A method, device and management system for estimating the capacity of a supercapacitor
By collecting multi-condition data from supercapacitors and constructing an adaptive noise reduction model using adversarial networks and neural networks, the problem of poor generalization of supercapacitors in diverse application scenarios is solved, and the accuracy and adaptability of full-cycle capacity assessment are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-04-07
AI Technical Summary
Existing supercapacitor capacity assessment methods have poor generalization ability in diverse and varied application scenarios, making it difficult to optimize and update them for different scenarios. Furthermore, data acquisition is complex and samples are insufficient, resulting in low efficiency when training models.
By collecting source domain data under multiple operating conditions, adversarial networks are used for knowledge fusion and feature extraction to construct an adaptive noise reduction model. Combined with prior knowledge and neural networks, a basic transfer learning capacity assessment model is constructed and fine-tuned in segments to form a full-cycle capacity assessment model.
This improved the adaptability and generalization of the capacity assessment model to different operating conditions, reduced the impact of noise, and achieved more accurate capacity estimation.
Smart Images
Figure CN121327784B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electrochemical energy storage, and specifically relates to a method, device and system for estimating the capacity of a supercapacitor. Background Technology
[0002] In recent years, with the rapid growth of modern society's demand for emerging technologies such as renewable energy, smart grids, and electric vehicles, traditional energy storage devices can no longer fully meet the needs for efficient, reliable, and sustainable energy storage. While lithium-ion batteries excel in energy density, they have limitations in power density, cycle life, and safety. Compared to lithium batteries, supercapacitors can simultaneously achieve high energy density and high power density, and also possess advantages such as long cycle life, high safety and stability, and a wide operating temperature range, making them ideal electrochemical energy storage devices for industrial applications. However, with increasing cycle count, supercapacitors exhibit aging phenomena such as capacity reduction. Therefore, accurately assessing the capacity of supercapacitors is crucial for the safety of power system equipment and for maximizing the residual value of individual supercapacitor cells.
[0003] Currently, capacity assessment methods are mainly divided into precision instrument-based measurement methods and model-based methods. Model-based methods are further divided into mechanistic model-based methods, equivalent model-based methods, and data-driven artificial intelligence-based methods. However, the application scenarios of supercapacitors are becoming increasingly complex and diverse. The variety of energy storage technologies and environments makes it difficult to obtain capacity labels for supercapacitors. For some new samples or new usage environments, the required measurement data volume is large, the time is long, and the testing conditions are demanding. Therefore, the complex data acquisition process increases the difficulty of data acquisition, resulting in insufficient samples for model training. Thus, in practical applications, even smaller samples are often used for training, making transfer learning an ideal solution. However, the application scenarios of supercapacitors are becoming increasingly diversified, and traditional methods have poor generalization capabilities, making it difficult to optimize and update them for different scenarios. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention proposes a method for estimating the capacity of a supercapacitor, comprising the following steps:
[0005] Source domain data of multiple supercapacitors under multiple operating conditions are collected, and multi-source features containing strong correlation features of capacity decay are extracted based on the source domain data.
[0006] Data from the target domain is collected and knowledge is fused through an adversarial network based on prior knowledge of supercapacitors and the multi-feature features of the source domain to obtain fused features, pattern matching results, and multi-feature contribution spectrum.
[0007] The noise reduction parameters of multiple noise reduction models are set based on the prior knowledge of the supercapacitor, the fusion features, and the pattern matching results.
[0008] A neural network is constructed to learn the denoising characteristics of the denoising models of the multiple modes, thereby obtaining an adaptive denoising model;
[0009] A basic transfer learning capacity assessment model is constructed, and the basic transfer learning capacity assessment model is trained using target domain data. The fusion features, pattern matching results, multi-feature contribution spectrum, adversarial network and adaptive noise reduction model are transferred to the basic transfer learning capacity assessment model to obtain the first-stage capacity assessment model.
[0010] Based on the prior knowledge of the supercapacitor, the first-stage capacity assessment model is finely adjusted in segments to obtain the full-cycle capacity assessment model.
[0011] The full-cycle capacity assessment model is integrated into the supercapacitor capacity management system.
[0012] In one embodiment of the above method, the acquisition of raw characteristic data of multiple supercapacitors under multiple operating conditions further includes:
[0013] The voltage, current, temperature, humidity, appearance images, and component structure characteristics data of the multiple supercapacitors are acquired.
[0014] In one embodiment of the above method, extracting source domain multivariate features containing strongly correlated capacitance decay features based on the source domain data and prior knowledge of the supercapacitor further includes:
[0015] At the acquisition end, basic data processing and basic anomaly detection are performed on the original feature data;
[0016] Based on the original feature dataset stored in the cloud, the source domain multivariate features containing features strongly correlated with capacity decay are transformed according to the basic data processing results of the original feature data.
[0017] In one embodiment of the above method, the process of collecting target domain data and fusing knowledge through an adversarial network based on prior knowledge of supercapacitors and the multi-dimensional features of the source domain further includes:
[0018] The source domain multivariate features are enhanced based on the physical characteristics of the supercapacitor using an attention mechanism to obtain enhanced features;
[0019] Based on the enhanced features, the source domain multi-features, and part of the target domain data, an adversarial network for fused knowledge is constructed to obtain the fused features.
