on physical model constraints can also be processed by an improved Kalman filter driven by a hypernet.

[0146] Step 106: Process the highly discrete data in the running data and the prediction results of the running data at the previous time step using an adaptive Kalman filter driven by a hypernetic network to obtain the prediction results of the running data at the current time step.

[0147] In an optional embodiment, the integrated circuit power supply redundancy self-healing control method further includes: The fused features are input into the supernetwork for mapping processing to generate an adaptive observation noise covariance matrix and an adaptive process noise covariance matrix, wherein the fused features include joint state features and discreteness features; The adaptive process noise covariance matrix and the posterior error covariance matrix of the previous time step are input into the Kalman filter for prediction processing to obtain the prior error covariance matrix of the current time step. Determine the error between the running data and the predicted running data at the current time, and calculate the Kalman gain based on the adaptive observation noise covariance matrix and the prior error covariance matrix at the current time; Based on the Kalman gain and the error, the prediction results of the current running data are corrected.

[0148] A schematic diagram of a data fusion process provided in the embodiments of this specification is shown below. Figure 4 As shown.

[0149] Specifically, as a "parameter generator," the hypernetwork does not directly output the prediction results, but instead uses the extracted fusion features. As input, it dynamically outputs the noise covariance matrix required for Kalman filtering. The "joint state" and "discretion" of the multi-sensor data at the current moment are encoded. This means that the input to the hypernetwork is not just the raw data, but high-dimensional features that have been abstracted and contain the data distribution characteristics. The hypernetwork is used to dynamically generate the key parameters of the Kalman Filter (KF), making it adaptable to the current data features. In other words, the essence of hypernetworks is to integrate the perceptual capabilities of deep learning (through...) The discreteness of the sensed data is transformed into parameters for traditional control algorithms. and (Matrix). Through this construction, the model can automatically increase the noise parameter when the data fluctuates drastically (high dispersion), making the Kalman filter more confident in the predicted values; conversely, it will be more confident in the observed values.

[0150] The specific steps are as follows: (1) Generation of hypernetwork parameters Specifically, the hypernetwork is constructed as a mapping function. The Kalman filter is responsible for mapping the high-dimensional feature space to the parameter space of the Kalman filter. To reduce computational complexity and the number of parameters, the network structure can be designed as having "3 fully connected layers," and the output covariance matrix is ​​a diagonal matrix, meaning it is not necessary to output the entire... For a matrix, only the diagonal elements need to be output. Each element. Furthermore, the network output layer uses the Softplus activation function to ensure that its output value is always positive. The specific formula is: .

[0151] The output of the hypernetwork consists of two time-varying matrices used to adjust the behavior of the Kalman filter in real time. Output 1 is the adaptive observation noise covariance matrix. Its dimensions are Its function is to tell the Kalman filter how much noise the "observed data" is at the current moment (i.e., how unreliable the data is). And output two: the adaptive process noise covariance matrix. Its dimensions: (in (For the state vector dimension). Its generation logic is the same as... It is exactly the same, used to tell the Kalman filter how much uncertainty there is in the "state prediction" at the current moment.

[0152] (2) Differentiable Kalman Filter This is a differentiable computation module. The core logic of this module is to dynamically adjust the hypernetwork. and Two key noise parameters are used to calculate the optimal system state using the standard Kalman filter formula (prediction + update). and use it directly as the output of that branch. This information is provided to subsequent fusion modules. The standard KF steps are performed using the parameters generated by the supernetwork. Specifically, this consists of a prediction step and an update step.

[0153] Phase 1: Prediction Steps.

[0154] The goal of this stage is to deduce the prior state and prior covariance of the current moment based on the state of the previous moment.

[0155] 1) State prediction:

[0156] Using the state transition matrix (Preset parameters) Estimate the posterior state from the previous time step. Mapping to the current time, we obtain the prior state estimate. .

[0157] 2) Covariance prediction:

[0158] Used to deduce the current prior error covariance matrix This step introduces an adaptive process noise covariance matrix generated by the hypernetwork. This represents the uncertainty in the model's predictions.

[0159] Phase Two: Update Steps.

[0160] The goal of this stage is to refine the prediction results by incorporating current observation data (sensor readings).

[0161] 1) Calculate the measurement residuals:

[0162] Used to calculate actual observed values Compared with predicted observations The differences between them. It is the observation matrix.

[0163] 2) Calculate the Kalman gain.

[0164] Kalman Gain The calculation of the model is the key link between "prediction" and "update". It determines to what extent the system should trust the current observation data (sensor readings) and to what extent it should retain the model's predictions when correcting the state.

[0165] Step 1: Calculate the residual covariance .

[0166] First, we need to calculate the prediction error covariance of the observed values. This matrix represents the sum of uncertainties in the difference between observed data and their predicted values.

[0167]

[0168] Parameter meaning: The observation matrix (with preset fixed parameters). This is the prior error covariance matrix obtained from the prediction step. This is the adaptive observation noise covariance matrix. This is a key parameter dynamically generated by the hypernetwork based on the current data characteristics. If... The value is very large (indicating high sensor noise), and the calculated value is... It will also get bigger.

[0169] Step 2: Calculate the inverse of the covariance matrix .

[0170] To calculate the gain, it is necessary to... Perform the inversion operation. To ensure that the entire network (including the supernetwork and Kalman filter) can undergo end-to-end backpropagation training, allowing gradients to be smoothly propagated back to the supernetwork to update parameters, The calculation must use differentiable matrix inversion. Cholesky decomposition can be used because the matrices in Kalman filtering are typically symmetric positive definite matrices, and the relevant constraints have already been applied. Residual covariance matrix. Depend on The calculated result indicates that it is theoretically a symmetric positive definite matrix. First, matrix regularization is performed. Next, perform Cholesky decomposition to decompose the matrix into a lower triangular matrix. And its transpose. This operation is differentiable in mainstream deep learning frameworks: ,Right now Then, solving the linear equation system replaces direct inversion to avoid numerical oscillations. This is done to obtain the Kalman gain. Using the already calculated The solution is obtained through two substitutions. This is equivalent to solving a system of linear equations. . Essentially, it's a weighting coefficient; when observation noise... When it is very large, Get bigger Get smaller It becomes smaller. The model reduces its reliance on observations when updating the state. The weighting of prior predictions is increased. When observing noise When I was very young, Get smaller Get bigger As the value increases, the model will increase the amount of data compared to the observed values ​​when updating the state. The weights are adjusted to quickly correct the state to closely match the sensor readings.

[0171] If used This represents the output of the Kalman filter branch. This output will then be combined with the output of the weighted average branch. Perform gating fusion.

[0172] (3) Gated fusion output The specific steps are as follows:

[0173] in, This is the final output of the network. Through soft hybridization, the smoothness of the fusion process during policy switching is ensured, and the entire network is end-to-end differentiable.

[0174] Step 108: Based on the running data, the fusion features, and the running data prediction results at the current moment, perform multi-task state evaluation based on the attention mechanism to generate evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply.

[0175] In an optional embodiment, the integrated circuit power supply redundancy self-healing control method further includes: Based on the fusion features, a weighted average is performed on the low-dispersion data in the running data to generate the corresponding processing result; The processing results and the current time-based operational data prediction results are then subjected to gating fusion to generate a corresponding fusion result; Based on the operational data, the fusion features, and the fusion results, a multi-task state evaluation based on an attention mechanism is performed to generate evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply.

[0176] In another optional implementation, the step of performing a multi-task state evaluation based on the operating data, the fused features, and the current operating data prediction results, and generating evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply, includes: The fused features are mapped to query vectors, and the running data is mapped to keywords and key values; The query vector, the keyword, and the key value are subjected to context enhancement processing based on the cross-attention mechanism to generate context enhancement features; The context enhancement features are input into the physical parameter evaluation layer in the two-level cascaded network structure for processing, generating evaluation results corresponding to the output voltage and ripple voltage of the secondary power supply, respectively. The context enhancement features and the evaluation results corresponding to the output voltage and ripple voltage of the secondary power supply are input into the health state inference layer in the two-level cascaded network structure for processing, thereby generating the evaluation results corresponding to the health state of the secondary power supply.

[0177] Furthermore, the integrated circuit power supply redundancy self-healing control method further includes: Obtain the parameter values ​​of at least two uncertainty parameters output by the physical parameter evaluation layer and the health status inference layer, wherein the at least two uncertainty parameters are used to evaluate the uncertainty of the evaluation result; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the evaluation result of the output voltage output by the physical parameter evaluation layer, calculate the first loss value of the normal-inverse gamma distribution negative log-likelihood loss function; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the evaluation result of the ripple voltage output by the physical parameter evaluation layer, calculate the second loss value of the normal-inverse gamma distribution negative log-likelihood loss function; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the health status assessment result output by the health status inference layer, calculate the third loss value of the normal-inverse gamma distribution negative log-likelihood loss function; The first loss value, the second loss value, and the third loss value are weighted and calculated to obtain the corresponding total loss value; The network parameters of the two-level cascaded network structure are optimized based on the total loss value.

