A method and equipment for capacitor health monitoring and inverter operation regulation

Abstract

This invention discloses a method and device for capacitor health monitoring and inverter operation adjustment. The method includes the following steps: training a pre-trained model containing a feature extraction layer based on a sample dataset of sample capacitors, the feature extraction layer being used to extract degradation features of the sample capacitors; connecting the feature extraction layer with a target adaptation layer to obtain an initial model; training the initial model based on a target dataset of a target capacitor and a target inverter to obtain a health monitoring model adapted to the target capacitor; using the health monitoring model to obtain the health assessment results of the target capacitor and / or the operating condition assessment results of the target inverter based on the real-time sampled data stream of the target capacitor; using the health monitoring model to obtain the health status value of the target capacitor based on the real-time dataset of the target capacitor, and adjusting the operating parameters of the inverter based on the health status value.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation technology, and in particular to a method and equipment for capacitor health monitoring and inverter operation regulation. Background Technology

[0002] As the core component of DC-side energy storage and filtering and AC-side harmonic suppression in inverters, the health status of capacitors directly affects the operational stability and output power quality of inverters. In related technologies, continuous monitoring of capacitor quality and health status is used to ensure the long-term reliable operation of inverters.

[0003] However, in complex operating scenarios characterized by drastic load fluctuations, significant changes in ambient temperature, and frequent variations in grid parameters, the limitations of these technologies are becoming increasingly apparent. For example, physical model-based monitoring methods are susceptible to operational interference and have large errors, failing to capture complex degradation characteristics; traditional machine learning methods have extremely low accuracy in scenarios with small amounts of field data; and fixed-condition regulation methods operate only according to rated parameters without considering the real-time health status of the capacitors for dynamic optimization. In other words, the aforementioned technologies are insufficient to effectively slow down capacitor degradation, which can lead to a decline in inverter output performance or even system failure. Summary of the Invention

[0004] One objective of this invention is to provide a capacitor health monitoring and inverter operation regulation method that can solve at least one defect in the above-mentioned background art.

[0005] Another object of the present invention is to provide an apparatus capable of implementing a capacitor health monitoring and inverter operation regulation method that solves at least one of the defects in the above-mentioned background art.

[0006] The first aspect of this invention provides a method for capacitor health monitoring and inverter operation adjustment, comprising the following steps: training a pre-trained model containing a feature extraction layer based on a sample dataset of sample capacitors, wherein the feature extraction layer is used to extract degradation features of the sample capacitors; connecting the feature extraction layer with a target adaptation layer to obtain an initial model; training the initial model based on a target dataset of a target capacitor and a target inverter to obtain a health monitoring model adapted to the target capacitor; obtaining a health assessment result of the target capacitor and / or an operating condition assessment result of the target inverter using the health monitoring model based on a real-time sampled data stream of the target capacitor; obtaining a health status value of the target capacitor using the health monitoring model based on a real-time dataset of the target capacitor; and adjusting the operating parameters of the inverter based on the health status value.

[0007] As a preferred embodiment, each target data in the target dataset is a multivariate time series, and the target data includes a subset of target electrical data for the target capacitor and a subset of target operating condition data for the target inverter; based on the subset of target operating condition data, the target dataset is divided into operating condition features to obtain at least two sub-class datasets, and each sub-class dataset is used to train the initial model.

[0008] As a preferred embodiment, dividing the target dataset to obtain the sub-datasets includes the following process: obtaining operating condition feature vectors based on the target operating condition data subsets, wherein the operating condition feature vectors at least represent the load state and / or thermal state; standardizing the operating condition feature vectors; and dividing the target dataset into at least two sub-datasets based on the standardized operating condition feature vectors using a clustering algorithm.

[0009] As a preferred embodiment, obtaining the health monitoring model includes the following process: freezing the bottom weights of the feature extraction layer in the initial model; training the initial model using each of the subclass datasets respectively, and minimizing the value of the loss function through an optimization algorithm to obtain the trained health monitoring model, which is used to establish the coupling relationship between the health indicators of the target capacitor and the operating conditions of the target inverter.

[0010] As a preferred embodiment, obtaining the pre-trained model based on the sample dataset includes the following steps: acquiring the sample dataset, which includes multivariate time series data of sample capacitors throughout their entire lifespan; normalizing the sample dataset; constructing a neural network architecture containing the feature extraction layer, which is used to extract degradation features from the normalized sample dataset; supervising the neural network architecture using health indicators of the sample capacitors as supervisory labels, and minimizing the value of the loss function through an optimization algorithm to obtain the pre-trained model.

[0011] As a preferred embodiment, connecting the feature extraction layer and the target adaptation layer to obtain the initial model includes the following process: removing or freezing the weights of the output layer of the pre-trained model, while retaining the feature extraction layer; connecting the target adaptation layer after the feature extraction layer to obtain the initial model, wherein the target adaptation layer contains one to two hidden units; the initial model is represented as Y=f new (f pre (a)), where f pre For the feature extraction layer, f new For the target adaptation layer, a is a feature in the target dataset input to the initial model, and Y is the health index of the target capacitor.