[0020] In one embodiment of the above method, the process of collecting target domain data and performing knowledge fusion through an adversarial network based on prior knowledge of supercapacitors and the multi-dimensional features of the source domain further includes:
[0021] To optimize feature selection for different working conditions, a probabilistic fusion learning matching method is used to perform feature matching on the fused features, resulting in the pattern matching results and an interpretable multi-feature contribution spectrum.
[0022] In one embodiment of the above method, setting the noise reduction parameters of the noise reduction model for multiple modes based on prior knowledge of the supercapacitor, the fusion features, and the mode matching results further includes:
[0023] Based on the fusion characteristics and prior knowledge of the supercapacitor, statistical analysis is performed on the data of the supercapacitor operating under different conditions to extract Gaussian filter parameters for the multiple modes.
[0024] In one embodiment of the above method, the denoising characteristics of the denoising model constructed by learning the multiple modes through a neural network further include:
[0025] The source domain multivariate features and the source domain multivariate features after denoising by the multiple modes are input into the neural network to train the neural network to learn the denoising characteristics of the multiple modes of denoising models.
[0026] In one embodiment of the above method, constructing a basic transfer learning capacity evaluation model and training the basic transfer learning capacity evaluation model using target domain data, and transferring the fusion features, pattern matching results, multi-feature contribution spectra, adversarial networks, and adaptive denoising models to the basic transfer learning capacity evaluation model further includes:
[0027] A domain adaptation transfer module is based on a fusion of convolutional neural networks (CNN) or transformer neural networks. The domain adaptation transfer module includes:
[0028] Extracting suitable input samples from the data sample can improve the quality of feature extraction.
[0029]
[0030] in, Represents the input sample of the source domain and , The number of samples in the source domain is represented by g, and the dimension is represented by g. Represents the input samples of the target domain and , Indicates the number of samples in the target domain. and Represents the loss function. This represents the feature extraction representation. Describes an auxiliary function. Represents a mapping function. Indicates the edge parameter.
[0031] In one embodiment of the above method, the segmented fine-tuning of the first-stage capacity assessment model based on prior knowledge of the supercapacitor further includes:
[0032] Based on the characteristics of the supercapacitor's entire life cycle in the source domain data, the target domain data of the supercapacitor is divided into multiple stages;
[0033] Set thresholds for the multiple stages;
[0034] The network layers of the first-stage capacity assessment model are fine-tuned based on the threshold values of the multiple stages.
[0035] The present invention also proposes a supercapacitor capacity estimation device for implementing any of the aforementioned methods, comprising:
[0036] The feature extraction and correlation analysis module is used to collect source domain data of multiple supercapacitors under multiple operating conditions, and extract source domain multivariate features containing strong correlation features of capacity decay based on the source domain data.
[0037] The knowledge data fusion pattern matching module is used to perform knowledge fusion through an adversarial network based on the target domain data, the prior knowledge of the supercapacitor, and the multi-feature features of the source domain to obtain fusion features, pattern matching results, and multi-feature contribution spectrum.
[0038] An adaptive noise reduction module is used to set the noise reduction parameters of multiple noise reduction models based on the prior knowledge of the supercapacitor, the fusion features, and the pattern matching results; and to construct a neural network to learn the noise reduction characteristics of the multiple noise reduction models to obtain an adaptive noise reduction model.
[0039] A two-stage transfer learning module is used to construct a basic transfer learning capacity evaluation model and train the basic transfer learning capacity evaluation model using target domain data. The fusion features, pattern matching results, multi-feature contribution spectrum, adversarial network, and adaptive noise reduction model are transferred to the basic transfer learning capacity evaluation model to obtain a first-stage capacity evaluation model. Furthermore, the first-stage capacity evaluation model is fine-tuned in segments based on the prior knowledge of the supercapacitor to obtain a full-cycle capacity evaluation model.
[0040] An integration module is used to integrate the full-cycle capacity assessment model into the supercapacitor capacity management system.
[0041] This invention also proposes a supercapacitor management system, comprising:
[0042] Supercapacitors
[0043] Controller
[0044] Multiple sensors, connected to the supercapacitor and / or controller, are used to acquire source domain data of the supercapacitor under multiple operating conditions, wherein...
[0045] The controller includes steps for implementing any of the aforementioned methods.
[0046] The present invention also proposes a storage medium for storing a computer control program for performing the steps of any of the methods described above.
[0047] Compared with existing technologies, the method disclosed in this invention (1) extracts features strongly correlated with capacity decay by using prior knowledge and data dual-drive in the data acquisition, preprocessing, and multi-physical feature extraction and correlation analysis modules for different operating conditions; (2) achieves more effective matching with corresponding patterns in the target domain through reasonable pattern matching; (3) realizes intelligent noise reduction for different operating conditions by learning the noise reduction characteristics of the noise reduction model of the multiple patterns through neural networks, thereby reducing the negative impact of adverse information on knowledge transfer; and (4) obtains a full-cycle capacity assessment model by segmented fine-tuning of the capacity assessment model. The above four aspects improve the adaptability and generalization of the capacity assessment model to different operating conditions.