[0178] Specifically, after data fusion, multi-task state evaluation based on an attention mechanism can be performed. The basic idea is: after fusion processing by GHF-Net, the best fused estimate (such as output current or voltage) is input into a multi-task learning module, and by introducing an attention mechanism, the importance weights of different evaluation tasks are dynamically assigned. Specific tasks include secondary power supply output voltage evaluation, ripple voltage evaluation, and health status scoring. However, to address the problems of neglecting the physical causal dependencies between tasks and lacking prediction confidence quantification in traditional multi-task learning, this specification proposes a hierarchical causal evaluation framework. This framework can not only dynamically calibrate features based on operating condition context, but also synchronously output state prediction values ​​and their uncertainty boundaries through Evidential Deep Learning (EDL). The specific implementation steps are as follows: 1) Enhanced working context based on cross-attention.

[0179] To address the problem that simply splicing control signals cannot effectively capture the dynamic nonlinear response of the system, a cross-attention mechanism is first introduced, which dynamically injects control signals (such as PWM duty cycle and load setting) into the state features as a "context".

[0180] Input definition: make Let represent the fused feature vector extracted by the preceding network (such as GHF-Net); , which represents the current system control signal vector.

[0181] Attention calculation: The fused features are mapped to query vectors, and the control signals are mapped to keywords and values:

[0182]

[0183] in, These are the learnable linear projection matrices. This is a scaling factor used to prevent the gradient from vanishing due to an excessively large dot product. This is a context-enhanced feature for the output. The original state features are recalibrated based on the current control strategy (e.g., high-frequency heavy load or low-frequency light load) to include operating condition information.

[0184] 2) Construct a causal hierarchical task decoupling structure.

[0185] Unlike traditional parallel multitasking output, the embodiments in this specification are based on "physical state deterioration". The objective causal law of "declining health" was used to construct a two-level cascaded network structure.

[0186] Level 1: Physical parameter evaluation layer.

[0187] This layer focuses on evaluating directly measurable physical quantities, including: output voltage. and ripple voltage Networks utilize enhanced features Extracting high-dimensional features specific to the physics task:

[0188] The formula does not directly output the final value, but retains the intermediate feature vectors. For use by the next level.

[0189] Level 2: Health Status Deduction Layer.

[0190] This layer is used to evaluate abstract health indicators. To simulate causal relationships, its input not only includes the original contextual features but also explicitly concatenates the physical features extracted from the previous level:

[0191]

[0192] Through this cascaded design, the health assessment branch can directly "sense" specific physical anomalies such as voltage instability or excessive ripple, thereby making the health score more physically interpretable.

[0193] 3) Uncertainty distribution prediction based on evidence theory.

[0194] To address the problem of "blind confidence" in the model, the output layer of each task is no longer a single scalar regression, but rather outputs four key parameters of a normal-inverse gamma distribution based on evidence theory.

[0195] Output layer definition: For any task (For example, voltage prediction), the output head is a fully connected layer containing 4 neurons, and the parameters are ensured to be positive through the Softplus activation function:

[0196] in, This represents the expected value of the predicted value. The evidence count, representing the observations, reflects the model's familiarity with the current input. The parameters controlling the shape of the distribution are used to derive the variance.

[0197] Based on these four parameters, two types of uncertainty can be decoupled: ① Cognitive uncertainty:

[0198] The above formula shows that the larger the value, the more likely the model has never seen this type of data (belonging to out-of-distribution samples, OOD), and the prediction results are unreliable.

[0199] ② Random uncertainty:

[0200] The above formula shows that this value reflects the inherent noise level of the data itself.

[0201] 3) Evidence loss function and joint optimization.

[0202] The training objective of this network is to maximize the model evidence from the observed data. End-to-end training is performed using a normal-inverse gamma distribution negative log-likelihood loss (NLL Loss) combined with an evidence regularization term. For each task, the loss function is defined as follows:

[0203] Among them, the NLL term: prompts the predicted distribution to fit the true label as closely as possible. Regularization term: When the prediction error... When the amount is large, mandatory punishment evidence count This forces the model to admit "I don't know" (outputting high uncertainty) rather than giving incorrect, confident predictions. Multi-task weighting: The final total loss is the weighted sum of the losses from each task.

[0204] By minimizing this loss, the model can simultaneously learn to accurately predict state values ​​and report its prediction risks.

[0205] Step 110: Based on the evaluation results and reinforcement learning algorithm, update the parameters of the PID controller, calculate and output control signals through the updated PID controller, and perform self-healing control on the output power of the secondary power supply.

[0206] In one optional implementation, updating the parameters of the PID controller based on the evaluation results and the reinforcement learning algorithm includes: A state space vector is constructed based on the evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply, as well as the rated voltage of the secondary power supply. The state space vector is input into the policy network in the reinforcement learning framework for processing to generate an action space vector, wherein the action space vector is composed of the PID parameter adjustment amount of the PID controller. The parameters of the PID controller are updated based on the action space vector.

[0207] Furthermore, the integrated circuit power supply redundancy self-healing control method further includes: The reward function value is calculated based on the state space vector and action space vector at the current moment; Determine the state space vector for the next moment, and store the current state space vector, action space vector, the next state space vector, and the function value into the experience replay buffer. Experience data is periodically collected in the experience replay buffer at a preset frequency, and the network parameters of the policy network and value network in the reinforcement learning framework are updated based on the experience data using the stochastic gradient descent algorithm.

[0208] A schematic diagram of a power self-healing control process provided in the embodiments of this specification is shown below. Figure 5 As shown.

[0209] Specifically, after generating the above evaluation results, output power self-healing control based on reinforcement learning and fuzzy PID (SARL-FPIDT) can be performed.

[0210] Traditional fuzzy PID controllers have fixed parameters (such as membership functions and rule tables). The SARL-FPIDT method places the fuzzy PID controller within a reinforcement learning framework, treating it as an environment and the mechanism for adjusting its internal parameters as an agent. The agent makes decisions (actions) based on the current system state (provided by GHF-Net and the state evaluation module) (how to adjust the PID parameters), then obtains rewards (rewards) from the environment (the quality of the control effect), and learns the optimal adjustment strategy through continuous trial and error.

[0211] (1) Define the state space state space vector A comprehensive description of the system's current state is needed, which comes directly from the output of the preceding modules:

[0212] in, Tracking error. It is the rated voltage of the secondary power supply. This is the evaluation result of the output voltage. Error change rate reflects the dynamic trend of error. The ripple voltage evaluation result represents the current control quality of the system. The health status assessment results represent the long-term reliability of the system. By incorporating the ripple voltage and health status assessment results into the state space vector, the agent can not only pursue tracking accuracy but also actively suppress ripple and delay system degradation, achieving true "self-healing".

[0213] (2) Define the action space Action space vector It is the amount by which the agent adjusts the parameters of the fuzzy PID controller:

[0214] in, These are the incremental adjustment values ​​for the proportional, integral, and derivative parameters of the PID controller, respectively.

[0215] The action space vector is continuous, and the range of each incremental change can be set to... .

[0216] The adjusted PID parameters are as follows:

[0217]

[0218]

[0219] In the formula, The parameters from the previous time step, or the baseline parameters given by the fuzzy rule table. Actions output by the policy network (Actor network) That is, the incremental adjustment value that the agent decides based on the current state. : The final proportional, integral, and differential coefficients actually used in the calculation at the current moment.

[0220] (3) Design the reward function reward function It is the core of reinforcement learning design, guiding the agent's learning direction. It is a multi-objective optimization function: in, To penalize tracking errors, reducing steady-state error is encouraged. To penalize drastic changes in error, a smooth response is encouraged, and overshoot is reduced. To penalize ripple, the controller is encouraged to actively select parameters that suppress ripple. This indicates a decrease in health as a punishment. The lower the value, the greater the penalty, encouraging the controller to adopt strategies beneficial to system health (such as avoiding frequent, large-amplitude control actions). This indicates that large parameter adjustments are penalized, while smooth parameter changes are encouraged to avoid overly abrupt control strategies. These are weighting coefficients used to balance the importance of different objectives.

[0221] As can be seen from the reward function, it deeply integrates secondary power supply performance indicators (error, ripple) and reliability indicators (health), making the learned control strategy not only high-performance but also sustainable and "healthy". Specifically, the function simultaneously penalizes tracking error, error variation (stability), ripple (quality), health degradation (reliability), and large parameter adjustments (smoothness), guiding the agent to learn the optimal multi-objective balance strategy.