[0012] As a preferred embodiment, the health monitoring model employs a multi-task architecture. When using the health monitoring model to obtain the health assessment results and the work condition assessment results, the following process is included: acquiring the real-time sampling data stream; standardizing the real-time sampling data stream; based on the standardized real-time sampling data stream, utilizing the feature extraction layer and processing it through a regression layer, outputting a quantified health assessment result, denoted as z. t Based on the standardized real-time sampled data stream, the feature extraction layer is used, and the output is processed by the classification layer to produce a working condition assessment result, which is denoted as q. pred,t ; Generate health status trajectory G=(z t ,q pred,t ,t), where t is the timestamp.

[0013] As a preferred embodiment, the process of obtaining the health status value using the health monitoring model and adjusting the inverter's operating parameters based on the health status value includes the following steps: acquiring the real-time dataset; calculating the cumulative damage amount by integration using the health monitoring model at fixed intervals based on the real-time dataset, subtracting the cumulative damage amount from a preset initial health value to obtain the health status value; constructing an electrothermal stress function that includes the actual switching frequency, and solving for the optimal switching frequency within a preset frequency range with the goal of minimizing the value of the electrothermal stress function when the rate of change of the health status value is less than a preset adjustment threshold.

[0014] As a preferred embodiment, after adjusting the operating parameters of the inverter, real-time electrical data of the target capacitor is recorded to form an adjustment dataset; when the amount of data in the adjustment dataset is greater than or equal to a preset data amount threshold, the health monitoring model is updated through incremental learning based on the adjustment dataset.

[0015] A second aspect of the present invention provides an apparatus comprising: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program to implement the capacitor health monitoring and inverter operation regulation method as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: In this application, a pre-trained model is obtained by training on a large sample dataset. The feature extraction layer of the pre-trained model is then connected to the target adaptation layer to obtain an initial model. Furthermore, the initial model is fine-tuned using a target dataset with a smaller dataset to obtain a health monitoring model. This transfer learning approach reduces the data requirements of the target dataset for the health monitoring model, thereby improving the accuracy and precision of the monitoring results in scenarios with limited data. In other words, it enhances the accuracy and precision of the health assessment results and the target inverter's operating condition assessment results. Moreover, the health monitoring model is used to obtain the health status value of the target capacitor, and the inverter's operating parameters are adjusted based on this health status value. In other words, proactively optimizing the inverter's operating parameters helps to slow down capacitor degradation. Attached Figure Description

[0017] Figure 1 This is a flowchart of a capacitor health monitoring and inverter operation adjustment method according to some embodiments of this application. Detailed Implementation

[0018] The present invention will now be further described in conjunction with specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0019] The terms “comprising” and “having”, and any variations thereof, in the specification and claims of this application are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0020] A method for capacitor health monitoring and inverter operation regulation, such as Figure 1 As shown, the process includes the following steps: First, based on a sample dataset of sample capacitors, a pre-trained model containing a feature extraction layer is trained. This feature extraction layer is used to extract degradation features of the sample capacitors. Second, the feature extraction layer is connected to a target adaptation layer to obtain an initial model. Third, based on a target dataset of the target capacitor and the target inverter, the initial model is trained to obtain a health monitoring model adapted to the target capacitor. Fourth, based on the real-time sampled data stream of the target capacitor, the health assessment result z of the target capacitor is obtained using the health monitoring model. t and / or the operating condition assessment results q of the target inverter pred,t Based on the real-time dataset of the target capacitance, a health monitoring model is used to obtain the health status value H of the target capacitance. c (t), based on health status value H c (t) Adjust the operating parameters of the inverter.

[0021] It can be understood that a pre-trained model is obtained by training on a large sample dataset, and then the feature extraction layer of the pre-trained model is connected to the target adaptation layer to obtain the initial model. Further, the initial model is fine-tuned using a target dataset with only a small amount of data to obtain the health monitoring model. In this way, through transfer learning, the data requirements of the target dataset for the health monitoring model can be reduced, thereby helping to improve the accuracy and precision of the monitoring results in scenarios with small amounts of data. In other words, it can improve the health assessment results. t The operating condition evaluation results of the target inverter q pred,t The accuracy and precision of the target capacitor are improved. Furthermore, a health monitoring model is used to obtain the health status value H of the target capacitor. c (t), and based on the health status value H c (t) Adjusting the inverter's operating parameters; in other words, actively optimizing the inverter's operating parameters helps to delay capacitor degradation.

[0022] It is worth mentioning that the sample capacitor and the target capacitor can be the same type of capacitor or similar types of capacitors. This helps to make the sample capacitor and the target capacitor more similar in terms of electrical characteristics, material structure, and failure mechanism, thereby improving the accuracy and precision of the monitoring results of the health monitoring model. In at least one embodiment, the target capacitor is a polypropylene film capacitor, and the sample capacitor is a polyester film capacitor.

[0023] Furthermore, the method in this embodiment can be used to train the initial model using different target datasets to obtain a health monitoring model suitable for different target capacitors and target inverters. In other words, after training to obtain the pre-trained model, the feature extraction layer of the pre-trained model can be reused multiple times, thereby improving the ease of development of the health monitoring model and reducing the development cost of the health monitoring model.