[0048] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description
[0049] Figure 1 A flowchart illustrating a supercapacitor capacity estimation method in one embodiment of the present invention is shown.
[0050] Figure 2 A diagram illustrating the overall architecture of a supercapacitor capacity estimation method in one embodiment of the present invention is provided.
[0051] Figure 3 A block diagram illustrating a supercapacitor capacity estimation device according to an embodiment of the present invention is shown.
[0052] Figure 4 A block diagram illustrating a supercapacitor capacity management system according to an embodiment of the present invention is shown.
[0053] In the attached figures, the following labels are used:
[0054] Steps S1-S7…
[0055] 1…Supercapacitor
[0056] 2…sensors
[0057] 3…Source Domain Data
[0058] 4…Source Domain Multivariate Features
[0059] 5… Fusion Features
[0060] 6…pattern matching results
[0061] 7…Multi-feature contribution spectrum
[0062] 8…Basic Transfer Learning Capacity Assessment Model
[0063] 81… Domain Adaptation Migration Module
[0064] 10…Supercapacitor Capacity Estimation Device
[0065] 11…Feature Extraction and Correlation Analysis Module
[0066] 12…Knowledge Data Fusion Pattern Matching Module
[0067] 13…Adaptive noise reduction module
[0068] 14…Two-stage transfer learning module
[0069] 15…Integrated Module
[0070] 20… controller
[0071] 100…Supercapacitor Management System Detailed Implementation
[0072] The following specific embodiments, in conjunction with the accompanying drawings, illustrate the implementation methods disclosed in this invention. Those skilled in the art can understand the advantages and effects of this invention from the content disclosed in this specification. However, the following disclosure is not intended to limit the scope of protection of this invention. Without departing from the spirit of the invention, those skilled in the art can implement this invention with other different embodiments based on different viewpoints and applications.
[0073] For clarity, the figures shown in this invention are simplified schematic diagrams illustrating the basic structure of the invention. Therefore, the structures shown in the figures are not drawn to scale according to the actual shape and size of the implementation. For example, the dimensions of certain components have been enlarged for ease of explanation.
[0074] Furthermore, it should be understood that when a component such as a layer, film, region, or substrate is referred to as being "on" or "connected" to another component, it may be directly on or connected to the other component, or an intermediate component may also be present. Conversely, when a component is referred to as being "directly on" or "directly connected" to another component, no intermediate component exists. As used herein, "connection" can refer to physical and / or electrical connections. Moreover, "electrical connection" or "coupling" can refer to the presence of other components between the two components.
[0075] Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having the same meaning as they have in the context of the relevant technology and this invention, and will not be interpreted as having an idealized or overly formal meaning unless expressly defined herein.
[0076] Furthermore, it should be understood that although the terms “first,” “second,” “third,” etc., may be used herein to describe various components, parts, regions, layers, and / or portions, these components, parts, regions, and / or portions should not be limited by these terms. These terms are used only to distinguish one component, part, region, layer, or portion from another. Therefore, the “first component,” “part,” “region,” “layer,” or “part” discussed below may be referred to as a second component, part, region, layer, or portion without departing from the teachings of this document.
[0077] It should be understood that references to "an embodiment," "embodiment," "example embodiment," etc., in the specification refer to the described embodiment, indicating that it may include specific features, structures, or characteristics, but does not necessarily include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.
[0078] The specification and subsequent claims use certain terms to refer to specific modules, components, or parts. Those skilled in the art will understand that users or manufacturers may use different names or terms to refer to the same module, component, or part. This specification and subsequent claims do not distinguish modules, components, or parts by differences in name, but rather by differences in function. The terms "comprising" and "including" used throughout the specification and subsequent claims are open-ended and should be interpreted as "including but not limited to." Furthermore, the term "connection" here includes any direct and indirect electrical connection means. Indirect electrical connection means include connections via other means.
[0079] Furthermore, in the following description and claims, numerous terms will be referenced, which should be defined as having the following meanings. The singular forms “a” and “the” include plural referents, unless the context clearly specifies otherwise. “Optional” or “optionally” indicates that an event or situation subsequently described may or may not occur, and the description includes both the scenario where the event occurs and the scenario where the event does not occur.
[0080] To address the issue of poor generalization of traditional methods when using transfer learning for supercapacitor capacity assessment due to the diverse and varied application scenarios, please refer to [link to relevant documentation]. Figure 1 and Figure 2 , Figure 1 A flowchart illustrating a supercapacitor capacity estimation method in one embodiment of the present invention is shown. Figure 2 A diagram illustrating the overall architecture of a supercapacitor capacity estimation method in one embodiment of the present invention is provided.
