Training method of small sample dc-dc converter parameter identification model for edge scene and application thereof

Abstract

The application discloses a kind of small sample DC-DC converter parameter identification model for edge scene Training method and its application, belong to the field of power electronic equipment parameter identification, the method uses the low-frequency sampling strategy of cross period that originally concentrates in 1 cycle N sampling points disperses to M cycle, without relying on high-frequency sampling hardware, adaptation edge device low computing power, low storage resource constraint, reduce data acquisition and processing cost, while meeting real-time demand;Through the gradient search rule of GBO, the input weight and bias initial value of ELM are dynamically adjusted, and the local escape operator of GBO is combined to avoid the initialization of ELM task learner into local optimum, realize the high-precision and high-robustness identification of filter inductance and output capacitor C of DC-DC converter, and support the predictive maintenance and reliable operation of power electronic equipment.

Description

Technical Field

[0001] This invention belongs to the field of power electronic equipment parameter identification, and more specifically, relates to a training method for a small-sample DC-DC converter parameter identification model for edge scenarios and its application. Background Technology

[0002] As a core component of power electronic systems, DC-DC converters are widely used in consumer electronics, electric vehicles, and renewable energy power generation systems. Their operational reliability directly determines the stability and safety of the overall system. With increasing operating time, the internal components of a DC-DC converter will experience performance degradation due to aging, wear, and other factors. Specifically, this manifests as: a decrease in the capacitance (C) of electrolytic capacitors or an increase in the equivalent series resistance (ESR), semiconductor device parameter drift, and increased losses in magnetic components. Among these, the degradation of capacitors and inductors is the primary cause of converter failure—a capacitor is considered to have failed when its capacitance decreases by 20% or its ESR increases by 180-200%; an inductor is considered to have failed when its inductance decreases by 20% or its ESR increases by 100%. This type of degradation significantly increases the risk of converter failure and threatens the safe operation of the system.

[0003] To ensure the reliable operation of DC-DC converters, it is necessary to monitor the component parameter status in real time through parameter identification technology, thereby enabling degradation assessment and predictive maintenance to reduce downtime losses in critical application scenarios. Currently, DC-DC converter component-level degradation assessment and parameter identification methods are mainly divided into three categories:

[0004] Model-based methods: These methods construct mathematical models by analyzing converter fault mechanisms to directly estimate parameters such as capacitance (C) and inductance (L). However, they rely on precise physical models, have poor adaptability to parameter drift under complex operating conditions, and the model construction process is cumbersome, making it difficult to quickly adapt to DC-DC converters with different topologies. Signal-based methods: These methods assess the degree of degradation by comparing the difference between the converter's current operating signal and its normal state signal. This method does not require complex modeling, but it requires high signal acquisition accuracy and is difficult to quantify the degree of parameter degradation. It can only achieve qualitative fault judgment and cannot meet the needs of refined maintenance. Data-driven methods: These methods construct parameter identification models based on statistical analysis or machine learning techniques. They do not require in-depth understanding of the converter structure and degradation mechanism, have strong versatility and efficiency, and have become a research hotspot in the field of parameter identification. Common methods include neural networks, stochastic processes, and deep learning.

[0005] Despite the significant advantages of data-driven methods, three core challenges remain in practical applications: First, weak generalization ability in small-sample scenarios. Collecting fault samples from DC-DC converters is difficult and costly, and practical engineering often faces the problem of limited training samples. Traditional data-driven methods are prone to overfitting due to insufficient samples, leading to a significant drop in parameter recognition accuracy. Second, poor adaptability for edge deployment. Edge devices (such as power electronic controllers and embedded DSPs) typically have limited computing resources and low sampling frequencies. Mature data-driven algorithms (such as complex deep learning models) are difficult to deploy efficiently in edge scenarios due to their high computing power requirements and reliance on high-frequency sampling signals. Third, insufficient model stability. Among existing data-driven methods, Extreme Learning Machines (ELM) have become a potential choice for edge deployment due to their simple structure and non-iterative training characteristics. However, the random initialization of the input weight layer in ELM can cause the model parameter optimization to fall into unstable solutions, further amplifying recognition errors in small-sample scenarios and failing to meet the reliability requirements of parameter recognition.

[0006] Therefore, there is an urgent need for a DC-DC converter parameter identification technology that balances small sample adaptability, lightweight edge deployment, and model stability, in order to overcome the application bottlenecks of existing methods and support the reliable operation and intelligent maintenance of power electronic equipment in edge scenarios. Summary of the Invention

[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a training method and application for a few-sample DC-DC converter parameter recognition model for edge scenarios. This solves the technical problems of existing DC-DC converter parameter recognition technologies, such as weak generalization ability in few-sample scenarios, unstable recognition results due to random initialization of Extreme Learning Machine (ELM), and poor adaptability to edge device deployment.