[0024] In some embodiments, such as Figure 1 As shown, each target data in the target dataset is a multivariate time series. The target data includes a subset of target electrical data for the target capacitor and a subset of target operating condition data for the target inverter. Based on the subset of target operating condition data, the target dataset is divided into operating condition features to obtain at least two sub-class datasets. Each sub-class dataset is used to train the initial model.

[0025] It is understandable that training the initial model with sub-datasets of different operating conditions can enable the obtained health monitoring model to better adapt to various operating conditions of the inverter, thereby improving the health assessment results. t The operating condition evaluation results of the target inverter q pred,tThe accuracy and precision of the model are improved. Furthermore, fine-tuning the initial model using different sub-datasets allows the health monitoring model to more accurately detect the degradation characteristics of the target capacitor under different operating conditions. This helps filter noise interference caused by changes in inverter operating conditions, thus facilitating the subsequent acquisition of the target capacitor's health status value H. c (t), and based on the health status value H c (t) This provides a more reliable basis for adjusting the inverter's operating parameters.

[0026] In at least one embodiment, the features in the target electrical data subset include the target capacitor voltage V. c (t), target current I c (t), target temperature T c (t) and target ripple current I ripple (t); The characteristics of the target operating condition data subset include the target output power P out (t), target switching frequency f sw (t) and target ambient temperature T env (t). Specifically, the target capacitor voltage V is collected. c (t), target current I c (t), target temperature T c (t), target ripple current I ripple (t), Target output power P out (t), target switching frequency f sw (t) and target ambient temperature T env (t) is integrated and time-series aligned based on timestamps to form a target dataset with temporal correlation. The target dataset can be represented as , where t i The sampling time, i.e., the timestamp, t i It can be implemented as a GPS timestamp, V c (t i ), I c (t i ), T c (t i ), I ripple (t i ), P out (t i ), f sw (t i ) and T env (t i ) are respectively t i The target capacitor voltage, target current, target temperature, target ripple current, target output power, target switching frequency, and target ambient temperature at any given time, with N being the total number of data entries.

[0027] It should be understandable that the target capacitor voltage V c (t), target current I c (t), target temperature T c (t) and target ripple current I ripple (t) can be measured and acquired by voltage sensor, current sensor, temperature sensor and ripple current sensor at the target capacitor terminal, respectively, and the target ripple current I ripple (t) can also be obtained through Fast Fourier Transform. Target output power P out (t) can be obtained through a power analyzer; the target switching frequency f sw (t) can be extracted from the switch drive signal, the target ambient temperature T env (t) can be acquired by a temperature sensor outside the inverter.

[0028] Furthermore, setting a uniform sampling frequency for all features of the target dataset, for example, to 10kHz, helps ensure compliance with the Nyquist criterion and simplifies the subsequent complexity of forming the target dataset. Even further, CRC checksums can be used during data transmission to enhance the integrity and reliability of the target dataset.

[0029] It is worth mentioning that those skilled in the art can adjust the types of features and acquisition methods in the target dataset according to the actual situation, and such adjustments all fall within the protection scope of this application.

[0030] In some embodiments, dividing the target dataset to obtain sub-datasets includes the following process: obtaining operating condition feature vectors based on the target operating condition data subsets, wherein the operating condition feature vectors at least represent the load state and / or thermal state; standardizing the operating condition feature vectors; and dividing the target dataset into at least two sub-datasets based on the standardized operating condition feature vectors using a clustering algorithm.

[0031] It is understandable that using operating condition feature vectors that characterize the load state and / or thermal state of the target inverter helps extract features related to the inverter's operating conditions from the target dataset, thus providing clear direction for subsequent target dataset partitioning. Furthermore, clustering algorithms can efficiently divide the target dataset into sub-datasets corresponding to different operating conditions, avoiding the subjectivity and limitations of manual partitioning and providing a higher-quality data foundation for subsequent initial model training. In addition, clustering algorithms reduce data redundancy, thereby improving the efficiency of subsequent initial model training and helping to avoid overfitting.

[0032] In at least one embodiment, the operating condition feature vector includes average load power. Average temperature and load fluctuation standard deviation Where W is the span of the sliding window; for example, when the sampling frequency is 10kHz and W is 1000, the corresponding time length is 0.1s. It should be understood that the combination of average load power, mean temperature, and standard deviation of load fluctuation reflects the inverter's load intensity, thermal state, and the degree of dynamic load change, thus helping to accurately divide the data into subclasses corresponding to different operating conditions using clustering algorithms. It is worth mentioning that those skilled in the art can adjust the specific content of the operating condition feature vector according to actual conditions, and such adjustments fall within the protection scope of this application.

[0033] Furthermore, the operating condition feature vector is represented as s. The operating condition feature vector s is then standardized to obtain the standardized operating condition feature vector. , where μ s Let σ be the mean of the characteristic vector s of the working condition. s Let be the standard deviation of the operating condition feature vector s. It should be understood that standardizing the operating condition feature vector s helps eliminate the influence of dimensions.

