Fault detection and classification method of energy storage battery and storage medium
By adopting a two-stage fault detection framework that combines unsupervised anomaly detection and supervised classification models, the problems of high false alarm rate, high false negative rate and unknown fault type in existing battery fault detection technologies are solved, achieving low latency and high accuracy in battery fault detection and classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUNAN CLOUD STORAGE RECYCLING NEW ENERGY TECH CO LTD
- Filing Date
- 2025-12-11
- Publication Date
- 2026-05-08
AI Technical Summary
Existing battery fault early warning technologies suffer from high false alarm rates, high false alarm rates, inability to adapt to battery aging and drift, and inability to pinpoint fault types. Especially when fault samples are scarce, it is difficult to achieve high-precision, low-latency fault detection and classification.
A two-stage fault detection framework is adopted. First, the features of the normal operation data of the battery are extracted using an unsupervised anomaly detection model. The reconstruction threshold is determined by a multi-channel LSTM autoencoder and a support vector data description algorithm to perform unsupervised anomaly detection. Then, after anomaly is detected, a supervised classification model such as a deep residual shrinking network is used to classify the fault type.
It achieves battery fault prediction and classification with low false alarm rate and low latency, can accurately identify fault types even when the amount of fault sample data is insufficient, and the model is easy to implement in engineering.
Smart Images
Figure CN121997114A_ABST
Abstract
Description
Technical Field
[0001] This invention mainly relates to the field of battery fault early warning technology, specifically a fault detection and classification method and storage medium for energy storage batteries. Background Technology
[0002] Lithium-ion batteries, with their advantages of high energy density, long lifespan, and low self-discharge rate, have become the preferred energy source in various fields. However, in recent years, fires and explosions caused by faults such as thermal runaway and internal short circuits have occurred frequently, resulting in serious economic losses and social impacts. The evolution of battery faults is characterized by its insidious, nonlinear, and time-varying nature, making it difficult for traditional methods to achieve high-precision, low-latency early warning. Therefore, early prediction and real-time monitoring technologies for battery faults still face significant challenges.
[0003] Existing battery fault early warning technologies mostly employ single thresholds or empirical models, but these suffer from high false alarm and false negative rates and cannot adapt to battery aging and drift. Some practitioners have proposed supervised deep learning methods, but these rely on a large number of fault labels, while actual fault samples are scarce, making the models prone to overfitting. Other practitioners have proposed unsupervised anomaly detection methods, which can provide early warnings but cannot identify the specific fault type or pinpoint the problem, making targeted maintenance difficult for operations personnel. When anomaly detection and fault classification are directly trained end-to-end, normal samples dominate the loss function, causing abnormal features to be overwhelmed and resulting in low classification accuracy.
[0004] Therefore, there is an urgent need for a two-stage fault detection framework that can be implemented in engineering, which "first uses an unsupervised model to fully explore the boundaries of normal data, and then uses a supervised model to finely distinguish fault types". Summary of the Invention
[0005] The technical problem to be solved by this invention is: in view of the technical problems existing in the prior art, this invention provides a fault detection and classification method and storage medium for energy storage batteries that is simple in principle, easy to implement, and has a wide range of applications.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A fault detection and classification method for energy storage batteries is proposed. It utilizes normal operation data to perform unsupervised anomaly detection, obtains fault features of the acquired abnormal data, and classifies the features. After an anomaly is triggered, a supervised classification model is invoked to provide the probability of the fault type.
[0007] As a further improvement of the present invention: the steps include: Step S1: Collect data on battery cell / module voltage, current, temperature, and SoC during actual operation. Collect normal operation data (D) and data when a fault occurs, and perform fault classification and labeling to obtain a fault label dataset. ; Step S2: For normal operation data The window data is channel normalized and divided into training set, test set and validation set; the training set is used to train the MC-LSTM-AE autoencoder model, the validation set is used to adjust the model parameters, and the test set is used to evaluate the model performance. Step S3: Train the MC-LSTM-AE autoencoder model using the training set from the normal operating data, and adjust the model parameters through training on the validation set to obtain the trained autoencoder model and the reconstruction matrix with the same dimensions as the training set; Step S4: Train the SVDD model using the error between the training dataset and the obtained reconstruction matrix to obtain the minimum hypersphere radius. ; Step S5: Transfer the fault label dataset The dataset is divided into a training set, a validation set, and a test set. The DRSN classification model is trained to perform fault diagnosis and classification on abnormal data. Step S6: Deploy the trained autoencoder model and classification model to achieve two-stage online fault diagnosis. For the data to be detected, the anomaly diagnosis module first determines whether the data is abnormal. If it is abnormal, the data to be detected is input into the fault diagnosis classification model to determine which type of fault it belongs to.