[0008] To achieve the above objectives, according to a first aspect of the present invention, a training method for a small-sample DC-DC converter parameter recognition model for edge scenarios is provided, comprising: S1, acquiring C sets of voltage signals output by the DC-DC converter under A degradation scenarios and B load conditions according to a preset sampling frequency, and extracting D time-domain features of the voltage signals; constructing D×B meta-tasks, each meta-task using a set of voltage signals as samples and the time-domain features of the set of voltage signals as sample pseudo-labels; each meta-task having A×C sets of samples; dividing the samples of each meta-task into a support set and a query set according to a preset ratio;

[0009] Wherein, the preset sampling frequency , For switching frequency, Let \(n\) be the number of sampling points in one switching period, \(M\) be the number of low-frequency sampling periods, and \(N < M\); in the degradation scenario, the filter inductor and / or output capacitor of the DC-DC converter have different degrees of degradation;

[0010] S2, randomly initialize the input weights and biases of the extreme learning machine (ELM);

[0011] S3, respectively process each meta-task as the target meta-task; the processing includes: using the pseudo-labels in the support set of the target meta-task as the supervision signal to pre-train the ELM; calculating the prediction loss of the hidden layer of the pre-trained ELM for the pseudo-labels in the query set of the target meta-task using the query set of the target meta-task, and iteratively optimizing the input weights and biases of the ELM using the gradient optimizer GBO according to the overall prediction loss of the query set of the target meta-task until the iteration termination condition of the current round is reached;

[0012] S4, use the current input weights and biases as the initial values of the input weights and biases of the ELM, and return to S3 for the next round of iterative optimization until the termination condition is reached to obtain the optimal input weights and biases;

[0013] S5, respectively collect the voltage signals output by the DC-DC converter under multiple target scenarios and multiple load conditions and the corresponding filter inductors and output capacitors; use them as the training data set, and train the target ELM with the goal of minimizing the difference between the predicted values and the true values of the filter inductors and output capacitors output by the target ELM according to the voltage signals in the training data set to obtain a trained DC-DC converter parameter identification model;

[0014] wherein, the target scenarios include degradation scenarios and non-degradation scenarios; the input weights and biases of the target ELM are the optimal input weights and biases

[0015] According to the second aspect of the present invention, an electronic device is provided, including: a computer-readable storage medium and a processor;

[0016] The computer-readable storage medium is used to store executable instructions;

[0017] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the method described in the first aspect.

[0018] According to the third aspect of the present invention, a computer-readable storage medium is provided, and the computer-readable storage medium stores computer instructions for causing a processor to execute the method described in the first aspect.

[0019] According to a fourth aspect of the invention, a computer program product is provided, comprising a computer program or instructions that, when executed by a processor, implement the method described in the first aspect.

[0020] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0021] To address the shortcomings of existing DC-DC converter parameter recognition technologies, such as weak generalization ability in small-sample scenarios, unstable recognition results due to random initialization of Extreme Learning Machine (ELM), and poor adaptability to edge device deployment, as well as the need for improved accuracy, real-time performance, and reliability of parameter recognition in industrial scenarios, this invention provides a training method and application of a small-sample DC-DC converter parameter recognition model for edge scenarios. The aim is to improve the accuracy and stability of parameter recognition under small-sample conditions, reduce data storage / processing costs in edge scenarios, and adapt to lightweight deployment requirements. Specifically, the method provided by this invention employs a technical chain of "low-frequency sampling - meta-task construction - GBO-optimized ELM - attention meta-learning," achieving high-precision recognition even in small-sample scenarios (400 parameter combinations). The low-frequency sampling strategy of "N sampling points / M periods" distributes the N sampling points, originally concentrated in a single period, across M periods, preserving key features of the voltage signal at low sampling frequencies. This eliminates the need for high-frequency sampling hardware, adapting to the resource constraints of low computing power and low storage on edge devices, reducing data acquisition and processing costs while meeting real-time requirements. The method dynamically adjusts the input weights and initial bias values ​​of the ELM using the gradient search rule of GBO, and combines this with the local escape operator of GBO to prevent the ELM task learner from getting trapped in local optima during initialization. Physical experiments and simulations demonstrate that the method provided by this invention achieves a mean square error (MSE) of only 1.32 × 10⁻⁶ for the identification of the filter inductor L. -4 The identification MSE of the output capacitor C is as low as 2.46 × 10⁻⁶. -5 With a mean percentage deviation (MPD) of ≤2%, it solves the problem of weak generalization ability of traditional data-driven methods with small samples. Simultaneously, by optimizing the initial parameters of the ELM through GBO, the standard deviation of the identification error is reduced by more than 99% compared to the traditional ELM in 100 random experiments, completely resolving the instability problem caused by random initialization of the ELM. It can operate stably in typical DC-DC topologies such as BUCK converters, and the identification results can directly support predictive maintenance of power electronic equipment (such as judging the remaining life of components based on the degradation trend of L and C parameters), providing a reliable parameter monitoring solution for edge scenarios such as vehicle power supplies and distributed energy nodes.