[0034] In at least one embodiment, the clustering algorithm is implemented as a k-means clustering algorithm, with the objective function being: Where K is the number of operating condition categories, μ k As the cluster center, C k Let be the dataset of the k-th class. Furthermore, each subclass dataset can be represented as... , where D i For the i-th data in the target dataset, s i Let be the working condition feature vector corresponding to the i-th data in the target dataset.

[0035] In a preferred embodiment, K is set to 4 to divide the target dataset into four sub-datasets, corresponding to the target inverter's light load condition, rated load condition, overload condition, and high / low temperature condition, respectively. It is worth noting that those skilled in the art can adjust the value of K according to actual conditions so that each sub-dataset corresponds to different inverter operating conditions; such adjustments fall within the protection scope of this application.

[0036] In a preferred embodiment, each subclass dataset is stored in a structured HDF (Hierarchical Data Format) format and categorized according to operating condition labels, so as to facilitate the extraction of the corresponding subclass dataset according to the inverter operating condition when training the initial model; in other words, the retrieval flexibility of the subclass dataset is improved, thereby improving the efficiency of training the initial model.

[0037] In some embodiments, obtaining a pre-trained model based on a sample dataset includes the following steps: acquiring a sample dataset, which includes multivariate time series data of sample capacitors throughout their lifespan; normalizing the sample dataset; constructing a neural network architecture containing a feature extraction layer, which is used to extract degradation features from the normalized sample dataset; supervising the neural network architecture using health indicators of the sample capacitors as supervisory labels, and minimizing the value of the loss function through an optimization algorithm to obtain a pre-trained model.

[0038] It should be understood that the sample dataset encompasses the time-series records of the sample capacitors from initial operation to eventual failure. Using capacitor health indicators as supervisory labels, pre-training the neural network containing the feature extraction layer using the sample dataset helps the feature extraction layer extract various degradation forms of the sample capacitors, such as abnormal degradation and long-term wear degradation, and form general degradation features. Consequently, the pre-trained model possesses good generalization ability and can be transferred and applied to target capacitors of similar types to the sample capacitors. In other words, when the feature extraction layer is connected to the target adaptation layer to build the initial model, the initial model possesses prior knowledge related to degradation features, thus helping to ensure high monitoring accuracy and precision of the health monitoring model even with limited data in the target dataset.

[0039] In at least one embodiment, the sample dataset contains more than 1000 data points, each sample being a multivariate time series containing features such as the capacitance value, equivalent series resistance, temperature, and operating voltage of the sample capacitor. In a preferred embodiment, the sampling frequency for each feature of the sample dataset is 1 Hz to capture the degradation process of the sample capacitor. The capacitance value, equivalent series resistance, temperature, and operating voltage of the sample capacitor can be acquired by an integrated sensor or obtained through other methods; this application does not impose specific limitations on this.

[0040] In at least one embodiment, the original values ​​of each feature in the sample dataset are denoted as x, and the normalization process is implemented by scaling each feature to the range [0,1]. The normalized feature values ​​are... , where x min x is the minimum value of this feature in the sample dataset. max This represents the maximum value of the feature in the sample dataset. It should be understood that normalizing the features in the sample dataset helps eliminate differences in units and also improves the convergence speed of training.

[0041] In at least one embodiment, considering the local pattern characteristics of the sample dataset, in other words, to more effectively extract the features of abnormal degradation of the sample capacitor, the neural network architecture includes a one-dimensional convolutional neural network. Specifically, the input side of the one-dimensional convolutional neural network is used to receive the normalized sample dataset. The dimension of the one-dimensional convolutional neural network is T×D, where T is the number of time steps and D is the feature dimension. In a preferred embodiment, T is set to 1000 and D is set to 4, corresponding to the four features of the sample capacitor: capacitance value, equivalent series resistance, temperature, and operating voltage. It is worth mentioning that those skilled in the art can adjust the types and number of features in the sample dataset, and correspondingly adjust the feature dimension of the one-dimensional convolutional neural network according to the actual situation. Such adjustments all fall within the protection scope of this application.

[0042] In at least one embodiment, considering the long-term dependency characteristics of the sample dataset—in other words, to more effectively extract the long-term loss and degradation features of the sample capacitance—the neural network architecture includes a Long Short-Term Memory (LSTM) network. Specifically, the hidden state update formula for the LSM network is h. t =LSTM(x t ,h t-1 ,c t-1 ), where h t The hidden state at the current time, x t c is the normalized feature vector at the current time. t In cellular state, h t-1 c is the previously hidden state. t-1 This represents the cell state at a previous time.

[0043] In at least one embodiment, the health metric of the sample capacitor is implemented as the remaining useful life percentage y. ’ , where y ’ ∈[0,1],y ’ When y = 0, it indicates that the capacitor is completely ineffective. ’ A value of 1 indicates that the capacitor is completely healthy. Furthermore, the loss function is implemented as mean squared error. Where N1 is the batch size. This represents the true percentage of the remaining lifespan of the sample capacitors. This is a predicted value representing the proportion of remaining lifespan of the sample capacitors. In a preferred embodiment, the Adam (Adaptive Moment Estimation) algorithm is used, with the learning rate set to 0.001, the batch size N1 set to 32 or 64, and the number of learning epochs set to 100, so that the mean squared error L... ’Minimize. It is worth mentioning that those skilled in the art can adjust the specific representation method of the health indicators of the sample capacitance and the specific settings of the corresponding optimization algorithm according to the actual situation, and such adjustments all fall within the protection scope of this application.