[0222] (4) Selecting a reinforcement learning algorithm The Soft Actor-Critic (SAC) Actor-Critic algorithm is employed to solve the problem of continuous states and action spaces. This is achieved through an Actor network (policy network). ), according to the state Select Action Through Critic networks (value networks) The state-action pairs are evaluated for quality. Through continuous interaction and policy gradient updates, an optimal policy is eventually learned. It can map any state s to an action a that yields the maximum cumulative reward.

[0223] 1) Actor Network (Policy Network) Architecture: Input layer: Multidimensional state vectors (e.g., 7-dimensional): .

[0224] Feature extraction layer: Branch A (Time Series Features): Processes error history and health trend.

[0225] Structure: 1-layer GRU (Gated Recurrent Unit) or LSTM.

[0226] Scale: 32 or 64 hidden units (compared to LSTM, GRU has fewer parameters, faster inference, and is more suitable for embedded control).

[0227] Branch B (Instantaneous Feature): Processes the current error value and duty cycle.

[0228] Structure: 1-2 fully connected / dense layers.

[0229] Size: 64 neurons per layer, ReLU activation function.

[0230] Feature fusion layer: combines the outputs of branch A and branch B.

[0231] Decision-maker: Mean Head (Mu Head): 1 fully connected layer -> Output action dimension (3-dimensional): )->Tanh activation function (limits the output to the [-1,1] interval).

[0232] Log_Std Head: 1 fully connected layer -> output action dimension (3D) -> Softplus activation function (ensuring the standard deviation is positive).

[0233] Output: Based on and The final action is obtained by sampling from a Gaussian distribution. .

[0234] 2) Critic Network (Value Network) Architecture: Input layer (concatenation layer): directly inputs the state vector and action vectors Concatenation is used as input.

[0235] Hidden layer: Structure: 3-layer fully connected network (MLP, standard practice), because the Critic does not need to output a sequence, but only needs to evaluate "whether the strategy is good in the current situation".

[0236] Size: Configured as [256,256,256] or lightweight [128,128].

[0237] Activation function: ReLU or Leaky ReLU.

[0238] Output layer: Structure: A fully connected layer with one neuron.

[0239] Activation function: None (Linear), directly outputs the real number Q value.

[0240] The parameters in the above design are kept to the level of several thousand to tens of thousands, so that it can be used in power electronic control loop processes and FPGA edge computing modules, achieving millisecond-level or even lower performance. Real-time reasoning at the highest level.

[0241] (5) Online learning and control.

[0242] 1) Initialization: Initialize the parameters of the Actor and Critic networks, and initialize the parameters of the fuzzy PID controller.

[0243] 2) Interactive loop (within each control cycle t): ① Observation status: Read from the status assessment module .

[0244] ② Generate an action: Input the Actor network to obtain the action (parameter adjustment amount). .

[0245] ③ Execution action: Adjust the parameters of the fuzzy PID controller.

[0246] ④ Apply control: The fuzzy PID controller applies control based on the adjusted fuzzy PID controller parameters and the current error. The control signal u(t) is calculated and output to the power converter.

[0247]

[0248] in, The target output voltage, also known as the rated voltage (the voltage value that the system expects to achieve). The output voltage is the evaluation result. To track errors and reflect the degree to which the current voltage deviates from the target. The error change rate reflects how fast and how much the voltage error changes (e.g., whether it is rapidly moving away from the target or approaching the target).

[0249] ⑤ Observing a new state: The system evolves to the next state. and received a reward .

[0250] ⑥ Storage experience: Store in the experience replay buffer.

[0251] ⑦ Network Update: Periodically sample a batch of empirical data from the buffer and update the parameters of the Actor and Critic networks using stochastic gradient descent. Specifically, for Critic updates, minimize the temporal difference error to improve the accuracy of value estimation. For Actor updates, adjust the policy parameters in the direction of increasing the expected return (i.e., the Critic's output). .

[0252] The integrated circuit power supply redundancy self-healing control method provided in the embodiments of this specification has high monitoring accuracy and reliability. By fusing data from multiple sensors, it effectively overcomes the errors and noise effects of a single sensor, providing monitoring accuracy far exceeding that of traditional methods. A dynamic data verification method is designed to dynamically adjust the monitoring threshold. Even if some sensors fail, the monitoring system can still work reliably, which is beneficial to improving the system's dynamic self-adaptation and accurate assessment capability of secondary power supply status in complex data scenarios.

[0253] Furthermore, the integrated circuit power supply analytical redundancy self-healing control method provided in the embodiments of this specification realizes the transformation from "traditional algorithm" to "AI enhancement", deeply integrating deep learning, generative models and fuzzy inference methods, designing a Kalman filter data fusion network based on hypernetwork, proposing a data generation fusion method based on physical constraints and generative adversarial networks and an output power self-healing control method based on reinforcement learning and fuzzy PID, and constructing an end-to-end, adaptive, highly learnable intelligent fusion system to realize online adaptive adjustment and control of secondary power module parameters.

[0254] In addition, the integrated circuit power supply redundancy self-healing control method provided in the embodiments of this specification realizes closed-loop self-healing of "sensing-diagnosis-decision-execution", without requiring any expert experience to manually adjust PID parameters. The controller can self-optimize online in real time. Furthermore, multi-objective optimization control can be performed, forming a "self-healing" system capable of self-sensing, self-evaluation, and self-adjustment by optimizing tracking accuracy, dynamic performance, ripple suppression, and system health.

[0255] This embodiment acquires the current operating data of the secondary power supply from various sensors in the aging bench monitoring system, as well as historical operating data from multiple historical time points. It extracts fusion features from the historical and current operating data. High-discreteness data in the operating data and the prediction results of the previous time point's operating data are processed to obtain the current time's operating data prediction result. Based on the operating data, fusion features, and the current time's operating data prediction result, a multi-task state assessment is performed, generating assessment results for the secondary power supply's output voltage, ripple voltage, and health status. Based on the assessment results, the PID controller's parameters are updated, and the updated PID controller calculates and outputs a control signal to perform self-healing control of the secondary power supply's output power. This embodiment effectively overcomes the errors and noise effects of a single sensor by fusing data from multiple sensors, thus improving the monitoring accuracy of the operating data. Furthermore, by processing the high-discreteness data through a Kalman filter data fusion network driven by a hypernetic network, and performing self-healing output power control based on reinforcement learning algorithms and a fuzzy PID controller, it not only facilitates accurate assessment of the secondary power supply system's state but also enables online adaptive adjustment and control of relevant secondary power supply parameters, thereby ensuring the stable operation of the secondary power supply system.

[0256] One embodiment of this specification also provides a method for assessing the health of a multi-channel secondary power supply on an integrated circuit aging bench, which specifically includes the following steps.

[0257] Step S1: During the real-time operation phase of the integrated circuit aging bench, collect multi-dimensional real-time data of the multi-channel secondary power supply of the integrated circuit aging bench under the current operating conditions; In one embodiment of this specification, after collecting multi-dimensional real-time data of the multi-channel secondary power supply of the integrated circuit aging bench under the current operating conditions, the method further includes: generating multiple real-time data segments with partial overlap by sliding slices of the multi-dimensional real-time data of the multi-channel secondary power supply under the current operating conditions according to a fixed window size, and obtaining a set of standardized real-time data segments by independently performing Z-Score standardization processing on each dimension of each real-time data segment.

[0258] Step S2: By inputting the multi-dimensional real-time data of the multi-channel secondary power supply under the current operating condition into the pre-trained deep neural network model, the fused feature vector, reconstructed waveform and prediction variance value of the multi-channel secondary power supply under the current operating condition are obtained. In one embodiment of this specification, the deep neural network model includes a hybrid spatiotemporal encoder, a reconstruction decoder, and an uncertainty estimation head; wherein the hybrid spatiotemporal encoder includes a temporal encoder and a spatial encoder.

[0259] In one embodiment of this specification, the step of inputting multi-dimensional real-time data of the multi-channel secondary power supply under the current operating condition into a pre-trained deep neural network model to obtain the fused feature vector, reconstructed waveform, and prediction variance of the multi-channel secondary power supply under the current operating condition includes: inputting the multi-dimensional real-time data of the multi-channel secondary power supply under the current operating condition into a time-domain encoder in the pre-trained deep neural network model to extract time-domain features, thereby obtaining a time-domain feature vector of the multi-channel secondary power supply under the current operating condition; inputting the multi-dimensional real-time data of the multi-channel secondary power supply under the current operating condition into a spatial encoder in the pre-trained deep neural network model to extract spatial features, thereby obtaining a spatial feature vector of the multi-channel secondary power supply under the current operating condition; and concatenating the time-domain feature vector output by the time-domain encoder with the spatial feature vector output by the spatial encoder to obtain the fused feature vector of the multi-channel secondary power supply under the current operating condition.