[0022] As a further preferred option, the method provided by this invention uses an attention mechanism to dynamically allocate the weights of meta-tasks: based on the loss value of each meta-task on the query set, the proportion is calculated after the loss value is processed by indexation (i.e., the weight of a single meta-task = the indexed loss of that task / the sum of the indexed losses of all meta-tasks), so that the low-loss meta-tasks with higher information value get higher weights, making the model highly robust to load fluctuations (0%~125% load change), with no significant decrease in recognition accuracy under different loads, and adapting to complex working conditions in industrial scenarios. Attached Figure Description

[0023] Figure 1 A flowchart illustrating the training method for a small-sample DC-DC converter parameter recognition model for edge scenarios provided in this embodiment of the invention;

[0024] Figure 2 This is a schematic diagram of the Buck circuit structure;

[0025] Figure 3 (a) and (b) in the figure are schematic diagrams of the output voltage ripple under high-frequency sampling (500 kHz) in two different degradation scenarios, respectively;

[0026] Figure 4 (a) and (b) in the figure are schematic diagrams of the output voltage ripple under low-frequency (40kHz) sampling in two different degradation scenarios, respectively;

[0027] Figures 5-8 The diagrams show the results of 100 independent parameter identifications using the PKAML-ELM parameter identification model proposed in this invention, traditional ELM, random forest (RF), and support vector regression (SVR), respectively.

[0028] Figure 9 A schematic diagram of the physical experimental platform for the DC-DCBUCK converter provided in an embodiment of the present invention.

[0029] Figure 10 A schematic diagram of the DSP and PC experimental platform for deploying the ELM model provided in this embodiment of the invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0031] This invention provides a training method for a few-sample DC-DC converter parameter identification model for edge scenarios, comprising:

[0032] S1. Collect C groups of voltage signals output by the DC-DC converter under A degradation scenarios and B load conditions at a preset sampling frequency, and extract D time-domain features of the voltage signals; construct D×B meta-tasks, each meta-task taking a group of voltage signals as samples and the time-domain features of this group of voltage signals as sample pseudo-labels; each meta-task has A×C groups of samples; divide the samples of each meta-task into a support set and a query set according to a preset ratio.

[0033] Among them, the preset sampling frequency , is the switching frequency, is the number of sampling points in one switching period, M is the number of low-frequency sampling periods, and N<M; the filter inductor and / or output capacitor of the DC-DC converter under the degradation scenario has different degrees of degradation.

[0034] Preferably, the multiple time-domain features include: maximum value, minimum value, mean value, standard deviation, root mean square amplitude, absolute mean value, skewness, kurtosis, root mean square, variance, kurtosis index, peak index, waveform index, pulse index, skewness index and margin index.

[0035] In step S1, based on the periodicity of the output voltage signal of the DC-DC converter, adopt a low-frequency sampling strategy to collect the output voltage signals of the DC-DC converter under various degradation scenarios, and extract multiple time-domain features (such as the above 16 time-domain features) of the collected voltage ripple signals. Then, respectively take the extracted time-domain features as pseudo-labels, and combine the voltage signals corresponding to the pseudo-labels to form a meta-task data set, and divide the support set and the query set from the meta-task data set according to a preset ratio. The preset ratio can be set according to the actual situation. For example, the ratio of the support set to the query set is 8:2. Each basic unit of the meta-task data set consists of "1 group of voltage signal samples + 16 time-domain feature pseudo-labels". Each time-domain feature corresponds to an independent meta-task, providing a task basis for subsequent meta-learning.

[0036] The low-frequency sampling strategy is the "N sampling points / M cycles" cross-cycle sampling strategy, that is, disperse the N sampling points originally concentrated in 1 cycle to M cycles, and collect 1 point every M / N cycles. Its sampling frequency is determined by the formula ( is the switching frequency of the DC-DC converter). This low-frequency sampling strategy retains the key features of the voltage signal at a low sampling frequency by dispersing the sampling points within a single cycle to multiple cycles, without relying on high-frequency sampling hardware.

[0037] When constructing the meta-task dataset, the ratio of the support set to the query set can be dynamically adjusted according to the feature recognition of the voltage ripple signal. If the signal features become blurred due to fluctuations in operating conditions, the proportion of the support set can be increased (e.g., from 8:2 to 9:1) to reduce the interference of low-quality samples on model training.

[0038] S2, randomly initialize the input weights and biases of the Extreme Learning Machine (ELM);

[0039] S3, each meta-task is processed as a target meta-task; the processing includes: pre-training the ELM using pseudo-labels in the support set of the target meta-task as supervision signals; calculating the prediction loss of the hidden layer of the pre-trained ELM on the pseudo-labels in the query set of the target meta-task using the query set of the target meta-task; iteratively optimizing the input weights and biases of the ELM using the gradient optimizer GBO based on the overall prediction loss of the query set of the target meta-task until the iteration termination condition of the current round is reached; (that is, in step S3, the ELM is pre-trained using pseudo-labels in the support set of each meta-task as supervision signals; the prediction loss of the hidden layer of the pre-trained ELM on the pseudo-labels in the query set of each meta-task is calculated using the query set of each meta-task; the input weights and biases of the ELM are iteratively optimized using the gradient optimizer GBO based on the overall prediction loss of each query set of each meta-task until the iteration termination condition of the current round is reached).