[0081] This invention proposes a method for estimating the capacity of a supercapacitor, comprising the following steps:
[0082] Step S1: Collect source domain data 3 of multiple supercapacitors 1 under multiple operating conditions, and extract source domain multivariate features 4 containing strong correlation features of capacity decay based on source domain data 3;
[0083] Step S2: Collect target domain data and perform knowledge fusion through adversarial network based on the prior knowledge of supercapacitor 1 and source domain multi-feature 4 to obtain fusion features 5, pattern matching results 6 and multi-feature contribution spectrum 7.
[0084] Step S3: Based on the prior knowledge of supercapacitor 1, fusion feature 5 and pattern matching result 6, set the noise reduction parameters of the noise reduction model for multiple modes;
[0085] Step S4: Construct a neural network to learn the denoising characteristics of multiple modes of denoising models to obtain an adaptive denoising model;
[0086] Step S5: Construct the basic transfer learning capacity evaluation model 8, and train the basic transfer learning capacity evaluation model 8 using target domain data. Transfer the fusion feature 5, pattern matching result 6, multi-feature contribution spectrum 7, adversarial network and adaptive noise reduction model to the basic transfer learning capacity evaluation model 8 to obtain the first-stage capacity evaluation model.
[0087] Step S6: Based on the prior knowledge of supercapacitor 1, the first-stage capacity assessment model is fine-tuned in segments to obtain the full-cycle capacity assessment model.
[0088] Step S7: Integrate the full-cycle capacity assessment model into the supercapacitor capacity management system.
[0089] In one embodiment, collecting raw characteristic data of multiple supercapacitors 1 under multiple operating conditions further includes:
[0090] Step S11: Acquire voltage, current, temperature, humidity, appearance images, and component structure feature data of multiple supercapacitors 1. The working scenarios of supercapacitors 1 are diverse and complex, covering rail transit, power grids, buildings, etc., involving different practical application environments. For application scenarios under different complex working conditions, it is necessary to deploy multi-point sensors 2 of supercapacitors 1 and introduce acquisition sensors 2 of supercapacitors 1, thereby acquiring raw feature data such as voltage, current, temperature, humidity, appearance images, and component structure of supercapacitors 1 in real time based on cameras, temperature and humidity sensors, thin films, patches, etc., which can be directly obtained through sensors 2.
[0091] In one embodiment, extracting source domain multivariate features 4, which includes features strongly correlated with capacitance decay, based on source domain data 3 and prior knowledge of the supercapacitor 1, further includes:
[0092] Step S12: Perform basic data processing and basic anomaly detection on the raw feature data at the acquisition end;
[0093] Step S13: Based on the original feature dataset stored in the cloud, transform the original feature data into source domain multivariate features 4 containing features strongly correlated with capacity decay, based on the basic data processing results of the original feature data.
[0094] Based on the collected data, a dual-end system is constructed, consisting of an on-device terminal and a cloud-based system. On the device terminal, basic data means, variance, and simple outlier detection are performed. On the cloud terminal, a supercapacitor dataset and knowledge base are formed, driven by both prior knowledge (expert knowledge) and data-driven knowledge. This allows for the extraction of more potentially similar datasets using knowledge, and further refinement of the data's means and variance through more effective preprocessing operations. This extracts features strongly correlated with capacity decay, including charging time (CT), discharging time (DT), equivalent DC internal resistance (DCIR), and median discharge voltage (MDV). Prior knowledge may include theoretical ranges for specific capacitance, empirical formulas for porosity-specific surface area, electrolyte decomposition voltage windows, self-discharge rate, operating temperature windows, maximum continuous current, and maximum allowable temperature rise.
[0095] In one embodiment, the process of collecting target domain data and performing knowledge fusion through an adversarial network based on the prior knowledge of the supercapacitor 1 and the multi-dimensional features of the source domain 4 further includes:
[0096] Step S21: Enhance the source domain multivariate features 4 based on the physical characteristics of the supercapacitor 1 using an attention mechanism to obtain enhanced features;
[0097] Step S22: Construct an adversarial network for fused knowledge based on enhanced features, source domain multi-feature 4, and some target domain data to obtain fused features 5.
[0098] The features extracted in step S1 can more effectively reduce noise and reveal the key capacitance variation characteristics of supercapacitor 1. However, to more effectively improve knowledge transfer, it is necessary to further enhance the potential source domain and improve interpretability. Specifically, the physical characteristics of supercapacitor 1 include ion diffusion and adsorption, hierarchical stripping, polarization degree, pore structure, and other physical characteristics. These are integrated into the attention mechanism construction. The process includes: forming a feature vector based on the input source domain multivariate features 4 and the physical characteristics knowledge. ( Belongs to the d-dimensional real space R d (Subsequent similar markers will not be explained again) and task embedding spliced as Construct an enhanced input representation. Among them, the feature vector It is formed by combining the multivariate features of the input source domain with physical property knowledge through processes such as preprocessing, sampling, lightweight encoding, and concatenation; task embedding. It is a vector obtained through, for example, distillation using an attention mechanism. Then, this enhanced input representation... The process involves a linear attention layer (i.e., a shared linear transformation), followed by layer normalization and softmax transformation, and then temperature coefficient adjustment. To smooth the distribution of attention:
[0099]
[0100] in It is the weight matrix for learning. This is a temperature coefficient used to control the smoothness of attention distribution. Smaller... A higher value will make the attention distribution more focused, suppressing unimportant features, while a larger value will... This value leads to a more even distribution of attention. It is a layer normalization function. It is attention weight.