[0260] In one embodiment of this specification, the step of inputting the multi-dimensional real-time data of the multi-channel secondary power supply under the current operating condition into a pre-trained deep neural network model to obtain the fused feature vector, reconstructed waveform, and prediction variance value of the multi-channel secondary power supply under the current operating condition includes: inputting the fused feature vector of the multi-channel secondary power supply under the current operating condition into the reconstruction decoder in the pre-trained deep neural network model to obtain the reconstructed waveform of the multi-channel secondary power supply under the current operating condition.

[0261] In one embodiment of this specification, the step of obtaining the fused feature vector, reconstructed waveform, and predicted variance of the multi-channel secondary power supply under the current operating condition by inputting the multi-dimensional real-time data of the multi-channel secondary power supply under the current operating condition into a pre-trained deep neural network model includes: obtaining the predicted variance of the multi-channel secondary power supply under the current operating condition by inputting the fused feature vector of the multi-channel secondary power supply under the current operating condition into the uncertainty estimation head of the pre-trained deep neural network model.

[0262] Step S3: Using the fused feature vector, reconstructed waveform, and predicted variance of the multi-channel secondary power supply under the current operating conditions, obtain the health assessment results of the multi-channel secondary power supply of the integrated circuit aging bench.

[0263] In one embodiment of this specification, obtaining the health assessment result of the multi-channel secondary power supply of the integrated circuit aging bench using the fused feature vector, reconstructed waveform, and prediction variance value of the multi-channel secondary power supply under the current operating condition includes: calculating the Mahalanobis distance between the fused feature vector of the multi-channel secondary power supply under the current operating condition and a pre-built health fingerprint database, and converting the Mahalanobis distance into a consistency health index; calculating the root mean square error between the reconstructed waveform and the original waveform of the multi-channel secondary power supply under the current operating condition, and converting the root mean square error into a reconstruction health index; calculating the dynamic weighting coefficient for assessing the current operating state of the integrated circuit aging bench using the prediction variance value of the multi-channel secondary power supply under the current operating condition; and calculating the comprehensive health index of the multi-channel secondary power supply under the current operating condition using the consistency health index, reconstruction health index, and dynamic weighting coefficient, and using the comprehensive health index as the health assessment result of the multi-channel secondary power supply of the integrated circuit aging bench.

[0264] The multi-channel secondary power supply health assessment scheme for the integrated circuit aging bench provided in this specification has the following advantages: (1) It has a deep channel correlation perception capability and avoids "island-style" assessment: The introduction of graph attention network (GAT) and mask collaborative reconstruction mechanism breaks the limitation of traditional methods that independently assess a single channel. The model not only focuses on the waveform distortion of a single channel, but also can keenly capture the abnormal linkage of "inconsistent pace" between multiple channels (such as abnormal current sharing and parallel imbalance), so as to identify early faults where the single channel indicators seem normal but the system has hidden dangers. (2) It uses a dynamic fusion strategy with uncertainty perception to significantly improve robustness: In response to the common load changes and electromagnetic interference in industrial sites, this specification uses Bayesian uncertainty estimation to quantify the "confidence" of the model in real time. Through the dynamic weight generation mechanism, the system can smoothly switch between "high sensitivity consistency detection under steady state" and "high robustness reconstruction detection under transient interference", which effectively solves the problem of false positives caused by operating condition drift in traditional methods. (3) A dual verification closed loop of "feature space + physical waveform" is constructed: This specification outputs a consistency index (CHI) based on statistical distribution in parallel. C ) and the physical signal-based reconstruction error index (CHI) R The former is sensitive to minor degradation, while the latter is reliable for serious faults. The two complement and verify each other, greatly improving the credibility and interpretability of the evaluation results. (4) It has extremely high sensitivity to early minor faults: By extracting high-dimensional spatiotemporal features through a hybrid spatiotemporal encoder and combining them with Mahalanobis distance statistical measures, this manual can amplify the minute offsets that are difficult to detect in the original signal in the feature space, thereby achieving predictive maintenance before the power supply performance deteriorates substantially.

[0265] The main purpose of the embodiments in this specification is to provide a method for health assessment of multi-channel secondary power supplies in an integrated circuit aging bench. This method aims to solve the problems of traditional methods in assessing multi-channel secondary power supplies, such as ignoring electrical coupling between channels, weak anti-interference capability under atypical operating conditions, and insufficient reliability of single assessment indicators. This will enable accurate identification and robust assessment of early minor faults in the secondary power supply system.

[0266] To achieve the above objectives, this specification provides an embodiment of a multi-channel secondary power supply health assessment method for an integrated circuit aging bench. This method includes two stages: offline model construction and training, and online real-time health assessment.

[0267] In the offline model building and training phase: First, multi-dimensional time-series data of multi-channel power supplies under different load conditions are collected and processed by sliding slicing and Z-score normalization. Second, a deep neural network including a hybrid spatiotemporal encoder, a reconstruction decoder, and an uncertainty estimation head is constructed and trained. The hybrid spatiotemporal encoder includes a temporal encoder and a spatial encoder: the temporal encoder uses a one-dimensional convolutional network (1D-CNN) and a Transformer encoder to extract high-frequency ripple and long-range temporal dependencies of a single channel; the spatial encoder uses a graph attention network (GAT) to construct a fully connected graph based on the electrical connections between channels, adaptively aggregating neighborhood information to capture the topological correlation features between channels. Simultaneously, a masked collaborative reconstruction mechanism is used for model training. By randomly masking some channel data, the model is forced to use the features of the remaining healthy channels and the learned inter-channel correlation weights to deduce the masked waveforms and simultaneously output the predicted uncertainty variance. Finally, steady-state data with low reconstruction errors are selected to construct a health fingerprint database, and the mean vector and covariance matrix in the latent feature space are calculated as health benchmarks.

[0268] During the online real-time evaluation phase: the real-time collected data is input into the trained deep neural network to perform a dynamic fusion evaluation based on two indicators: First, the Consistency Health Index (CHI) is calculated. C First, the Mahalanobis distance between the latent feature vectors (i.e., fused feature vectors) of real-time data and the health fingerprint database is used to quantify the degree of consistency between the current state and the healthy distribution. Second, the reconstruction error health index (CHI) is calculated. RThe first step involves quantifying the deep neural network's interpretability of the current data by using the root mean square error between the reconstructed waveform and the original waveform, based on the output of the deep neural network. The second step is to perform adaptive fusion based on uncertainty awareness: the uncertainty estimation head of the deep neural network outputs the prediction variance in real time, generating dynamic weight coefficients through nonlinear mapping. When the system is in a steady-state, high-confidence scenario, the focus is on consistency health indicators. When the system encounters transient disturbances or atypical operating conditions leading to a surge in uncertainty, the dependence on feature distance is automatically reduced, and the focus shifts to reconstruction error indicators, ultimately generating a comprehensive health index (CHI). The original waveform is obtained by preprocessing real-time data.

[0269] The embodiments in this specification provide a complete and rigorous technical loop, from data processing, construction and training of deep neural network models to multidimensional evaluation. This loop can provide unprecedented in-depth health insights for critical power systems, effectively overcome the limitations of traditional methods, and provide a scientific basis for the maintenance and management of power systems.

[0270] One embodiment of this specification provides a method for assessing the health of a multi-channel secondary power supply on an integrated circuit aging bench, including: Step 1. Data acquisition and preprocessing; Multi-dimensional time-series data of a multi-channel power supply under different load points (no load, 25% load, 50% load, 75% load, and full load) are collected, including output voltage waveforms, output current waveforms, input voltage, input current, and temperature of key modules. The sampling frequency is at least 10-20 times the power supply switching frequency to fully capture high-frequency details, ultimately resulting in a multi-channel time-series dataset. .

[0271] Collected long sequence data Slicing is performed using a fixed window size (2048 sampling points) to generate partially overlapping data segments. Z-score standardization (subtracting the mean and dividing by the standard deviation) is performed independently on each dimension of each segment, and the standardization parameters (mean and standard deviation) are saved for subsequent online use. This results in a processed set of data segments. .

[0272] Step 2. Offline model building and training; 2.1 Offline model construction based on spatiotemporal graph attention; Traditional methods extract timing features from a single channel independently, neglecting the electrical coupling characteristics of multi-channel power supplies operating in parallel or cascaded configurations. This step constructs and trains a deep neural network using a standardized sample set. The network takes a high-dimensional feature vector (i.e., fused feature vector), a reconstructed waveform, and a predicted variance as input. It consists of three parts: a hybrid spatiotemporal encoder responsible for spatiotemporal feature extraction, a reconstruction decoder responsible for waveform reconstruction, and an uncertainty estimator head responsible for confidence assessment.