[0040] S4: Use the current input weights and biases as the initial values ​​of the input weights and biases of ELM, return to S3 for the next round of iterative optimization, until the termination condition is met and the optimal input weights and biases are obtained.

[0041] The above steps S2~S4 include pre-training the ELM based on fault prior knowledge, and using the gradient optimizer GBO to iteratively optimize the input weights and biases of the ELM during the process. The purpose of pre-training is to enable the hidden layer of the ELM to learn the fault mechanism feature mapping, and to iteratively optimize the input weights and biases of the ELM through GBO based on the prediction loss of the pseudo-labels of the query set of the meta-task.

[0042] In the above process, the initial values ​​of the input weights and biases of the ELM are dynamically adjusted through the gradient search rule of the GBO. Simultaneously, the local escape operator of the GBO is used to prevent the ELM learner from getting trapped in local optima during initialization. The candidate solution set generated by the local escape operator of the GBO includes: the optimal parameters for the current iteration of ELM initialization, the parameters obtained from the gradient search rule, and randomly generated parameters. By using these multiple candidate solutions, the local optimum constraint is broken, preventing the ELM initialization from getting trapped in local optima. The parameters are the input weights and biases of the ELM.

[0043] In the gradient search rule of GBO, the direction and step size of the adjustment of ELM input weights and initial bias values ​​are dynamically controlled by introducing gradient adjustment coefficients to accelerate the convergence of parameters to the optimal solution; at the same time, the abnormal situation of the denominator being zero during the gradient calculation is avoided by setting small perturbation values.

[0044] Preferably, the minimum value β_min of the gradient adjustment coefficient is set to 0.2 and the maximum value β_max is set to 1.2. The small perturbation value ε, which is used to avoid the denominator of the gradient calculation being zero, is set in the interval [0,0.1].

[0045] The iteration termination condition for the current round in step S3 is: when the number of GBO iterations reaches the preset number, or when the mean square error (MSE) of the pseudo-label prediction of ELM on the query set is less than the preset value, the optimization stops, and the input weights and biases at this time are saved as the initial values ​​of the input weights and biases of ELM in the next round.

[0046] The termination condition in step S4 is: the number of iterations reaches the preset number of iterations.

[0047] The overall prediction loss of the query set in step S3 (also known as the query set loss) can be the sum of the prediction losses of each pseudo-label in the query set.

[0048] As a further preferred option, the overall prediction loss of the query set in step S3 is the weighted sum of the prediction losses of each pseudo-label in the query set. That is, before calculating the overall prediction loss of the query set, an attention mechanism is introduced to dynamically allocate the weights of each meta-task based on the loss of each meta-task (i.e., assign weights to the prediction losses of each pseudo-label in the query set). The formula for calculating the weights of the prediction losses of each pseudo-label in the query set is:

[0049]

[0050] in, For the query set The weight of each pseudo-label This is the sharpness control coefficient for the weight distribution. For the size of the query set, For the query set The prediction loss for each pseudo-label, For the query set The prediction loss for each pseudo-label.

[0051] During the pre-training phase of fault prior knowledge, the loss function for the prediction loss of pseudo-labels can be any existing loss function to learn the feature mapping of the fault mechanism, such as the mean squared error loss function.

[0052] S5: Collect the voltage signals output by the DC-DC converter under multiple target scenarios and multiple load conditions, as well as the corresponding filter inductors and output capacitors; use these as training datasets; train the target ELM with the goal of minimizing the difference between the predicted and true values ​​of the filter inductors and output capacitors output by the voltage signals in the training dataset, and obtain a trained DC-DC converter parameter recognition model.

[0053] The target scenarios include degraded scenarios and non-degraded scenarios; the input weights and biases of the target ELM are the optimal input weights and biases.

[0054] In step S5, the pre-trained model is fine-tuned using the real parameter labels of the DC-DC converter parameters to obtain the PKAML-ELM parameter recognition model, i.e., the DC-DC converter parameter recognition model.

[0055] During the fine-tuning stage of the true parameter labels, any existing loss function can be used to optimize the parameter identification accuracy, such as the cross-entropy loss function.

[0056] Preferably, the candidate solution set generated by the local escape operator of GBO includes: the optimal parameters of the current iteration, the parameters obtained by the gradient search rule, and randomly generated parameters; wherein, the optimal parameters of the current iteration are selected from the parameters obtained by each iteration optimization in the current iteration based on the principle of minimizing the overall prediction loss of the query set; the parameters include the input weights and biases of ELM.