[0044] In at least one embodiment, the sample data in the sample dataset is divided into a training set, a validation set, and a test set in a 7:1.5:1.5 ratio, respectively for training, validating, and testing the pre-trained model, thereby reducing the risk of overfitting in the pre-trained model. Furthermore, an early stopping strategy is implemented during training; in other words, when the mean squared error L is detected... ’ Training is terminated when descent stops, which helps avoid ineffective training, improves the quality of degenerate features extracted by the feature extraction layer, and helps ensure that the feature extraction layer has good generalization ability and robustness. It is worth mentioning that those skilled in the art can adjust the ratio of training set, validation set and test set according to actual conditions, and such adjustments fall within the protection scope of this application.

[0045] In some embodiments, connecting the feature extraction layer and the target adaptation layer to obtain an initial model includes the following steps: removing or freezing the weights of the output layer of the pre-trained model, while retaining the feature extraction layer; connecting the target adaptation layer after the feature extraction layer to obtain the initial model, wherein the target adaptation layer contains one to two hidden units; the initial model is represented as Y=f new (f pre (a)), where f pre For feature extraction layer, f new For the target adaptation layer, 'a' represents the features in the target dataset of the initial input model, and 'Y' represents the health index of the target capacitor.

[0046] It is understandable that the target adaptation layer can be implemented as a regression layer or a classification layer. Furthermore, the target adaptation layer contains only one or two hidden units, which reduces the number of trainable parameters and lowers the risk of overfitting when training the initial model using the target dataset. This helps ensure that the health monitoring model has high monitoring accuracy and precision, even with limited target dataset data, to detect the impact of different inverter operating conditions on target capacitor degradation. In addition, this setup reduces the complexity of the initial model, thereby improving the training convergence speed and helping to ensure that the trained health monitoring model retains prior knowledge related to degradation features.

[0047] In some embodiments, obtaining a health monitoring model includes the following process: freezing the bottom weights of the feature extraction layer in the initial model; training the initial model using each subclass dataset and minimizing the value of the loss function through an optimization algorithm to obtain a trained health monitoring model. The health monitoring model is used to establish the coupling relationship between the health index Y of the target capacitor and the operating condition of the target inverter.

[0048] As mentioned earlier, the target dataset is divided into sub-datasets corresponding to different operating conditions, with each sub-dataset containing a relatively small amount of data, for example, less than 100 samples. By training the initial model on each sub-dataset separately, a health monitoring model adapted to the target capacitor can be obtained in a relatively short time. In other words, the degradation features of the sample capacitors retained by the feature extraction layer can be transformed into degradation features adapted to the target capacitor in a relatively short time, thereby enabling the health monitoring model to have high monitoring accuracy and precision.

[0049] In at least one embodiment, the health index Y of the target capacitor is implemented as the remaining useful life percentage y, where y∈[0,1], y=0 indicates that the capacitor is completely failed, and y=1 indicates that the capacitor is completely healthy. Further, the loss function is implemented as mean squared error. Where N2 is the batch size, y i This is a true label indicating the percentage of remaining lifespan of the target capacitor. This represents the model prediction of the remaining lifespan percentage of the target capacitor. In a preferred embodiment, the Adam algorithm is used to minimize the mean squared error L. Specifically, the weight update rule is as follows: , where θ contains f new weights and f pre At least some of the top-level weights, where η1 is the learning rate. The loss gradient is set; further, the learning rate η1 is set to 0.0001, the batch size N2 is set to a natural number between 4 and 8, and the number of learning epochs is set to a natural number between 10 and 20. This setup is understandable as it adapts to subclass datasets with limited data volume, while also helping to prevent the destruction of prior knowledge related to degenerate features in the feature extraction layer, thereby reducing the risk of catastrophic forgetting in the health monitoring model during training through smooth transfer.

[0050] It is worth mentioning that those skilled in the art can adjust the specific representation method of the health index Y of the target capacitor and the specific settings of the corresponding optimization algorithm according to the actual situation, and such adjustments all fall within the protection scope of this application.

[0051] In at least one embodiment, the coupling relationship between the health index Y of the target capacitor and the operating condition of the target inverter is expressed as Y=g(s)+ε, where s is the operating condition feature vector mentioned above, ε is the error term, g is a relational expression containing s, and g is obtained by the health monitoring model during the training process.

[0052] In some embodiments, such as Figure 1 As shown, the health monitoring model has a multi-task architecture. The model is used to obtain health assessment results z. t and operating condition assessment results q pred,t The process includes the following steps: acquiring real-time sampled data streams and standardizing them; based on the standardized real-time sampled data streams, using a feature extraction layer and a regression layer, outputting a quantified health assessment result z. t Based on the standardized real-time sampled data stream, a feature extraction layer is used, followed by a classification layer to output the operating condition assessment result q. pred,t ; Generate health status trajectory G=(z t ,q pred,t ,t), where t is the timestamp.