[0273] Hybrid spatiotemporal encoder construction: This process aims to extract deep features z that contain "temporal evolution" and "channel coupling". The encoder contains two parallel branches, specifically including: Temporal feature branch: Continuing with the "local convolution-global self-attention" architecture. The input sequence for each channel i... Local high-frequency ripple features are extracted using 1D-CNN, and then long-range temporal dependencies are extracted using TransformerEncoder to obtain temporal feature vectors. .

[0274] Spatial interaction branch: Introducing a graph attention mechanism. Each channel of the multi-channel secondary power supply is treated as a graph node. The electrical connection relationship between channels is defined as edge Considering the fully coupled characteristics of the secondary power supply system, a fully connected graph is constructed.

[0275] Calculate the attention coefficient between channel i and channel j The formula is as follows:

[0276] Here, || represents a splicing operation. W and W are learnable weight matrices.

[0277] Aggregate neighborhood information based on attention coefficients to generate spatial feature vectors containing global topological information. ,in, This step can capture consistency patterns in whether channel B responds as expected when channel A experiences fluctuations.

[0278] 2.2 Feature fusion and multi-task collaborative training; To obtain a high-dimensional representation that includes both waveform details and topological relationships, and to simultaneously train the model's reconstruction capability and uncertainty awareness, the following strategy is adopted: Strategy 1, Feature Fusion: The temporal feature vector output by the temporal encoder is fused together... Spatial feature vectors output by the spatial encoder By concatenating the data, we obtain the final latent feature vector that incorporates spatiotemporal information. :

[0279] The feature vector is It serves as the core basis for all subsequent evaluation tasks.

[0280] Strategy 2, Mask Co-reconstruction: To force the model to truly understand the electrical coupling relationships between channels, rather than merely memorizing the waveform of a single channel, the embodiments in this specification employ a mask co-training mechanism: Random mask: In each training iteration, a channel k in the input multi-channel data is randomly selected, and all its sampled points are set to zero (mask) to simulate the extreme case of complete loss of data in that channel.

[0281] Collaborative inference requires the decoder to utilize only the feature information of the remaining healthy channels, as well as the inter-channel correlation weights learned by the encoder at the graph attention layer. The goal is to reverse-engineer and complete the original waveform of the masked channel k. If the model can accurately reconstruct the masked waveform, it proves that the model has grasped the implicit functional relationship between channels A, B, and C and channel D.

[0282] Strategy 3, Uncertainty Estimation Head: To endow the model with adaptive perception of operating conditions, this embodiment integrates a lightweight uncertainty estimation head on the parallel branch of the reconstruction decoder. This uncertainty estimation head aims to capture the inherent uncertainty in the data, i.e., quantify the model's confidence level in the current reconstruction result. This uncertainty estimation head uses the high-dimensional feature vector output by the hybrid spatiotemporal encoder. For input. Feature vector After compression and mapping by the fully connected layer, the final output is an unbounded scalar value s. When the input data contains complex noise or severe transient interference, the network will express "low confidence" by outputting a large s value. This output will be directly used as a parameter in the subsequent loss function calculation, providing a mathematical basis for the model's anti-interference training.

[0283] Strategy 4, Multi-objective Loss Function Optimization: In multi-channel secondary power supply systems, the signal and noise distributions under different operating conditions exhibit significant heteroscedasticity. To enable the model to adaptively handle this difference, the embodiments in this specification introduce Bayesian deep learning theory, modeling the model's predictions as a probabilistic process following a Gaussian distribution. Specifically, the model output includes two headers: one for outputting the reconstructed waveform. (The mean of the corresponding Gaussian distribution), another uncertainty parameter s used for output prediction, takes... The network weights are updated by minimizing the following multi-objective loss function:

[0284] Among them, the first item This is a consistency reconstruction constraint. It forces the model's decoder to rely on the inter-channel coupling features (i.e., W and ...) extracted by the spatiotemporal encoder. The captured topological relationships are used to reconstruct the masked waveforms, thereby ensuring that the model truly grasps the physical laws governing the collaborative operation of multi-channel power supplies. (Second item) This is an uncertainty regularization term. For "atypical operating conditions" commonly encountered in industrial scenarios, such as load jumps or electromagnetic noise, the model struggles to accurately reconstruct the data using normal coupling rules. This term allows the model to automatically reduce the weight of the reconstruction error in the total loss by increasing the prediction variance *s*, providing a reliable mathematical basis for "adaptive dynamic weight generation" in the subsequent online evaluation stage.

[0285] 2.3 Construction of the Health Fingerprint Database (HealthPool): A healthy baseline is constructed using the trained encoder: 10% of data segments from the training set, representing the period when the system is in a steady state and exhibiting the lowest model reconstruction error, are selected as "healthy samples." These samples are then input into the encoder to extract a set of feature vectors. .

[0286] Calculate the mean vector of this set. The sum and covariance matrix ∑. Storage ( The dataset, ∑, serves as a health fingerprint database, representing the standard statistical distribution of the system's health status. Simultaneously, the prediction variances of these health samples in the network uncertainty estimation header output are obtained and their averages are calculated, denoted as the baseline uncertainty. .

[0287] Step 3: Real-time online data evaluation.

[0288] During real-time operation, the system executes the following process to calculate the Comprehensive Health Index (CHI): 3.1. Consistent Health Indicators (CHI) C ) The real-time window data is encoded to obtain the feature vector z. real Calculate z real Mahalanobis distance from the health fingerprint database (μ,∑) : ; Convert the distance to a consistency metric of 0-100: ; Where D thresh The threshold is set based on the 3σ statistical law.

[0289] 3.2 Reconstruction Error Health Index (CHI) R ) The original waveform is processed by hybrid spatiotemporal coding and then decoded to obtain the reconstructed waveform. Calculate the root mean square error between the original waveform and the reconstructed waveform. .

[0290] Convert the error to a reconstruction metric of 0-100: .

[0291] 3.3 Adaptive Dynamic Fusion Strategy Based on Bayesian Uncertainty In real-world industrial scenarios, power systems often face complex operating conditions such as transient load changes and electromagnetic interference. Under these "atypical conditions," the consistency index based on statistical distance (CHI) is crucial. C This model is highly susceptible to false positives due to temporary shifts in data distribution. To address this, this embodiment proposes a dynamic fusion mechanism based on uncertainty awareness, utilizing the model's "confidence level" in the current data. "To adjust the evaluation strategy in real time."

[0292] Real-time uncertainty measurement: During online evaluation, real-time data is input into a trained deep neural network. In addition to outputting fused feature vectors and reconstructed waveforms, the uncertainty estimation head integrated into the network outputs a scalar value—the uncertainty parameter s—in real time, which is then exponentially reduced. Obtain the predicted variance value The prediction variance value It quantifies the difference between the current input data and the model training data distribution; when When the value is relatively small: This indicates that the current data matches the "healthy pattern" learned by the model, and the model is highly confident in its predictions. A larger value indicates that the model has encountered interference, transients, or extreme noise that it has not seen before, and the model's prediction results are at high risk.

[0293] Dynamic weight generation mechanism: Utilizing the non-linear characteristics of the Sigmoid function, the predicted variance value is... The dynamic weight coefficient λ is mapped to a value between 0 and 1.

[0294] Integrated decision-making logic: Based on the calculated dynamic weighting coefficient λ, the final comprehensive health index (CHI) is constructed. .

[0295] The integrated circuit power supply switching self-healing control method provided in the embodiments of this specification includes the following steps: Step 1: Smooth and denoise the first health index data of the collected sample power sources using the Kalman filter algorithm to obtain the historical health index dataset.

[0296] The integrated circuit power supply power switching self-healing control method provided in this specification can be applied to electronic devices. The electronic devices are equipped with an integrated circuit power supply power switching self-healing control computer program. When the computer program is executed by the processor, it implements the integrated circuit power supply power switching self-healing control method of the embodiments in this specification.

[0297] In this embodiment, the first health index data of the sample power source is used as the basic data for the supervised learning dataset. In actual implementation, health index data for the entire lifecycle of the sample power source can be collected, or only the health index data for the latter half of the lifecycle can be collected. This embodiment does not impose specific limitations on this. The collected first health index data of the sample power source can be regarded as the original health index (HI) data sequence. To eliminate the influence of sensor noise and environmental interference on the original health index (HI) data sequence and improve the input quality and prediction accuracy of the subsequent prediction model, this scheme first uses a Kalman filter to smooth and denoise the real-time collected original HI data.

[0298] In the specific implementation process, the true health index of the power supply circuit can be regarded as a system state that changes over time, while the data collected by the sensor is a noisy observation. Using a one-dimensional Kalman filter, at each time step, the current state is first "predicted" based on the optimal estimate of the previous time step, and then the current "observation" is used to "correct" this prediction, thereby obtaining the optimal estimate for the current time step.