[0057] This invention provides a method for identifying parameters of a small-sample DC-DC converter in edge scenarios, comprising:

[0058] The DC-DC converter parameter identification model trained using the method described in any of the above embodiments is burned into the edge device. The voltage signal of the DC-DC converter to be identified is input to the edge device. The edge device calls the DC-DC converter parameter identification model to identify its parameters and obtains the filter inductor and output capacitor of the DC-DC converter.

[0059] Specifically, taking the edge device as DSP hardware as an example, the PKAML-ELM model trained using the above training method is reconstructed into the DSP hardware through matrix operations to adapt to the computing power constraints of the edge hardware. The DSP receives the voltage signal to be identified in real time through a preset data receiving interface, calls the PKAML-ELM model to calculate and output the identification results of the filter inductor L and output capacitor C of the DC-DC converter.

[0060] The matrix operation reconstruction includes: storing the input weight matrix ω, bias matrix b, and output weight matrix beta of the PKAML-ELM model into DSP hardware; the model inference process is implemented using the formula H=sofplus(X·ω+b) (where H is the ELM hidden layer output, sofplus is the activation function, and X is the input voltage signal); the output of the PKAML-ELM model is calculated using "Y=H·beta", where H is generated by processing the "product of the input signal and ω plus b" through the ELM activation function. The DSP hardware can be an embedded processor with edge real-time computing capabilities (such as the TMS320F28377D), capable of supporting matrix operation reconstruction and inference execution of the PKAML-ELM model, retaining only the core logic of model inference, and eliminating redundant training modules.

[0061] The period adaptation parameters of low-frequency sampling and the signal preprocessing logic of time-domain feature extraction can be dynamically adjusted according to the topology type of the DC-DC converter (such as BUCK, BOOST) to ensure feature integrity under different topology scenarios.

[0062] The DSP deployment module also has hardware resource adaptation and optimization functions: it can perform lightweight compression of the core parameter matrix (ω, b, β) of the PKAML-ELM model according to the actual memory and computing speed of the DSP, retaining key elements that contribute highly to the recognition accuracy, and reducing hardware resource consumption without reducing accuracy; at the same time, it solidifies the matrix operation process of model inference, reducing the data reading time during real-time calls.

[0063] The method provided in the embodiments of the present invention will be further illustrated below with a specific example.

[0064] The implementation focuses on a typical BUCK topology DC-DC converter (its circuit structure is as follows). Figure 2 As shown, the input voltage 、 Output voltage 、 Switching frequency 、 load resistor The core focus is on parameter identification of the converter's filter inductor (nominal value 47μH) and output capacitor (nominal value 22μF). The specific steps are as follows:

[0065] (1) Low-frequency sampling and temporal feature extraction

[0066] Based on the periodicity (switching period) of the output voltage signal of the BUCK converter The output voltage ripple signal is acquired using a "4-point / 5-cycle" low-frequency sampling strategy, that is, the sampling frequency is calculated according to the formula. Calculate the sampling frequency. (Sampling interval) By distributing the original four sampling points concentrated in one cycle to five cycles (one point is sampled every 1.25 cycles, and the signal is spliced ​​across cycles), the amplitude fluctuation and periodic characteristics of the voltage ripple are preserved at a low sampling frequency, without relying on... High-frequency sampling hardware.

[0067] Signal acquisition hardware: An industrial-grade voltage sampling module (sampling accuracy ±0.1%) is connected to the output of the BUCK converter to synchronously acquire the output voltage ripple signal. The acquisition time is 1 second, generating 40,000 sampling points for each operating condition (corresponding to...). );

[0068] Time-domain feature extraction: For the acquired voltage ripple signal, 16 time-domain features are extracted using Python's scipy library. These features include: maximum value (reflecting signal peak fluctuations), minimum value (reflecting signal valley characteristics), mean (reflecting signal DC bias), standard deviation (reflecting signal dispersion), square root amplitude (reflecting signal energy distribution), absolute mean (reflecting signal average amplitude), skewness (reflecting signal symmetry), kurtosis (reflecting signal peaking), root mean square (reflecting signal RMS value), variance (reflecting signal fluctuation intensity), kurtosis index (normalized ratio of kurtosis to variance), peak index (ratio of peak value to root mean square), waveform index (ratio of root mean square to absolute mean), impulse index (ratio of peak value to absolute mean), skewness index (ratio of skewness to cube root variance), and margin index (ratio of peak value to square root amplitude). Each feature is normalized using "(feature value - mean) / standard deviation" to avoid dimensional interference.