[0101] By attention weight With input feature vector Element-wise multiplication is performed to calculate the weighted eigenvector. Finally, the weighted feature vector is compared with the original input feature vector. The features are then concatenated again to obtain the final enhanced feature representation. :
[0102]
[0103] This enables feature enhancement based on attention mechanisms.
[0104] The steps for constructing an adversarial network based on enhanced features, source domain multi-feature 4, and a small amount of target domain data to obtain fused features 5 are as follows:
[0105] Construct an adversarial network (GAN) consisting of a generator G and a discriminator D engaging in an adversarial game, with the objective function being:
[0106]
[0107] Where x is the output, z is the noise, D is the discriminator, G is the generator, and V(D,G) represents the value function. E x∼pdata(x) This represents the expectation calculation for a sample x on the true data distribution pdata(x). z∼pz(z) This indicates that the expectation of a sample z on the input noise distribution pz(z) of the generator is calculated.
[0108] To improve its performance in different scenarios for supercapacitor 1 and to achieve more suitable results in scenarios with limited data, this invention employs an improved generator, constructs a frozen layer, and increases physical constraint loss.
[0109] The improved solution for the generator is as follows:
[0110]
[0111] in, For generator functions, This represents the solution of a partial differential equation, where F is the neural network parameterization function and N(s) is the random noise field. This indicates a parameterized definition. These are the initial and final times.
[0112] To fully utilize the knowledge learned from other operating conditions, some parameters can be frozen to find better parameters. :
[0113]
[0114] As for the final losses, in addition to the existing losses from the fight against the enemy... In addition, physical constraint loss was also added. The total loss is defined as follows:
[0115]
[0116] Enhance features After the source domain multi-feature 4 and some (small amount) target domain data are passed through the constructed adversarial network, the fused feature 5 is obtained.
[0117] In one embodiment, the process of collecting target domain data and performing knowledge fusion through an adversarial network based on the prior knowledge of the supercapacitor 1 and the multi-dimensional features of the source domain 4 further includes:
[0118] Step S23: To optimize feature selection for different working conditions, the probabilistic fusion learning matching method is used to perform feature matching on the fused feature 5, resulting in pattern matching result 6 and an interpretable multi-feature contribution spectrum 7.
[0119] Specifically, a probabilistic fusion learning matching method is used to perform feature matching analysis with the source domain selected by fusion feature 5, which has rich data information. The formula based on the Bayesian model is as follows:
[0120]
[0121] Where D is the given data source, and x is the selected probability characteristic, such as mean, variance, distribution, and other probability statistical properties. The weights selected based on knowledge are either weights stored in the cloud or weights stored on the device. These are candidate artificial intelligence models. This includes model frameworks such as CNN, RNN, and Transformer, upon which subsequent complete artificial intelligence networks are built, as detailed later. This indicates that in x and the k-th candidate model Under the given conditions, the probability distribution of the target variable y is the prediction of the target variable by a single model. This indicates the selection of the k-th candidate model given a data source D. The probability reflects the model M k The degree of matching with data source D. Pattern matching result 6 is determined by... The magnitude of the value reflects the selection of a value decision model and data source based on a higher probability.
[0122] By selecting features, that is, by fusing features obtained through expert knowledge and adversarial networks, an interpretable multi-feature contribution spectrum 7 is completed. The feature selection process is visualized using visualization tools such as tsfresh + plotly, AutoML + fviz, Vertex Explainable AI, etc. This invention is not limited to these tools. The influence of all features on the state of supercapacitor 1 is further analyzed and clarified, thereby further refining the features.
[0123] It should be noted that the mode of a supercapacitor can be understood as a set of one or more feature vectors under a specific operating condition, such as terrain features like slopes and straight routes, speed changes of the train on different sections, climate conditions, train load characteristics, and features such as voltage, current, temperature, and humidity of the supercapacitor.
[0124] Based on the above, the optimal fusion feature 5 is obtained, and pattern matching is performed with the most similar source domain dataset to obtain pattern matching result 6 (i.e., ).
[0125] Through iterative updates using the aforementioned intelligent methods, more reliable features are clarified, more similar datasets are found, different types of datasets are matched, and their parameter design rules are clarified to achieve pattern recognition.
[0126] In one embodiment, setting the denoising parameters of the denoising model for multiple modes based on prior knowledge of the supercapacitor 1, fusion feature 5, and mode matching result 6 further includes:
[0127] Step S31: Based on the prior knowledge of fusion feature 5 and supercapacitor 1, perform statistical analysis on the data of supercapacitor 1 operating under different conditions, and extract Gaussian filter parameters for multiple modes.