[0299] In an optional embodiment, the method of smoothing and denoising the first health index data of the collected sample power supply using the Kalman filter algorithm to obtain the historical health index dataset may include the following sub-steps: Sub-step 1: For each time step in the first health index data of the collected sample power supply, determine the measured health index value and the predicted health index value at the current time. Sub-step 2: Based on the current measured health index value and the current predicted health index value, and using the Kalman gain, determine the optimal estimated value of the health index at the current time; The optimal estimate of the health index at the current moment includes: the mean of the optimal estimate of the health index at the current moment, and the uncertainty of the optimal estimate of the health index at the current moment; the optimal estimate of the health index at the current moment is used as the predicted value of the health index at the next moment.

[0300] Sub-step 3: Use the optimal estimate of the health index corresponding to each time step as the historical health index dataset.

[0301] One specific way to generate a historical health index dataset is as follows: First, the parameters are defined as shown in the table below:

[0302] Secondly, the goal is to find the optimal estimate of the health index at the current moment. First, we need to obtain the current health index measurement value. and the predicted value of the health index at the current moment .

[0303] In one-dimensional Kalman filtering, the formulas for calculating the mean and variance of the optimal estimate of the health index are as follows:

[0304]

[0305] The health index after Kalman filtering at time t is ,in , which is the Kalman gain.

[0306] Among the parameters used in the above calculations, the predicted health index value at the current moment is taken as the optimal estimate of the health index at the previous moment, and the formula is as follows: The variance of the current health index prediction is the sum of the variance of the best estimate from the previous time step and a fixed prediction variance, as shown in the formula: .

[0307] By iterating repeatedly using the above formula, the mean of the optimal estimate of the health index at each moment can be obtained. The curve that changes over time is the health index after Kalman filtering.

[0308] Step 2: Use a sliding window approach to extract the historical health index dataset into multiple first data sequences of preset window lengths, and determine the actual remaining lifespan label corresponding to each first data sequence.

[0309] The preset window length can be flexibly adjusted by those skilled in the art. This specification does not impose specific limitations on this aspect in the embodiments. For example, the preset window length can be set to include 5 time steps, 10 time steps, etc.

[0310] This specification employs a particle swarm optimization-based LSTM-Attention model for power supply lifetime prediction in its embodiments. After obtaining the historical health index dataset after Kalman filtering in step one, it is necessary to determine the initial parameters of the particle swarm optimization-based LSTM-Attention model and generate a supervised learning dataset based on the historical health index dataset. When generating the supervised learning dataset, a sliding window approach is used to truncate the historical health index dataset into multiple first data sequences of preset window lengths, and the true remaining lifetime label corresponding to each first data sequence is determined.

[0311] To overcome the problems of error accumulation and low computational efficiency in traditional rolling forecasting methods, this specification proposes a more advanced end-to-end remaining useful life (HI) prediction framework. The core idea of ​​this framework is to redefine the model's prediction objective: instead of predicting the health index at the next time step, it directly predicts the time from the current moment until the HI value first reaches a preset failure threshold. The required time span. Therefore, the first step is to generate labels for the data sequence, which includes the following sub-steps: Sub-step 1: On the complete historical HI dataset, extract an input sequence of length L using a sliding window method. , where t is the starting point of the window.

[0312]

[0313] Sub-step 2: For each extracted input sequence, i.e., the first data sequence Determine the true remaining useful life label corresponding to the first number sequence. .

[0314] In one optional embodiment, the method for determining the true remaining useful life tag corresponding to each first data sequence may specifically include the following sub-steps: A1: Determine the end time point of the window corresponding to the first data sequence; Determine the time step corresponding to the end time point of the window. Each point in time corresponds to a time step.

[0315] A2: Starting from the end time point, scan the supervised learning dataset in the future direction to obtain the first time point that meets the preset conditions; Preset conditions can be set to , This is a preset failure threshold.

[0316] from Starting from this moment, scan the historical HI dataset towards the future to find the first one that satisfies the condition. Time step That is, the first time point.

[0317] A3: Calculate the time step difference between the first time point and the end time point; The time step difference can be calculated using the following formula.

[0318] A4: Generate the actual remaining useful life label corresponding to the first data sequence based on the time step difference and the time difference between time steps.

[0319] The first data sequence can be calculated using the following formula. RUL time stamp: Where △t represents the time for one step.

[0320] Step 3: For each first data sequence, store the first data sequence and its corresponding real remaining lifespan label to generate a supervised learning dataset.

[0321] All generated The sample pairs are combined to form the final supervised learning dataset used for model training.

[0322] Step 4: Determine the initial parameters of the LSTM-Attention prediction model based on the particle swarm optimization algorithm and the supervised learning dataset.

[0323] After obtaining a smoothed health index sequence through Kalman filtering, in order to accurately predict the remaining effective lifetime of the main power supply and determine the optimal switching time, this specification proposes a Long Short-Term Memory (LSTM) prediction model based on Particle Swarm Optimization (PSO) in the embodiments.

[0324] This solution addresses key parameters of traditional LSTM models, such as the time window length L and the number of hidden layer neurons. Learning rate For problems that are difficult to determine, the particle swarm optimization algorithm, which has global optimization capabilities and fast convergence characteristics, is used to adaptively optimize the model hyperparameters to determine the initial parameter values ​​of the model, in order to train a model with better performance.

[0325] In one alternative embodiment, determining the initial parameters of the LSTM-Attention prediction model based on the particle swarm optimization algorithm and the supervised learning dataset may include the following sub-steps: B1: Set up a particle swarm and randomly initialize the position and velocity of each particle in the swarm within a preset range; Each particle corresponds to a combination of parameters.

[0326] In practical implementation, the particle swarm size can be set to... (For example =20), maximum number of iterations is Each particle Represents a set of potential parameter combinations, whose position vector is denoted as The velocity vector is represented as The position and velocity of each particle are randomly initialized within a preset range.

[0327] B2: For each particle, the particle is used as a parameter of the LSTM-Attention prediction model. The LSTM-Attention prediction model is iteratively trained using a supervised learning dataset. The corresponding prediction error is then calculated. Specifically, the root mean square error (RMSE) on the supervised learning dataset can be used as the fitness function. For each parameter combination represented by a particle, a corresponding LSTM-Attention model is constructed, trained, and its prediction error is calculated.

[0328] in To determine the number of samples in the validation set, For the actual remaining lifespan, These are the model's predicted values. The smaller the fitness value, the better the parameter combination.

[0329] B3: During the iterative training of the LSTM-Attention prediction model using a supervised learning dataset, the historical best position and the global historical best position of the particle are recorded in each iteration. B4: Update the particle's velocity and position based on the particle's corresponding historical best position and global historical best position; In each iteration, the best position in the history of the individual particle is recorded. and the global historical best position In actual implementation, the velocity and position of each particle can be updated according to the following formula:

[0330]

[0331] in, For inertial weights, and The learning factor (usually 2.0) , It is a random number between [0,1].

[0332] B5: After completing the iterative training of each particle in the particle swarm, determine the parameter combination of the particle corresponding to the global optimal position as the initial parameters of the LSTM-Attention prediction model.

[0333] When the maximum number of iterations is reached If the fitness value is less than a preset threshold, the iteration of particles in the particle swarm is considered complete. The global optimal position is output. Corresponding parameter combinations As the initial parameters for the final LSTM-Attention model.

[0334] Step 5: Optimize and iteratively train the initial parameters of the LSTM-Attention prediction model based on the supervised learning dataset to obtain the target LSTM-Attention prediction model.

[0335] To accurately capture the complex nonlinear dynamics and long-term dependencies in health index data sequences, this specification's embodiments construct a deep prediction model, namely the LSTM-Attention prediction model, which integrates a Long Short-Term Memory network and an attention mechanism. This LSTM-Attention prediction model includes an input layer, an LSTM encoding layer, an attention layer, and a fully connected output layer.

[0336] The specific process of optimizing and iteratively training the initial parameters of the LSTM-Attention prediction model based on the supervised learning dataset can be referenced from existing methods of training models based on training samples, and no specific restrictions are imposed here.

[0337] Step 6: Collect the second health index data of the power supply to be monitored within a preset window length.

[0338] The data collected in this step is the latest data at a preset window length, and the purpose is to predict the latest health status of the power supply under test.

[0339] The preset window length is the same as the preset window length set when the historical health index dataset is truncated into a first data sequence of multiple preset window lengths using a sliding window method.

[0340] Step 7: Based on the target LSTM-Attention prediction model, predict the second health index data to obtain the remaining lifespan of the power supply to be monitored.

[0341] In one optional embodiment, the method of predicting the remaining lifespan of the power supply to be monitored based on the target LSTM-Attention prediction model may include the following sub-steps: Sub-step 1: Generate a second data sequence based on the second health index data and input it into the input layer.

[0342] In practical implementation, the health index data can also be filtered and smoothed using the Kalman filter algorithm.