[0069] (2) Construction of meta-task dataset

[0070] The 16 time-domain features extracted from S1 are used as pseudo-labels and combined with the corresponding voltage signal samples to form the meta-task dataset. The specific implementation is as follows:

[0071] Dataset source: Degradation scenarios of L and C in a simulated BUCK converter. The degradation range of L is 0%~19% (20 levels set in 1% steps, actual values ​​are 47μH~37.97μH), and the degradation range of C is 0%~19% (20 levels set in 1% steps, actual values ​​are 22μF~17.82μF). 400 sets of core parameter combinations are generated by combining "L level × C level". For each parameter combination, 4 sets of voltage signal samples are collected under each load, covering 5 load conditions (load resistance R corresponds to 0%~125% of rated load, i.e. 2.4Ω~3.75Ω), finally generating 400×5×4=8000 original samples.

[0072] Meta-task partitioning: Each original sample is bound to its 16 temporal feature pseudo-labels, constructing a total of 16×5=80 meta-tasks; each meta-task has 400×4=1600 samples. The 1600 samples of each meta-task are divided into a support set and a query set of the meta-task according to a preset ratio of 8:2. The support set is used for model pre-training on the meta-task, and the query set is used to verify the learning effect of the meta-task and back-optimize the meta-learning parameters.

[0073] Data format processing: Voltage signal samples are converted into 8-dimensional vectors (each 8 consecutive sampling points is an input segment), and time-domain feature pseudo-labels are converted into 16-dimensional vectors to ensure that the meta-task dataset is compatible with the input dimension of the subsequent ELM model.

[0074] (3) Pre-training on prior knowledge of faults

[0075] To avoid local optima, the initial values ​​of the input weights and biases of the ELM are adjusted using a gradient optimizer (GBO), including:

[0076] ELM basic parameter settings: ELM adopts a single hidden layer feedforward structure, with 8 input layer nodes (matching voltage signal segment dimension) and a hidden layer node count of... (Determined through cross-validation, balancing accuracy and computational power), the output layer has 2 nodes (corresponding to the recognition results of L and C), and the sofplus activation function is selected. , Input for the sample;

[0077] GBO Optimization Logic: Gradient Search Rule (GSR): Setting Gradient Adjustment Coefficient minimum value Maximum value Small perturbation value (To avoid the denominator of gradient calculation being zero); dynamically adjust the ELM input weight matrix using GSR. The initial values ​​of the bias matrix b (dimension 8×213) and the bias matrix b (dimension 1×213) are adjusted by the step size determined by the difference between the optimal parameter and the random parameter in the current iteration, which accelerates the convergence of the parameters to the global optimum.

[0078] Local Escape Operator (LEO): Generates a set of three types of candidate solutions: ① the optimal parameters of the current iteration (obtained by minimizing the query set loss, i.e., by selecting the parameters obtained from each iteration in the current iteration based on the principle of minimizing the overall prediction loss of the query set), ② the parameters calculated by GSR, and ③ random parameters uniformly distributed in the interval (-1,1). LEO is triggered every 10 iterations to select the parameter with the minimum loss from the candidate solutions as the new initial value, thus breaking the local optimum constraint.

[0079] The optimization termination condition for the current iteration round is: when the GBO iteration count reaches 100, or the mean squared error (MSE) of the pseudo-label prediction on the query set by ELM is ≤10. -4 Stop optimization and save the current state. b is used as the initial parameter of ELM.

[0080] Attention mechanism weight allocation: For each meta-task in the query set, calculate its pseudo-label prediction loss. (Mean squared error), through " ( Control the sharpness of the weight distribution. For the size of the query set, For the query set The weight of each meta-task, The query set The pseudo-label prediction loss of each meta-task is used to calculate the meta-task weights; the top 20% of meta-tasks with the smallest loss (such as tasks with stable load and clear features) account for a total weight of 60%, ensuring that the model focuses on high-value task information.

[0081] Using the temporal feature pseudo-labels of the support set as the supervision signal, the mean square error loss function is employed. , These are genuine and fake labels. To predict pseudo-labels, the overall loss of the query set is calculated by combining the weights of the meta-task, and a GBO-based ELM is trained. After 50 iterations, the model learns the mapping relationship between voltage signals and fault features, and obtains the optimal values ​​of the input weights and biases of the ELM.

[0082] (4) ELM fine-tuning based on real parameter labels

[0083] The voltage signals of the DC-DC converter and the corresponding L and C are collected under both degraded and non-degraded scenarios.

[0084] The true values ​​of L and C are used as supervision signals (obtained by a high-precision LCR meter with a measurement accuracy of ±0.5%), and the cross-entropy loss function is employed. , For actual parameter labels, The ELM model was fine-tuned (for the prediction parameters) for 20 iterations with a learning rate of 0.001, resulting in the final PKAML-ELM parameter recognition model.

[0085] Model Validation: The model performance was validated on the query set, where the recognition MSE for L was 1.32 × 10⁻⁶. -4 The MSE for C is 2.46 × 10⁻⁶. -5 The mean percentage deviation (MPD) is ≤2%, which meets the accuracy requirements.