[0053] It should be understood that the health monitoring model in this embodiment is able to simultaneously output health assessment results. t and operating condition assessment results q pred,t And generate a health state trajectory G, through which the health assessment result z of the target capacitor can be obtained. t Working condition assessment results q pred,t Correspondingly, by using timestamps to record the historical evolution process, monitoring efficiency can be improved, and real-time monitoring can be achieved, while providing a more comprehensive data foundation for predictive maintenance of inverters.

[0054] In at least one embodiment, the real-time sampled data stream includes the operating voltage v of the target capacitor. t Current i t Temperature T t and equivalent series resistance R esr,t Features such as these are represented by real-time sampled data as x. tr =[v t i t ,T t ,R esr,t ] T In a preferred embodiment, the sampling frequency of various features of the real-time sampled data is 1 kHz. Wherein, the operating voltage v of the target capacitor... t Current i t Temperature T t and equivalent series resistance R esr,tThese features can be acquired by integrated sensors or by other means, and this application does not impose any specific restrictions on them.

[0055] Furthermore, regarding the real-time sampled data x tr Standardization processing is performed to obtain standardized real-time sampling data. , where μ and σ are the feature mean and standard deviation calculated during training to obtain the health monitoring model, respectively. It should be understood that by analyzing real-time sampled data x... tr Standardization helps eliminate the influence of dimensions.

[0056] It is worth mentioning that those skilled in the art can adjust the types of features in the real-time sampling data stream according to the actual situation, and such adjustments all fall within the protection scope of this application.

[0057] In at least one embodiment, a fully connected regression layer is used to assess the health results. , where w z For the weight vector, This is the transpose of the weight vector. This is the feature extraction layer for the health monitoring model, used to output high-dimensional feature vectors, b z This is a bias term. Furthermore, b z ∈[0,1],b z When the value is 0, it indicates that the capacitor is completely ineffective. z A value of 1 indicates that the capacitor is completely healthy. This setting is intended to visually reflect the degree of degradation of the target capacitor, allowing operators to quickly assess its remaining lifespan.

[0058] In at least one embodiment, a softmax classification layer is used to evaluate the working condition results. ,in, K is the number of operating condition categories mentioned above, q t Operating condition categories, such as light load condition, rated load condition, overload condition, and high / low temperature condition, w q This is the weight vector for the current working condition category. b is the transpose of the weight vector for the current working condition category. q w is the bias term for the current operating condition category. k This is the weight vector corresponding to each working condition category when traversing all working condition categories. b is the transpose of the weight vector corresponding to each working condition category when traversing all working condition categories. k This refers to the bias term corresponding to each operating condition category when traversing all operating condition categories.

[0059] In at least one embodiment, the health status trajectory In other words, the timestamp t of G in the health status trajectory can be a relative sequence number, which helps avoid cumbersome time alignment and format conversion operations, thereby improving the real-time processing efficiency and response speed of the health monitoring model. In at least one other embodiment, the timestamp t of G in the health status trajectory can be a GPS timestamp, so as to associate the health status trajectory G with other working logs of the inverter, providing a reliable time basis for subsequent fault analysis and maintenance tracing.

[0060] In at least one embodiment, the health monitoring model is optimized into a lightweight format using a frame conversion tool, and the optimized health monitoring model is deployed to the edge computing unit or host computer system of the target inverter. It should be understood that this helps ensure the health monitoring model runs in real-time on computationally limited devices, reducing inference latency, for example, enabling inference latency to be controlled within milliseconds.

[0061] In some embodiments, such as Figure 1 As shown, a health status value H is obtained using a health monitoring model. c (t), based on health status value H c (t) Adjusting the inverter's operating parameters includes the following processes: acquiring real-time data sets; adjusting parameters at fixed intervals. Based on real-time datasets, a health monitoring model is used to calculate the cumulative damage amount F through integration. The difference between the preset initial health value H0 and the cumulative damage amount F is then used to obtain the health status value H. c (t); Construct a system containing the actual switching frequency f swr The electrothermal stress function S(f) swr When the health status value H c When the rate of change of (t) is less than the preset adjustment threshold, the electrothermal stress function S(f) is minimized. swr The objective is to find the optimal switching frequency f within a preset frequency range. best .

[0062] It should be understood that this embodiment is based on the health status value H. c The rate of change of (t) triggers the regulation mechanism, and the electrothermal stress function S(f) is constructed. swr To solve for the optimal switching frequency f best This allows the inverter to proactively adjust its operating parameters based on the real-time aging status of the target capacitor, reducing its electrothermal stress and thus helping to slow down its degradation while maintaining high power output quality. It is worth mentioning that, in obtaining the optimal switching frequency f... best Then, the optimal switching frequency f can be determined. bestThe frequency is converted to the carrier frequency and updated via the PWM (Pulse Width Modulation) module, thereby adjusting the drive signals of the inverter's power devices in real time.