[0008] As a further improvement of the present invention, the fault types include internal short circuit, electrolyte leakage, thermal runaway, and excessive battery aging.
[0009] As a further improvement of the present invention: the unsupervised anomaly detection adopts a multi-channel LSTM autoencoder model. The multi-channel LSTM autoencoder model is used to extract multi-dimensional time series features from the normal battery operation status data, enhance information extraction through the attention layer, learn to capture the core features of the normal battery operation data, and simultaneously use the support vector data description algorithm to calculate the reconstruction error threshold of the MC-LSTM-AE model. The judgment of abnormal data is realized based on the threshold.
[0010] As a further improvement of the present invention: the LSTM structure based on the multi-channel LSTM autoencoder model controls the updating and utilization of historical sequence data through input gates, output gates, and forget gates. It uses multi-channel LSTM as the encoder and combines it with an attention mechanism to realize the connection between different feature channels of long-term sequences.
[0011] As a further improvement of the present invention: the data input to the encoder in the multi-channel LSTM autoencoder model includes the battery cells within the time window. The characteristic data, including voltage, current, temperature, and SoC, are shown below:
[0012]
[0013]
[0014]
[0015] In the formula, , , , Representing the first The time range of individual battery cells Voltage, current, temperature, and SoC characteristic data between them; In the encoder, for each feature data, an intermediate feature representation is obtained through LSTM and a linear layer; this includes voltage data features. ,in and These are the weights and biases of the linear layer, respectively; the same operation is performed on the current, temperature, and SoC feature channel data to obtain... , , Then, the data from all feature channels are concatenated to obtain... This means that the input for the multi-head attention mechanism is obtained.
[0016] As a further improvement of the present invention: In the multi-head attention mechanism, for each attention head, the input... Transform into three different matrices: , , ,in, If the weights are trainable, then the attention is calculated as follows:
[0017] in, yes The dimension; Feature fusion is performed using a multi-head attention mechanism, expressed as follows:
[0018] in, , These are the trainable weights of the attention head.
[0019] As a further improvement of the present invention: during the training process of deep networks, a residual normalization layer is used, which is represented as:
[0020] in, It is the characteristic normalization function. It is the output of the attention layer; The output of the attention layer The battery characteristics will be further optimized through a linear residual normalization layer; these characteristics are calculated using the following formula:
[0021] Among them, This represents the output of the linear residual normalization layer. and These represent the training weights and biases of that layer, respectively. at last, Through a name The flattening function transforms the input features into a low-dimensional space, and then the encoder output is obtained through a linear layer. Its expression is:
[0022] in, and These represent the training weights and biases of the last linear layer, respectively.
[0023] As a further improvement of the present invention: the supervised classification model is used to obtain fault features of the acquired abnormal data using a deep residual shrinking network (DRSN) and to classify the features.
[0024] The present invention also provides a storage medium that can be read by a computer or processor, wherein the storage medium stores a computer program for executing any of the above methods.
[0025] Compared with the prior art, the advantages of the present invention are as follows: 1. The fault detection and classification method and storage medium for energy storage batteries of this invention are simple in principle, easy to implement, and widely applicable. It adopts a two-stage intelligent diagnostic framework to achieve "detection first, classification later." The first stage utilizes readily available normal operating data to complete unsupervised anomaly detection; the second stage, after an anomaly is triggered, calls a supervised classification model to provide the probability of the fault type. Through techniques such as "detection-classification" decoupling and threshold adaptation, it achieves low false alarms, low latency, and interpretable battery fault prediction and classification. This solves the problems of high false alarm rates and unknown fault types in existing technologies.
[0026] 2. The fault detection and classification method and storage medium for energy storage batteries of this invention employ a two-stage fault detection framework: first, an unsupervised model is used to fully explore the boundaries of normal data, and then a supervised model is used to finely distinguish fault types. This framework is practical and can be implemented in engineering. It solves the problem of sample imbalance caused by insufficient fault sample data. At the same time, it proposes an autoencoder model (MC-LSTMA-AE) based on multi-channel LSTM and attention mechanism. The autoencoder is trained using normal operating data, and the reconstruction threshold of the MC-LSTMA-AE model is calculated using the support vector data description algorithm, thereby realizing anomaly detection of online operating data. The soft threshold in the Deep Residual Shrinking Network (DRSN) is improved to reduce the loss of some effective information due to constant bias and solve the problem that the network may stop updating when all input features are below the threshold. Attached Figure Description
[0027] Figure 1 This is a schematic diagram illustrating the implementation principle of the present invention in a specific embodiment.
[0028] Figure 2 This is a schematic diagram of the unsupervised anomaly detection module in a specific embodiment of the present invention.
[0029] Figure 3 This is a schematic diagram of the RSBU structure in a specific embodiment of the present invention.