[0086] (5) DSP reconstruction of PKAML-ELM model

[0087] The TMS320F28377D DSP (32-bit floating-point arithmetic, 200MHz clock speed, 1MB on-chip RAM) was selected as the edge deployment hardware. The model was reconstructed through matrix operations, specifically as follows: Core parameter matrix storage: The PKAML-ELM model... The 8×2¹³ dimension, b (1×2¹³ dimension), and output weight matrix β (2¹³×2 dimension) are stored in the on-chip RAM in the DSP's memory-aligned format (32-bit floating-point number). The total space occupied by the matrix is ​​approximately (8×2¹³+1×2¹³+2¹³×2)×4Byte=11.4KB, which is adapted to the DSP's memory constraints.

[0088] Solidified reasoning logic: The model reasoning process is converted into basic matrix operations, that is:

[0089] 1) Input voltage signal segment X (8-dimensional vector) and Matrix multiplication yields an intermediate result of 8×2¹³ dimensions;

[0090] 2) Stack the bias matrix b, and calculate the hidden layer output H (1×2^13 dimension, H=sofplus(X·ω+b)) using the sofplus activation function.

[0091] 3) Multiply H and β matrices to obtain the recognition result Y of L and C (1×2 dimension, Y=H·β);

[0092] Deployment verification: Voltage signals are received through the DSP's ADC interface (12-bit precision, 1MHz sampling rate). The model inference process does not depend on external computing power, and the single-sample diagnosis time is stable at 1.1ms, meeting the real-time requirements of edge scenarios.

[0093] (6) Real-time parameter identification and result output

[0094] The DSP receives the voltage signal acquired by S1 in real time through a preset UART interface (baud rate 115200bps) (receiving one group of 8-dimensional signal segments every 100ms), and calls the reconstructed PKAML-ELM model to perform the following operations:

[0095] Signal preprocessing: Normalize the received voltage signal segments (using the mean and standard deviation parameters in step (1)) to ensure consistency with the distribution of the model training data;

[0096] Parameter calculation: Calculate the recognition values ​​of L and C according to the matrix operation logic in step (5). If the MPD of the recognition results is greater than 2% for three consecutive recognitions, trigger model fine-tuning (call the fine-tuning logic of S4 to quickly optimize the β matrix based on the current 5 new samples), and re-burn the fine-tuned model to the DSP.

[0097] Output results: The DSP outputs the current values ​​of L and C and the degradation level in real time (e.g., Corresponding to a 4.26% degradation, (corresponding to a 9.09% degradation), supporting subsequent predictive maintenance decisions.

[0098] The data acquisition and feature extraction consist of an industrial-grade voltage sampling module (model AD7606, 8-channel 16-bit precision) and an embedded MCU (STM32F103). The MCU controls the sampling module to acquire voltage signals according to the "4-point / 5-cycle" strategy, and extracts 16 time-domain features through a preset algorithm. The feature extraction time is ≤5ms / sample.

[0099] The meta-task dataset is constructed by a building module deployed on a PC (CPU i7-12700H, memory 16GB). It receives “voltage signal + time domain feature” data transmitted from the MCU via USB interface, automatically combines them to form the meta-task dataset, and divides it into support set and query set in an 8:2 ratio. The dataset is stored in CSV format, and each sample occupies approximately 128 bytes of space.

[0100] GBO-ELM optimization and fine-tuning are implemented through a module integrated into a Python environment on the PC (relying on TensorFlow and SciPy libraries). The former optimizes the parameters of GBO for ELM, and the latter performs fine-tuning. The total training time is about 15 minutes.

[0101] DSP hardware: Composed of TMS320F28377D DSP, the parameter matrix and inference logic of the PKAML-ELM model are burned into the DSP through the JTAG interface, and the deployment process takes ≤2 minutes;

[0102] Real-time identification is achieved by a module embedded in the DSP firmware, which receives sampling signals through the ADC interface, calls the model to complete parameter identification, and outputs the results on the PC through the Ethernet / UART interface. The module has an average power consumption of ≤1.5W, which is suitable for the low power consumption requirements of edge devices.

[0103] Through 100 randomized experiments (each experiment randomly divided the training / test set and randomly initialized the GBO parameters), and compared with traditional ELM, random forest (RF), and support vector regression (SVR), the performance of the PKAML-ELM model of this invention is shown in Table 1:

[0104] Table 1 Test Results

[0105]

[0106] As shown in Table 1, the method provided by this invention has significantly better recognition accuracy than traditional methods in small sample scenarios, and the model stability (MSE standard deviation ≤ 3.76 × 10⁻⁶) is also superior. -4 Both the model and its real-time performance (single sample diagnosis time ≤ 1.5ms) meet the requirements of edge scenarios; at the same time, under the conditions of load fluctuation (0%~125%) and sampling frequency variation (10kHz~500kHz), the model recognition accuracy does not decrease significantly (mean percentage error (MPD) fluctuation ≤ 0.5%), proving its strong robustness.