[0128] Although the aforementioned steps have already performed preliminary noise reduction in supercapacitor 1, various types of noise still exist during transmission and collection. In one embodiment, a Gaussian filtering model is used for noise reduction, but this invention is not limited thereto. Specifically, a large amount of source domain data 3 (including fused feature 5 and prior knowledge of supercapacitor 1) is acquired in the laboratory and in actual tests, and statistical analysis, hypothesis testing analysis, cluster analysis, and correlation analysis are performed on the large amount of data. Through statistical analysis, statistical regularities are clarified, making it easier to extract common patterns in the data. Based on hypothesis testing analysis, distribution patterns are clarified, making it easier to extract data phenomena and support data characteristic-related decisions. Based on cluster analysis, data classification is clarified, making it easier to discover the internal common characteristics and structure of the data. Based on correlation analysis, correlation patterns are clarified, making it easier to eliminate the influence of irrelevant data based on cluster analysis.
[0129] Based on the data and prior knowledge of supercapacitor 1, different Gaussian filter parameters are analyzed according to the extracted different modes.
[0130]
[0131] in, It is the standard deviation of the m-th mode. It is the distance between the extracted point and the center point in the m-th mode. These are the optimal parameters for matching after pattern recognition. `mode()` represents a pattern matching analysis algorithm that uses empirical knowledge and statistical characteristics to obtain the pattern membership relationship for a new working condition. This invention does not limit the specific pattern matching analysis algorithm used.
[0132] The Gaussian filtering formula of this invention is as follows:
[0133]
[0134] in, It is the value of the Gaussian function in the m-th mode.
[0135] In one embodiment, the denoising characteristics of the denoising model that constructs a neural network to learn multiple modes further include:
[0136] The neural network is trained to learn the denoising characteristics of the denoising models of multiple modes by inputting source domain multivariate features 4 and denoising source domain multivariate features 4 after denoising through multiple modes.
[0137] Because Gaussian filtering suffers from high computational cost, difficulty in capturing key features, and poor generalization, a convolutional neural network (CNN) is introduced. By inputting a large amount of undenoised data under different modes and denoised data after Gaussian filtering into the CNN, the self-learning capability of the CNN learns the denoising characteristics of Gaussian filtering, achieving intelligent updates for different modes, thereby constructing a unified intelligent adaptive denoising model. This invention does not limit the architecture or related parameters of the CNN. Based on the model, multi-physical characteristic analysis is achieved in the target domain by analyzing fused physical knowledge such as electrochemical characteristics, electrode characteristics, and kinetic characteristics. Contribution spectrum analysis is performed through empirical selection of fusion and statistical regularities to obtain the most similar features and patterns of fused physical knowledge. This utilizes rich source domain data under different operating conditions to achieve an adaptive denoising model in the target domain under specific modes, reducing noise, improving the similarity of potential source domain data under different target domain conditions, and enhancing the potential transferability between different domains.
[0138] In one embodiment, constructing a basic transfer learning capacity evaluation model 8 and training the basic transfer learning capacity evaluation model 8 using target domain data, and transferring the fused feature 5, pattern matching result 6, multi-feature contribution spectrum 7, adversarial network, and adaptive denoising model to the basic transfer learning capacity evaluation model 8 further includes:
[0139] A domain adaptation transfer module 81 is based on a convolutional neural network (CNN) or a Transformer neural network. The domain adaptation transfer module 81 includes:
[0140]
[0141] in, Represents the input sample of the source domain and , The number of samples in the source domain is represented by g, and the dimension is represented by g. Represents the input samples of the target domain and , Indicates the number of samples in the target domain. and Represents the loss function. This represents the feature extraction representation. Describes an auxiliary function. Represents a mapping function. Indicates the edge parameter.
[0142] Based on the aforementioned MDD formula, the domain adaptation transfer learning module 81 will help select more suitable data samples, and the fusion feature 5 will further refine more suitable input samples from the data samples to improve feature extraction. In addition, the aforementioned adaptive denoising model will also reduce the impact of adverse noise, thereby avoiding incorrect calculation of the loss function in the MDD formula. Based on the aforementioned feature extraction, pattern matching, and denoising model, the domain adaptation transfer learning module constructed by the MDD formula can effectively reduce the boundary gap between different domains, thereby aligning the distribution of the source domain and the target domain to complete the knowledge transfer from the source domain to the target domain, thus achieving early capacity assessment.
[0143] In one embodiment, the segmented fine-tuning of the first-stage capacity assessment model based on prior knowledge of the supercapacitor 1 further includes:
[0144] Based on the characteristics of the entire life cycle of supercapacitor 1 in the source domain data, the target domain data of supercapacitor 1 is divided into multiple stages; for example, the target domain data can be segmented according to the supercapacitor capacity retention rate, equivalent series resistance, leakage current, etc. in the source domain data, but the present invention is not limited thereto.
[0145] Set thresholds for multiple stages;
[0146] Fine-tuning of some network layers in the first-stage capacity assessment model is performed based on thresholds from multiple stages.
[0147] Specifically, knowledge such as rated cycle life, operating environment, and process characteristics is introduced to set different segments for supercapacitor 1, that is, thresholds are set through artificial intelligence methods. The specific process is as follows:
[0148]
[0149] in, Indicates the rated cycle life. Indicates the work environment. Indicates process characteristics, Let represent the threshold of the i-th stage, and ARI represent the mapping relationship of the neural network.