[0063] In at least one embodiment, the features in the real-time dataset include the real-time temperature T of the target capacitor. cr Real-time voltage stress V sr and real-time ripple current I rr Furthermore, the cumulative damage amount Health status value Where H0 is the preset initial health value, α is the aging coefficient, and λ1, λ2, and λ3 are weighting coefficients, generated by the health monitoring model. It should be understood that the features in the real-time dataset may also include the real-time load current I of the target inverter. Lr and ambient temperature T envr It is worth mentioning that the real-time dataset and the real-time sampling data stream mentioned above can be merged, which helps to avoid dual-path reading and data redundancy.

[0064] In a preferred embodiment, the sampling frequency of various features of the real-time dataset is 1 kHz, with a fixed period. The value is 10ms; in other words, the health status value H is continuously calculated and updated using real-time datasets acquired every 10ms time window. c (t), thus providing a basis for subsequent calculation of the health status value H. c The rate of change of (t) provides reliable data support, helping to ensure timely adjustment of the actual switching frequency f of the target inverter. swr It is worth mentioning that those skilled in the art can adjust the fixed cycle according to the actual situation. The values ​​of these adjustments all fall within the scope of protection of this application.

[0065] In at least one embodiment, the health status value H c The rate of change of (t) is expressed as Where β is a preset adjustment threshold, and its value is negative. It should be understood that, over time, the health status value H... c (t) gradually decreases, therefore the health status value H c The rate of change of (t) is negative when the health status value H c When the rate of change of (t) is less than the adjustment threshold β, it indicates that the health status value H c When the absolute value of the rate of change of (t) increases, that is, the degradation rate of the target capacitor accelerates, it is necessary to adjust the actual switching frequency f of the target inverter. swr Adjustments will be made.

[0066] In at least one embodiment, the adjustment threshold β is set to -0.01 / s. Furthermore, each time the health status value H is updated... c (t), then calculate the fixed period. Health status value H within c The rate of change of (t) allows for timely detection of accelerated degradation of the target capacitor. It is worth noting that those skilled in the art can adjust the value of the adjustment threshold β and calculate the health status value H according to the actual situation. c The specific calculation method of the rate of change of (t), and such adjustments all fall within the protection scope of this application.

[0067] In at least one embodiment, the electrothermal stress function Where γ1 and γ2 are stress coefficients, which can be preset values ​​or generated by a health monitoring model. Furthermore, the actual switching frequency f... swr The preset frequency range is [f min ,f max ], where f min To ensure the lower frequency limit for stable operation of the target inverter, for example, 5kHz, f max The upper frequency limit to ensure stable operation of the target inverter is set, for example, 20kHz.

[0068] In at least one embodiment, the optimal switching frequency f is solved iteratively using a gradient descent algorithm. best Specifically, the optimal switching frequency f is determined. best The problem can be represented as That is, within the preset frequency range [f min ,f max Within this range, a practical switching frequency f that minimizes the electrothermal stress function S is obtained. swr The actual switching frequency f swr As the optimal switching frequency f best Iteratively solve for the optimal switching frequency f. best The process can be represented as Where η2 is the learning rate and n is the number of iterations. In a preferred embodiment, η2 is 0.1.

[0069] In some embodiments, such as Figure 1 As shown, after adjusting the inverter's operating parameters, real-time electrical data of the target capacitor is recorded to form an adjustment dataset. When the amount of data in the adjustment dataset is greater than or equal to a preset data threshold, the health monitoring model is updated through incremental learning based on the adjustment dataset. It should be understood that this setting allows the health monitoring model to continuously update in accordance with the actual state of the target inverter and target capacitor, thereby helping to ensure that the health monitoring model maintains a high degree of adaptability to the target inverter and target capacitor over a long period.

[0070] In at least one embodiment, the characteristics of the adjustment dataset include the real-time temperature T of the target capacitor. cr (t), Real-time voltage stress V sr (t), Real-time ripple current I rr (t), actual switching frequency f swr and the electrothermal stress function value S; in other words, the adjustment dataset can be expressed as .

[0071] In at least one embodiment, during incremental learning, the health monitoring model is able to optimize the health status value H using the least squares method. c (t) Calculate the weighting coefficients λ1, λ2, and λ3 in the formula. Specifically, solving for the weighting coefficients λ1, λ2, and λ3 can be expressed as: Where λ = [λ1, λ2, λ3] T , Health status value H c The observed values ​​of (t), Health status value H c The predicted value of (t). It should be understood that by optimizing the weighting coefficients λ1, λ2, and λ3, the health status value H can be improved. c The accuracy of the calculation of (t) can be improved, thus more accurately reflecting the degree of degradation of the target capacitor.

[0072] In a preferred embodiment, the data volume threshold is set to 1000 records. This helps ensure that the health monitoring model can obtain sufficient data to improve the accuracy and precision of monitoring, while also reducing the computation time and resource consumption of incremental learning.

[0073] An apparatus includes a processor and a memory; the memory stores a computer program, and the processor executes the computer program to implement the above-described capacitor health monitoring and inverter operation regulation method.

[0074] The basic principles, main features, and advantages of this invention have been described above. Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made without departing from the spirit and scope of the invention, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection claimed by this invention is defined by the appended claims and their equivalents.