[0343] The HI data sequence is smoothed using Kalman filtering and then segmented by a sliding window (window length L). For example, at time t, the input is... The output is a three-dimensional tensor with dimensions [batch size, time step, feature dimension], which will be used as the input to the LSTM encoding layer.

[0344] Sub-step 2: The input layer sequentially inputs the second data sequence into the LSTM encoding layer according to the time steps.

[0345] Sub-step 3: The LSTM encoding layer generates a sequence of hidden state vectors containing all time steps for the input second data sequence.

[0346] The hidden state vector sequence includes: batch size, time step, and number of LSTM hidden units.

[0347] Input sequence That is, the second data sequence is fed into the LSTM layer step by step. At each time step... The LSTM unit will adjust according to the current input. and the hidden state of the previous moment With cell state Calculate and update the hidden state at the current time step. and cell state This process encodes and transmits historical information. Among other things, The range is from 1 to L.

[0348] A sequence of hidden state vectors containing all time steps, denoted as... The hidden state vector sequence has dimensions of [batch size, time step L, number of LSTM hidden units]. The hidden state vector sequence H completely preserves the contextual information of the original input sequence, i.e., the second data sequence, at different time points, and is the basis for information filtering in subsequent Attention layers.

[0349] Sub-step 4: The attention layer dynamically weights the hidden state vector sequence to generate a context vector.

[0350] The context vector contains key time step information.

[0351] The attention layer in the embodiment provided in this specification simulates the human attention mechanism, dynamically weighting the hidden state vector sequence H output by the LSTM encoding layer. This enables the target LSTM-Attention prediction model to automatically identify and focus on key historical time points that contribute most to the prediction of future HI values, such as early signs of failure or turning points in degradation trends, while ignoring secondary or irrelevant information.

[0352] In one optional embodiment, the attention layer dynamically weights the hidden state vector sequence to generate the context vector, which may specifically include the following sub-steps: Sub-step 41: Calculate the attention score; Each hidden state is computed using a small feedforward neural network. Importance score This network is learnable; it automatically learns how to score different hidden states during training. The calculation formula is as follows:

[0353] in, , and These are the learnable weight matrix, bias terms, and context vectors in the network. This is the activation function, used to introduce nonlinearity.

[0354] Sub-step 42: Attention weight normalization; To transform the scores into a probability distribution, the Softmax function is used to calculate the scores at all time steps. Normalization is performed to obtain the final attention weights. :

[0355] Weight The value is between 0 and 1, and the sum of all weights is 1. The larger the value, the higher the hidden state at time step i. The more important it is for the final prediction.

[0356] Sub-step 43: Generate context vector; The obtained attention weights With the corresponding LSTM hidden state Perform a weighted summation to generate a single, fixed-length context vector. .

[0357]

[0358] Context vector It is a summary of the entire input sequence, which aggregates information from all time steps, but highlights the information at the moments that the model considers most critical.

[0359] Sub-step 5: The fully connected output layer generates the remaining lifetime of the power supply to be monitored based on the key time step information in the context vector.

[0360] The fully connected output layer serves as the model's final predictor, taking the highly condensed context vector generated by the Attention layer. Mapped to the final prediction result. Outputs a single numerical value, the predicted remaining lifespan.

[0361] In an optional embodiment, to eliminate false triggering of switching due to predicted fluctuations, this embodiment of the specification includes a confirmation logic based on a time window. Specifically, a threshold for the startup time of backup power supply intervention is set. The system calculates the RUL in real time, and only if the RUL predicted in n consecutive times satisfies If the power supply to be monitored is about to fail, a switching command will be immediately sent to the soft-switching control circuit.

[0362] The power supply under monitoring is the main power supply, while the power supply switched via a soft-switching transistor control circuit is the backup power supply. In the event of an impending failure of the main power supply, timely switching to the backup power supply enables self-healing control of the integrated circuit power supply.

[0363] The integrated circuit power supply power switching self-healing control method disclosed in the embodiments of this specification uses a Kalman filter algorithm to smooth and denoise the first health index data of the collected sample power supply to obtain a historical health index dataset; it uses a sliding window method to truncate the historical health index dataset into multiple first data sequences of preset window lengths, and determines the true remaining service life label corresponding to each first data sequence; it stores the first data sequences and their corresponding true remaining service life labels to generate a supervised learning dataset; it determines the initial parameters of the LSTM-Attention prediction model based on the particle swarm optimization algorithm and the supervised learning dataset; it optimizes and iteratively trains the initial parameters of the LSTM-Attention prediction model according to the supervised learning dataset to obtain the target LSTM-Attention prediction model; it collects the second health index data of the power supply to be monitored for a preset window length; and it predicts the remaining service life of the power supply to be monitored based on the target LSTM-Attention prediction model. The solution provided in the embodiments of this specification achieves several advantages. First, it employs an end-to-end prediction framework, enabling the model to directly learn the complex nonlinear mapping from the circuit health state sequence to its remaining lifetime, fundamentally avoiding the error accumulation problem inherent in traditional rolling prediction methods. Second, it combines an Attention mechanism that adaptively focuses on key fault features with Kalman filtering preprocessing that effectively removes noise, ensuring the accuracy and stability of the RUL prediction results. Third, it transforms high-precision prediction capabilities into precise and dynamic switching control. Since the model directly outputs RUL values ​​in time units, it provides a direct and quantitative basis for switching decisions, solving the problems of coarse time granularity and uncertain switching timing in traditional prediction methods. This allows the system to achieve power handover before a serious main power supply failure occurs, effectively protecting the safety of the downstream load circuits. In summary, the solution disclosed in the embodiments of this specification transforms the reliability assurance of the power system from a traditional passive post-fault response mode to an active pre-fault warning and avoidance mode, improving the reliability and safety of integrated circuits.

[0364] Corresponding to the above method embodiments, this specification also provides an embodiment of an integrated circuit power supply redundancy self-healing control device. Figure 6 This specification illustrates a schematic diagram of an integrated circuit power supply redundancy self-healing control device according to one embodiment. Figure 6 As shown, the device includes: The acquisition module 602 is configured to acquire the operating data of the secondary power supply at the current moment and the historical operating data at multiple historical moments collected by at least two sensors in the aging bench monitoring system. Extraction module 604 is configured to extract fusion features from the historical running data and the running data; The processing module 606 is configured to process the highly discrete data in the running data and the prediction result of the running data at the previous time step through an adaptive Kalman filter driven by a hypernetic network to obtain the prediction result of the running data at the current time step. The generation module 608 is configured to perform multi-task state evaluation based on the running data, the fusion features and the running data prediction results at the current moment, and generate evaluation results corresponding to the output voltage, ripple voltage and health status of the secondary power supply. The control module 610 is configured to update the parameters of the PID controller based on the evaluation results and the reinforcement learning algorithm, and to calculate and output a control signal through the updated PID controller to perform self-healing control on the output power of the secondary power supply.

[0365] Optionally, the generation module 608 is further configured to: Based on the fusion features, a weighted average is performed on the low-dispersion data in the running data to generate the corresponding processing result; The processing results and the current time-based operational data prediction results are then subjected to gating fusion to generate a corresponding fusion result; Based on the operational data, the fusion features, and the fusion results, a multi-task state evaluation based on an attention mechanism is performed to generate evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply.

[0366] Optionally, the processing module 606 is further configured to: The fused features are input into the supernetwork for mapping processing to generate an adaptive observation noise covariance matrix and an adaptive process noise covariance matrix, wherein the fused features include joint state features and discreteness features; The adaptive process noise covariance matrix and the posterior error covariance matrix of the previous time step are input into the Kalman filter for prediction processing to obtain the prior error covariance matrix of the current time step. Determine the error between the running data and the predicted running data at the current time, and calculate the Kalman gain based on the adaptive observation noise covariance matrix and the prior error covariance matrix at the current time; Based on the Kalman gain and the error, the prediction results of the current running data are corrected.

[0367] Optionally, the generation module 608 is further configured to: The fused features are mapped to query vectors, and the running data is mapped to keywords and key values; The query vector, the keyword, and the key value are subjected to context enhancement processing based on the cross-attention mechanism to generate context enhancement features; The context enhancement features are input into the physical parameter evaluation layer in the two-level cascaded network structure for processing, generating evaluation results corresponding to the output voltage and ripple voltage of the secondary power supply, respectively. The context enhancement features and the evaluation results corresponding to the output voltage and ripple voltage of the secondary power supply are input into the health state inference layer in the two-level cascaded network structure for processing, thereby generating the evaluation results corresponding to the health state of the secondary power supply.