[0150] Based on the above formula, different thresholds can be constructed. In the basic module obtained in the first stage, there are layers such as the feature extraction layer in the frozen part. When the health state SOH is less than... At the same time, the network layers of key parts are fine-tuned, and key information of knowledge fusion is captured through mechanisms such as attention. Then, new target domain data acquired online can be used to effectively and reliably extract data information through fine-tuning, thereby achieving more reliable capacity assessment. This enables the transfer learning model for capacity estimation to be optimized and updated in stages at each stage throughout the entire life cycle, thus achieving accurate supercapacitor capacity status assessment.
[0151] The present invention also proposes a supercapacitor capacity estimation device for implementing any of the aforementioned methods, comprising:
[0152] The feature extraction and correlation analysis module 11 is used to collect source domain data 3 of multiple supercapacitors under multiple operating conditions, and extract source domain multivariate features 4 containing strong correlation features of capacity decay based on the source domain data 3.
[0153] The knowledge data fusion pattern matching module 12 is used to perform knowledge fusion through an adversarial network based on the target domain data, the prior knowledge of the supercapacitor, and the multi-feature features of the source domain 4, to obtain the fusion features 5, the pattern matching result 6, and the multi-feature contribution spectrum 7.
[0154] The adaptive noise reduction module 13 is used to set the noise reduction parameters of multiple noise reduction models based on the prior knowledge of the supercapacitor, the fusion feature 5 and the pattern matching result 6; and to construct a neural network to learn the noise reduction characteristics of multiple noise reduction models to obtain an adaptive noise reduction model.
[0155] The two-stage transfer learning module 14 is used to construct the basic transfer learning capacity evaluation model 8 and train the basic transfer learning capacity evaluation model 8 using target domain data. The fusion feature 5, pattern matching result 6, multi-feature contribution spectrum 7, adversarial network and adaptive noise reduction model are transferred to the basic transfer learning capacity evaluation model 8 to obtain the first-stage capacity evaluation model. Furthermore, the first-stage capacity evaluation model is fine-tuned in segments according to the prior knowledge of supercapacitors to obtain the full-cycle capacity evaluation model.
[0156] Integration module 15 is used to integrate the full-cycle capacity assessment model into the supercapacitor capacity management system.
[0157] The present invention also proposes a supercapacitor management system 100, comprising:
[0158] Supercapacitor 1,
[0159] Controller 20,
[0160] Multiple sensors 2, connected to the supercapacitor and / or controller, are used to acquire source domain data 3 of the supercapacitor under multiple operating conditions, wherein,
[0161] The controller contains steps for implementing any of the aforementioned methods.
[0162] The present invention also proposes a storage medium for storing a computer control program, which is used to execute the steps of any of the foregoing methods.
[0163] Compared with existing technologies, the method disclosed in this invention (1) extracts features strongly correlated with capacity decay by using prior knowledge and data-driven approaches for different operating conditions in the data acquisition, preprocessing, and multi-physical feature extraction and correlation analysis modules; (2) achieves more effective matching with corresponding patterns in the target domain through reasonable pattern matching; (3) realizes intelligent noise reduction for different operating conditions by learning the noise reduction characteristics of multiple pattern noise reduction models through neural networks, thereby reducing the negative impact of adverse information on knowledge transfer; and (4) obtains a full-cycle capacity assessment model by segmented fine-tuning of the capacity assessment model. These four aspects improve the adaptability and generalization of the capacity assessment model to different operating conditions.
[0164] The above-disclosed content is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of the patent application of the present invention. Therefore, any equivalent technical changes made based on the contents of the present invention specification and drawings fall within the scope of the patent application of the present invention.
Claims
1. A method for estimating the capacity of a supercapacitor, characterized in that, Including the following steps: Source domain data of multiple supercapacitors under multiple operating conditions are collected, and multi-source features of the source domain, including those with strong correlation to capacitance decay, are extracted based on the source domain data and prior knowledge of the supercapacitors. Data from the target domain is collected and knowledge is fused through an adversarial network based on the prior knowledge of the supercapacitor and the multi-feature features of the source domain to obtain fused features, pattern matching results, and multi-feature contribution spectrum. In order to optimize feature selection for different working conditions, a probabilistic fusion learning matching method is used to perform feature matching on the fused features to obtain the pattern matching results and the interpretable multi-feature contribution spectrum. The noise reduction parameters of multiple noise reduction models are set based on the prior knowledge of the supercapacitor, the fusion features, and the pattern matching results. A neural network is constructed to learn the denoising characteristics of the denoising models of the multiple modes, thereby obtaining an adaptive denoising model; A basic transfer learning capacity assessment model is constructed, and the basic transfer learning capacity assessment model is trained using target domain data. The fusion features, pattern matching results, multi-feature contribution spectrum, adversarial network and adaptive noise reduction model are transferred to the basic transfer learning capacity assessment model to obtain the first-stage capacity assessment model. Based on the prior knowledge of the supercapacitor, the first-stage capacity assessment model is finely adjusted in segments to obtain the full-cycle capacity assessment model. The full-cycle capacity assessment model is integrated into the supercapacitor capacity management system; wherein... Constructing a basic transfer learning capacity evaluation model and training the basic transfer learning capacity evaluation model using target domain data, further comprising transferring the fusion features, pattern matching results, multi-feature contribution spectra, adversarial networks, and adaptive denoising models to the basic transfer learning capacity evaluation model, includes: A domain adaptation transfer module is based on a fusion of convolutional neural networks (CNN) or transformer neural networks. The domain adaptation transfer module includes: Extracting input samples from data samples improves the quality of feature extraction. in, Represents the input sample of the source domain and , The number of samples in the source domain is represented by g, and the dimension is represented by g. Represents the input samples of the target domain and , Indicates the number of samples in the target domain. and Represents the loss function. This represents the feature extraction representation. Describes an auxiliary function. Represents a mapping function. Indicates edge parameters.