[0030] Figure 4 This is a schematic diagram comparing different threshold functions in a specific embodiment of the present invention.
[0031] Figure 5 This is a flowchart illustrating the present invention in a specific embodiment. Detailed Implementation
[0032] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0033] like Figure 1 and Figure 5 As shown, this invention discloses a fault detection and classification method for energy storage batteries, applicable to electrochemical energy storage batteries, the process of which includes: Unsupervised anomaly detection is performed using normal operation data. For the acquired abnormal data, the fault characteristics of the abnormal data are obtained, and the characteristics are classified. After an anomaly is triggered, a supervised classification model is invoked to provide the probability of the fault type.
[0034] As can be seen from the above, the above-described solution of the present invention constitutes a two-stage battery fault diagnosis architecture of "detection first, classification later", and the overall architecture is as follows: Figure 1 As shown.
[0035] In specific application examples, the normal operation data is generally easily obtainable normal operation data.
[0036] In the above method of the present invention, the fault detection classification architecture includes an unsupervised anomaly detection module and a supervised classification module; wherein: The unsupervised anomaly detection module adopts a multi-channel LSTM autoencoder model (MC-LSTM-AE). The multi-channel LSTM autoencoder model is used to extract multi-dimensional time series features from battery normal operation state data. The attention layer enhances information extraction and learns to capture the core features of battery normal operation data. At the same time, the support vector data description algorithm is used to calculate the reconstruction error threshold of the MC-LSTM-AE model. The judgment of abnormal data is realized based on the threshold. The supervised classification module uses a deep residual shrinking network (DRSN) to obtain the fault features of the acquired abnormal data and then classifies these features.
[0037] To address the lack of actual fault labeling data in the daily operation of energy storage power stations, this invention employs an unsupervised method to detect fault anomalies. Specifically, for operational data such as voltage, current, temperature, and system-on-chip (SoC) collected in actual energy storage power stations, this invention proposes a multi-channel LSTM autoencoder model (MC-LSTMA-AE). This is an autoencoder model based on multi-channel LSTM and an attention mechanism. It trains the autoencoder using normal operation data and simultaneously uses a support vector data description algorithm to calculate the reconstruction threshold of the MC-LSTMA-AE model, thereby achieving anomaly detection in online operational data. In specific application examples, such as... Figure 2 The diagram shown is an architecture diagram of the unsupervised anomaly detection module.
[0038] In the LSTM structure, the updating and utilization of historical sequence data are controlled by input gates, output gates, and forget gates. This invention uses a multi-channel LSTM as an encoder and combines it with an attention mechanism to realize the connection between different feature channels of long-term sequences.
[0039] The data input to the encoder includes the battery cells within the time window. The characteristic data, including voltage, current, temperature, SoC, etc., are shown below:
[0040]
[0041]
[0042]
[0043] In the formula, , , , Representing the first Individual battery cells within a time range Voltage, current, temperature, and SoC characteristic data between them.
[0044] In the encoder, for each feature data, an intermediate feature representation is obtained through LSTM and a linear layer; for example, for voltage data features, i.e. ,in and These are the weights and biases of the linear layer, respectively; the same operation is performed on other feature channel data such as current, temperature, and SoC to obtain the corresponding values. , , Then, the data from all feature channels are concatenated to obtain... This means that the input for the multi-head attention mechanism is obtained.
[0045] In multi-head attention mechanisms, the goal is to effectively capture long-range dependencies in the battery cell sequence. For each attention head, the input... Transform into three different matrices: , , ,in, If the weights are trainable, then the attention is calculated as follows:
[0046] in, yes The dimension of the feature fusion is determined by the need to independently calculate the attention values for different battery characteristics such as voltage, current, SOC, and temperature. Therefore, a multi-head attention mechanism is employed for feature fusion, which can be expressed as:
[0047] in, , These are the trainable weights of the attention head.
[0048] Subsequently, a residual normalization layer is used to alleviate the problems of gradient vanishing and gradient exploding during deep network training. This mechanism can be expressed as:
[0049] in, It is the characteristic normalization function. It is the output of the attention layer.
[0050] Subsequently, the output of the attention layer The battery characteristics will be further optimized through a linear residual normalization layer; these characteristics can be calculated using the following formula:
[0051] Among them, This represents the output of the linear residual normalization layer. and These represent the training weights and biases of that layer, respectively.
[0052] at last, Through a name The flattening function transforms the input features into a low-dimensional space, and then the encoder output is obtained through a linear layer. Its expression is:
[0053] in, and These represent the training weights and biases of the last linear layer, respectively.
[0054] The data reconstruction process of the decoder is a mirror image of that of the encoder. To avoid repetition, this invention defines each operational layer sequentially as: unwinding layer. Linear layer Residual normalization layer and multi-head attention layer Therefore, the data reconstruction process can be represented as:
[0055]
[0056] By training the network using normal battery data to optimize the trainable matrix in the above formula, the proposed attention mechanism autoencoder can effectively reduce the difference between the original normal battery features and the reconstructed features.