[0368] Furthermore, the generation module 608 is also configured to: Obtain the parameter values ​​of at least two uncertainty parameters output by the physical parameter evaluation layer and the health status inference layer, wherein the at least two uncertainty parameters are used to evaluate the uncertainty of the evaluation result; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the evaluation result of the output voltage output by the physical parameter evaluation layer, calculate the first loss value of the normal-inverse gamma distribution negative log-likelihood loss function; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the evaluation result of the ripple voltage output by the physical parameter evaluation layer, calculate the second loss value of the normal-inverse gamma distribution negative log-likelihood loss function; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the health status assessment result output by the health status inference layer, calculate the third loss value of the normal-inverse gamma distribution negative log-likelihood loss function; The first loss value, the second loss value, and the third loss value are weighted and calculated to obtain the corresponding total loss value; The network parameters of the two-level cascaded network structure are optimized based on the total loss value.

[0369] Optionally, the control module 610 is further configured to: A state space vector is constructed based on the evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply, as well as the rated voltage of the secondary power supply. The state space vector is input into the policy network in the reinforcement learning framework for processing to generate an action space vector, wherein the action space vector is composed of the PID parameter adjustment amount of the PID controller. The parameters of the PID controller are updated based on the action space vector.

[0370] Furthermore, the control module 610 is also configured to: The reward function value is calculated based on the state space vector and action space vector at the current moment; Determine the state space vector for the next moment, and store the current state space vector, action space vector, the next state space vector, and the function value into the experience replay buffer. Experience data is periodically collected in the experience replay buffer at a preset frequency, and the network parameters of the policy network and value network in the reinforcement learning framework are updated based on the experience data using the stochastic gradient descent algorithm.

[0371] The above is a schematic scheme of an integrated circuit power supply redundancy self-healing control device according to this embodiment. It should be noted that the technical solution of this integrated circuit power supply redundancy self-healing control device and the technical solution of the aforementioned integrated circuit power supply redundancy self-healing control method belong to the same concept. Details not described in detail in the technical solution of the integrated circuit power supply redundancy self-healing control device can be found in the description of the technical solution of the aforementioned integrated circuit power supply redundancy self-healing control method.

[0372] Figure 7 A structural block diagram of a computing device 700 according to one embodiment of this specification is shown. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0373] The computing device 700 also includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these 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. The access device 740 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.

[0374] In one embodiment of this specification, the above-described components of the computing device 700 and Figure 7 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 7The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0375] The computing device 700 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 700 can also be a mobile or stationary server.

[0376] The processor 720 is used to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-described integrated circuit power supply redundancy self-healing control method.

[0377] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the integrated circuit power supply redundancy self-healing control method described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the integrated circuit power supply redundancy self-healing control method described above.

[0378] An embodiment of this specification 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 power supply redundancy self-healing control method.

[0379] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the integrated circuit power supply redundancy self-healing control method described above. Details not described in detail in the technical solution of the storage medium can be found in the description of the technical solution of the integrated circuit power supply redundancy self-healing control method described above.

[0380] An embodiment of this specification 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 power supply redundancy self-healing control method.

[0381] The above is an illustrative scheme of a computer program according to this embodiment. It should be noted that the technical solution of this computer program and the technical solution of the above-described integrated circuit power supply redundancy self-healing control method belong to the same concept. For details not described in detail in the technical solution of the computer program, please refer to the description of the technical solution of the above-described integrated circuit power supply redundancy self-healing control method.

[0382] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0383] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0384] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0385] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0386] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A self-healing control method for analytical redundancy in integrated circuit power supplies, comprising: Acquire the operating data of the secondary power supply at the current moment and the historical operating data at multiple historical moments collected by at least two sensors in the aging bench monitoring system; Extract fusion features from the historical operation data and the operation data; The highly discrete data in the running data and the prediction results of the running data at the previous time step are processed by an adaptive Kalman filter driven by a hypernetic network to obtain the prediction results of the running data at the current time step. Based on the operational data, the fusion features, and the operational data prediction results at the current moment, a multi-task state evaluation based on the attention mechanism is performed to generate evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply. Based on the evaluation results and reinforcement learning algorithm, the parameters of the PID controller are updated. The updated PID controller calculates and outputs a control signal to perform self-healing control of the output power of the secondary power supply.

2. The integrated circuit power supply redundancy self-healing control method according to claim 1 further includes: Based on the fusion features, a weighted average is performed on the low-dispersion data in the running data to generate the corresponding processing result; The processing results and the current time-based operational data prediction results are then subjected to gating fusion to generate a corresponding fusion result; Based on the operational data, the fusion features, and the fusion results, a multi-task state evaluation based on an attention mechanism is performed to generate evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply.

3. The integrated circuit power supply redundancy self-healing control method according to claim 1 further includes: The fused features are input into the supernetwork for mapping processing to generate an adaptive observation noise covariance matrix and an adaptive process noise covariance matrix, wherein the fused features include joint state features and discreteness features; The adaptive process noise covariance matrix and the posterior error covariance matrix of the previous time step are input into the Kalman filter for prediction processing to obtain the prior error covariance matrix of the current time step. Determine the error between the running data and the predicted running data at the current time, and calculate the Kalman gain based on the adaptive observation noise covariance matrix and the prior error covariance matrix at the current time; Based on the Kalman gain and the error, the prediction results of the current running data are corrected.

4. The integrated circuit power supply analytical redundancy self-healing control method according to claim 1, wherein the step of performing multi-task state evaluation based on the operating data, the fusion characteristics, and the current operating data prediction results, and generating evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply, includes: The fused features are mapped to query vectors, and the running data is mapped to keywords and key values; The query vector, the keyword, and the key value are subjected to context enhancement processing based on the cross-attention mechanism to generate context enhancement features; The context enhancement features are input into the physical parameter evaluation layer in the two-level cascaded network structure for processing, generating evaluation results corresponding to the output voltage and ripple voltage of the secondary power supply, respectively. The context enhancement features and the evaluation results corresponding to the output voltage and ripple voltage of the secondary power supply are input into the health state inference layer in the two-level cascaded network structure for processing, thereby generating the evaluation results corresponding to the health state of the secondary power supply.

5. The integrated circuit power supply redundancy self-healing control method according to claim 4 further includes: Obtain the parameter values ​​of at least two uncertainty parameters output by the physical parameter evaluation layer and the health status inference layer, wherein the at least two uncertainty parameters are used to evaluate the uncertainty of the evaluation result; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the evaluation result of the output voltage output by the physical parameter evaluation layer, calculate the first loss value of the normal-inverse gamma distribution negative log-likelihood loss function; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the evaluation result of the ripple voltage output by the physical parameter evaluation layer, calculate the second loss value of the normal-inverse gamma distribution negative log-likelihood loss function; Based on the parameter values ​​of at least two uncertainty parameters corresponding to the health status assessment result output by the health status inference layer, calculate the third loss value of the normal-inverse gamma distribution negative log-likelihood loss function. The first loss value, the second loss value, and the third loss value are weighted and calculated to obtain the corresponding total loss value; The network parameters of the two-level cascaded network structure are optimized based on the total loss value.

6. The integrated circuit power supply analytical redundancy self-healing control method according to claim 1, wherein updating the parameters of the PID controller based on the evaluation results and reinforcement learning algorithm includes: A state space vector is constructed based on the evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply, as well as the rated voltage of the secondary power supply. The state space vector is input into the policy network in the reinforcement learning framework for processing to generate an action space vector, wherein the action space vector is composed of the PID parameter adjustment amount of the PID controller. The parameters of the PID controller are updated based on the action space vector.

7. The integrated circuit power supply redundancy self-healing control method according to claim 6 further includes: The reward function value is calculated based on the state space vector and action space vector at the current moment; Determine the state space vector for the next moment, and store the current state space vector, action space vector, the next state space vector, and the function value into the experience replay buffer. Experience data is periodically collected in the experience replay buffer at a preset frequency, and the network parameters of the policy network and value network in the reinforcement learning framework are updated based on the experience data using the stochastic gradient descent algorithm.

8. An integrated circuit power supply redundancy self-healing control device, comprising: The acquisition module is configured to acquire the operating data of the secondary power supply at the current moment and the historical operating data at multiple historical moments collected by at least two sensors in the aging bench monitoring system. The extraction module is configured to extract fusion features from the historical running data and the running data; The processing module is configured to process the highly discrete data in the running data and the prediction results of the running data at the previous time step through an adaptive Kalman filter driven by a hypernetic network to obtain the prediction results of the running data at the current time step. The generation module is configured to perform multi-task state evaluation based on the running data, the fusion features, and the running data prediction results at the current moment, and generate evaluation results corresponding to the output voltage, ripple voltage, and health status of the secondary power supply. The control module is configured to update the parameters of the PID controller based on the evaluation results and reinforcement learning algorithm, and calculate and output a control signal through the updated PID controller to perform self-healing control of the output power of the secondary power supply.

9. 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 integrated circuit power supply redundancy self-healing control method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the integrated circuit power supply redundancy self-healing control method according to any one of claims 1 to 7.

Images