2. The method according to claim 1, characterized in that, The acquisition of source domain data from multiple supercapacitors under multiple operating conditions further includes: The voltage, current, temperature, humidity, appearance images, and component structure characteristics data of the multiple supercapacitors are acquired.
3. The method according to claim 1, characterized in that, Based on the source domain data and prior knowledge of supercapacitors, the extracted source domain multivariate features, which include strong correlation features with capacitance decay, further include: Data processing and anomaly detection are performed on the source domain data at the acquisition end; Based on the original feature dataset stored in the cloud, the data processing results of the source domain data are transformed into source domain multivariate features containing features strongly correlated with capacity decay.
4. The method according to claim 3, characterized in that, The process of collecting target domain data and fusing it with prior knowledge of supercapacitors and the multi-dimensional features of the source domain through an adversarial network further includes: The source domain multivariate features are enhanced based on the physical characteristics of the supercapacitor using an attention mechanism to obtain enhanced features; Based on the enhanced features, the source domain multi-features, and part of the target domain data, an adversarial network for fused knowledge is constructed to obtain the fused features.
5. The method according to claim 1, characterized in that, The denoising parameters of the denoising model for multiple modes, based on the prior knowledge of the supercapacitor, the fusion features, and the mode matching results, further include: Based on the fusion characteristics and prior knowledge of the supercapacitor, statistical analysis is performed on the data of the supercapacitor operating under different conditions to extract Gaussian filter parameters for the multiple modes.
6. The method according to claim 5, characterized in that, The denoising characteristics of the denoising model constructed by learning the multiple modes through a neural network further include: The source domain multivariate features and the source domain multivariate features after denoising by the multiple modes are input into the neural network to train the neural network to learn the denoising characteristics of the multiple modes of denoising models.
7. The method according to claim 1, characterized in that, Further fine-tuning of the first-stage capacity assessment model based on prior knowledge of the supercapacitor includes: Based on the characteristics of the supercapacitor's entire life cycle in the source domain data, the target domain data of the supercapacitor is divided into multiple stages; Set thresholds for the multiple stages; The network layers of the first-stage capacity assessment model are fine-tuned based on the threshold values of the multiple stages.
8. A supercapacitor capacity estimation device for implementing the method as described in any one of claims 1 to 7, characterized in that, include: The feature extraction and correlation analysis module is used to collect source domain data of multiple supercapacitors under multiple operating conditions, and extract source domain multivariate features containing strong correlation features of capacity decay based on the source domain data. The knowledge data fusion pattern matching module is used to perform knowledge fusion through an adversarial network based on the target domain data, the prior knowledge of the supercapacitor, and the multi-feature features of the source domain to obtain fusion features, pattern matching results, and multi-feature contribution spectrum. In order to optimize feature selection for different working conditions, a probabilistic fusion learning matching method is used to perform feature matching on the fusion features to obtain the pattern matching results and the interpretable multi-feature contribution spectrum. An adaptive noise reduction module is used to set the noise reduction parameters of multiple noise reduction models based on the prior knowledge of the supercapacitor, the fusion features, and the pattern matching results; and to construct a neural network to learn the noise reduction characteristics of the multiple noise reduction models to obtain an adaptive noise reduction model. A two-stage transfer learning module is used to construct a basic transfer learning capacity evaluation model and train the basic transfer learning capacity evaluation model using target domain data. The fusion features, pattern matching results, multi-feature contribution spectrum, adversarial network, and adaptive noise reduction model are transferred to the basic transfer learning capacity evaluation model to obtain a first-stage capacity evaluation model. Furthermore, the first-stage capacity evaluation model is fine-tuned in segments based on the prior knowledge of the supercapacitor to obtain a full-cycle capacity evaluation model. An integration module is used to integrate the full-cycle capacity assessment model into the supercapacitor capacity management system.
9. A supercapacitor management system, characterized in that, include: Supercapacitors Controller Multiple sensors, connected to the supercapacitor and / or controller, are used to acquire source domain data of the supercapacitor under multiple operating conditions, wherein... The controller includes steps for implementing any one of the methods described in claims 1 to 7.
10. A storage medium for storing a computer control program, characterized in that, The computer control program is used to perform the steps of the method as described in any one of claims 1 to 7.