[0107] This invention provides an electronic device, including: a computer-readable storage medium and a processor;

[0108] The computer-readable storage medium is used to store executable instructions;

[0109] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the training method or parameter identification method as described in any of the above embodiments.

[0110] Furthermore, the instruction can also implement the following auxiliary logic during execution: First, it records the parameters of the construction process of the meta-task dataset, including the extraction accuracy of temporal features, the ratio of support set to query set and the corresponding number of samples, so as to facilitate the tracking of the impact of different dataset construction parameters on the training effect of PKAML-ELM model; Second, it records the key parameters during training in real time, including the initial parameter adjustment results of GBO-optimized ELM and the meta-task weight allocation data of attention mechanism. The optimal training process can be quickly reproduced based on these records, or fine-tuning can be quickly started based on historical training parameters when small sample data is incrementally supplemented, without having to re-execute the complete prior knowledge pre-training process.

[0111] This invention provides a computer-readable storage medium storing computer instructions for causing a processor to execute a training method or parameter identification method as described in any of the above embodiments.

[0112] This invention provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the training method or parameter identification method as described in any of the above embodiments.

[0113] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A training method for a few-sample DC-DC converter parameter identification model for edge scenarios, characterized in that, include: S1, Collect C sets of voltage signals output by the DC-DC converter under A degradation scenarios and B load conditions according to the preset sampling frequency, and extract D time-domain features of the voltage signals; Construct D×B meta-tasks, each meta-task using a set of voltage signals as samples and the time-domain features of that set of voltage signals as sample pseudo-labels; each meta-task has A×C sets of samples; The samples of each meta-task are divided into a support set and a query set according to a preset ratio; Among them, the preset sampling frequency , is the switching frequency, is the number of sampling points in a switching period, M is the number of low-frequency sampling periods, and N < M; the filter inductor and / or output capacitor of the DC-DC converter in the degradation scenario have different degrees of degradation; S2, randomly initialize the input weights and biases of the Extreme Learning Machine (ELM); S3, each meta-task is processed as a target meta-task; the processing includes: using pseudo-labels in the support set of the target meta-task as supervision signals to pre-train the ELM; using the query set of the target meta-task to calculate the prediction loss of the hidden layer of the pre-trained ELM for the pseudo-labels in the query set of the target meta-task; using the gradient optimizer GBO to iteratively optimize the input weights and biases of the ELM according to the overall prediction loss of the query set of the target meta-task, until the iteration termination condition of the current round is reached; S4, use the current input weights and biases as the initial values ​​of the input weights and biases of ELM, return to S3 for the next round of iterative optimization, until the termination condition is met and the optimal input weights and biases are obtained. S5: Collect the voltage signals output by the DC-DC converter under multiple target scenarios and multiple load conditions, as well as the corresponding filter inductors and output capacitors; use these as training datasets; train the target ELM with the goal of minimizing the difference between the predicted and true values ​​of the filter inductors and output capacitors output by the voltage signals in the training dataset, and obtain a trained DC-DC converter parameter recognition model. The target scenarios include degraded scenarios and non-degraded scenarios; the input weights and biases of the target ELM are the optimal input weights and biases.

2. The method as described in claim 1, characterized in that, The candidate solution set generated by the local escape operator of GBO includes: the optimal parameters of the current iteration, the parameters obtained by the gradient search rule, and randomly generated parameters; wherein, the optimal parameters of the current iteration are selected from the parameters obtained by each iteration optimization in the current iteration based on the principle of minimizing the overall prediction loss of the query set; the parameters include the input weights and biases of ELM.

3. The method as described in claim 2, characterized in that, The overall prediction loss of the query set is the weighted sum of the prediction losses of each pseudo-label in the query set; wherein the weight of the prediction loss of each pseudo-label in the query set is calculated using the following formula: in, For the query set The weight of each pseudo-label This is the sharpness control coefficient for the weight distribution. For the size of the query set, For the query set The prediction loss for each pseudo-label, For the query set The prediction loss for each pseudo-label.

4. The method as described in claim 1, characterized in that, The multiple time-domain features include: maximum value, minimum value, mean, standard deviation, square root amplitude, absolute mean, skewness, kurtosis, root mean square, variance, kurtosis index, peak value index, waveform index, impulse index, skewness index, and margin index.

5. A method for identifying parameters of a small-sample DC-DC converter in edge scenarios, characterized in that, include: The DC-DC converter parameter identification model trained using the method described in any one of claims 1-4 is burned into the edge device. The voltage signal of the DC-DC converter to be identified is input to the edge device. The edge device calls the DC-DC converter parameter identification model to identify its parameters and obtains the filter inductor and output capacitor of the DC-DC converter.

6. An electronic device, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the training method as described in any one of claims 1-4 or the parameter identification method as described in claim 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to execute the training method as described in any one of claims 1-4 or the parameter identification method as described in claim 5.

8. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the training method as described in any one of claims 1-4 or the parameter identification method as described in claim 5.

Images