Claims

1. A method for capacitor health monitoring and inverter operation regulation, characterized in that, Includes the following steps: Based on a sample dataset of sample capacitance, a pre-trained model containing a feature extraction layer is trained, which is used to extract degradation features of sample capacitance. The feature extraction layer is connected to the target adaptation layer to obtain the initial model; The initial model is trained based on the target dataset of the target capacitor and the target inverter to obtain a health monitoring model adapted to the target capacitor. Based on the real-time sampling data stream of the target capacitor, the health assessment results of the target capacitor and / or the operating condition assessment results of the target inverter are obtained using the health monitoring model. Based on the real-time dataset of the target capacitor, the health status value of the target capacitor is obtained using the health monitoring model. The operating parameters of the inverter are adjusted based on the health status value, including the following process: constructing an electrothermal stress function that includes the actual switching frequency; when the rate of change of the health status value is less than a preset adjustment threshold, minimizing the value of the electrothermal stress function is the objective, and the optimal switching frequency is solved within a preset frequency range.

2. The capacitor health monitoring and inverter operation adjustment method according to claim 1, characterized in that, Each target data in the target dataset is a multivariate time series, and the target data includes a subset of target electrical data for the target capacitor and a subset of target operating condition data for the target inverter; Based on the target working condition data subset, the target dataset is divided into working condition features to obtain at least two sub-class datasets, and each sub-class dataset is used to train the initial model.

3. The capacitor health monitoring and inverter operation adjustment method according to claim 2, characterized in that, Dividing the target dataset into sub-datasets includes the following process: Based on the target operating condition data subset, an operating condition feature vector is obtained, wherein the operating condition feature vector at least represents the load state and / or thermal state. The operating condition feature vector is standardized. Based on the standardized operating condition feature vector, the target dataset is divided into at least two sub-datasets using a clustering algorithm.

4. The capacitor health monitoring and inverter operation adjustment method according to claim 2, characterized in that, Obtaining the health monitoring model includes the following process: Freeze the bottom weights of the feature extraction layer in the initial model; The initial model is trained using each of the aforementioned subclass datasets, and the loss function is minimized using an optimization algorithm to obtain the trained health monitoring model. The health monitoring model is used to establish the coupling relationship between the health indicators of the target capacitor and the operating conditions of the target inverter.

5. The capacitor health monitoring and inverter operation adjustment method according to any one of claims 1-4, characterized in that, Obtaining the pre-trained model based on the sample dataset includes the following process: Obtain the sample dataset, which includes a multivariate time series of sample capacitance over its entire lifespan; The sample dataset is normalized. Construct a neural network architecture that includes the feature extraction layer, which is used to extract degenerate features from the normalized sample dataset; Using the health indicators of sample capacitance as supervisory labels, the neural network architecture is trained under supervision, and the loss function is minimized through an optimization algorithm to obtain the pre-trained model.

6. The capacitor health monitoring and inverter operation adjustment method according to any one of claims 1-4, characterized in that, Connecting the feature extraction layer to the target adaptation layer to obtain the initial model includes the following process: Remove the output layer of the pre-trained model or freeze the weights of the output layer, while retaining the feature extraction layer; The target adaptation layer is connected after the feature extraction layer to obtain the initial model, wherein the target adaptation layer contains one to two hidden units; The initial model is represented as Y=f new (f pre (a)) wherein f pre is the feature extraction layer, f new is the target adaptation layer, a is a feature in the target dataset input to the initial model, and Y is a health indicator of the target capacitance.

7. The capacitor health monitoring and inverter operation adjustment method according to any one of claims 1-4, characterized in that, The health monitoring model is a multi-task architecture, and the process of obtaining the health assessment results and the working condition assessment results using the health monitoring model includes the following: Acquire the real-time sampling data stream and perform standardization processing on the real-time sampling data stream; Based on the standardized real-time sampling data stream, the feature extraction layer is utilized, and the quantitative health assessment result is output via a regression layer processing, denoted as z t ; Based on the real-time sampling data stream after standardization processing, the feature extraction layer is utilized, and the working condition evaluation result is output through the classification layer processing, and the working condition evaluation result is denoted as q pred,t ; A health state trajectory G = (z t , q pred,t , t) is generated, where t is a time stamp.

8. The capacitor health monitoring and inverter operation adjustment method according to any one of claims 1-4, characterized in that, The process of obtaining the health status value using the health monitoring model and adjusting the inverter's operating parameters based on the health status value includes the following steps: Obtain the real-time dataset; At a fixed period, based on the real-time dataset, the health monitoring model is used to calculate the cumulative damage amount through integration, and the difference between the preset initial health value and the cumulative damage amount is obtained to obtain the health status value.

9. The capacitor health monitoring and inverter operation adjustment method according to any one of claims 1-4, characterized in that, After adjusting the inverter's operating parameters, record the real-time electrical data of the target capacitor to form an adjustment dataset; When the amount of data in the adjustment dataset is greater than or equal to a preset data threshold, the health monitoring model is updated through incremental learning based on the adjustment dataset.

10. A device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program to implement the capacitor health monitoring and inverter operation regulation method as described in any one of claims 1-9.

Images