[0057] In a preferred embodiment, the present invention further employs a reconstruction error threshold calculation, including using a Support Vector Data Description (SVDD) algorithm to determine the reconstruction threshold. This involves mapping the data to a high-dimensional space using a kernel function, and training to find a hypersphere that is as small as possible, such that as many normal data samples as possible fall within the hypersphere. The objective is to minimize the following objective function:
[0058] In the formula, For training samples; Let be the radius of the hypersphere. The center of the hypersphere, As a relaxation factor, It is a nonlinear mapping, responsible for mapping samples to a high-dimensional nonlinear space. The error penalty coefficient is used to adjust the number of normal samples outside the hypersphere and the algorithm complexity.
[0059] The SVDD model is trained using the error between the training data and the reconstructed data, resulting in a trained SVDD model and the corresponding minimum hypersphere radius. As a threshold for reconstruction error.
[0060] The Deep Residual Shrinking Network (DRSN) classification network consists of an input layer, convolutional layers, multiple residual shrinking units (RSBUs), batch normalization (BN) layers, rectified linear units (ReLU), global average pooling (GAP), and fully connected layers. An improved method is proposed for the soft thresholding function in the RSBU layers of the DRSN network. The improved RSBU structure is shown below. Figure 3 As shown.
[0061] The convolutional layer is responsible for extracting features from the input data, which can be represented as:
[0062] Batch normalization (BN) layers calculate the mean of features in a batch. and variance The mechanism reduces the internal covariate bias, and then normalization is performed according to the following formula.
[0063]
[0064]
[0065]
[0066]
[0067] In the above formula, Indicates batch size. , , Let represent the input, normalized output, and output of the i-th observation, respectively; It is a small value that approaches 0. and These are the parameters used for scaling. A semilinear transformation is then performed using the ReLU activation function: By calculating the average value for each channel, the GAP layer minimizes the number of parameters, thereby reducing the risk of poor fitting.
[0068] Combination Figure 4 Orange represents the fitting curve of the traditional soft thresholding function, and blue represents the fitting curve of the improved soft thresholding function. In the traditional soft thresholding function, the data exhibits good continuity, but when… hour, and A constant bias exists, which causes the loss of some useful information, leading to a certain degree of signal distortion and reducing diagnostic accuracy. The improved soft threshold function introduces a parameter... , is the slope coefficient, which is a very small positive value, when When, the output feature is Through parameters Achieving a smoother transition effectively solves the problem that the network might stop updating when all input features are below a threshold. In this case, the output features exhibit a non-linear relationship with the input data, rather than a constant deviation, thus enriching the output feature attributes. The improved soft thresholding function is:
[0069] Combination Figure 5 In a specific application, taking the application of this invention to an energy storage power station as an example, such as... Figure 5 As shown, the detailed process of this invention includes: Step S1: Collect data such as voltage, current, temperature, and SoC of individual battery cells / modules during actual operation of the energy storage power station. Data on normal operation (D) and fault occurrences of energy storage power stations are collected and classified by professionals to obtain a fault label dataset. The types of faults include internal short circuit, electrolyte leakage, thermal runaway, and excessive battery aging. Step S2: For normal operation data Channel normalization is performed on the window data, and the data is divided into training set, test set and validation set. The training set is used to train the MC-LSTM-AE autoencoder model, the validation set is used to adjust the model parameters, and the test set is used to evaluate the model performance. Step S3: Train the MC-LSTM-AE autoencoder model using the training set from the normal operating data, and adjust the model parameters through training on the validation set to obtain the trained autoencoder model and the reconstruction matrix with the same dimensions as the training set.
[0070] Step S4: Train the SVDD model using the error between the training dataset and the obtained reconstruction matrix to obtain the minimum hypersphere radius. .
[0071] Step S5: Transfer the fault label dataset The dataset is divided into a training set, a validation set, and a test set. The DRSN classification model is trained to perform fault diagnosis and classification on abnormal data.
[0072] Step S6: Deploy the trained autoencoder model and classification model to achieve two-stage online fault diagnosis. For the data to be detected, the anomaly diagnosis module first determines whether the data is abnormal. If it is abnormal, the data to be detected is input into the fault diagnosis classification model to determine which type of fault it belongs to.
[0073] The present invention also provides a storage medium that can be read by a computer or processor, wherein the storage medium stores a computer program for executing any of the above methods.
[0074] Those skilled in the art will understand that the above embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0075] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A fault detection and classification method for energy storage batteries, characterized in that, Unsupervised anomaly detection is performed using normal operation data. For the acquired abnormal data, the fault characteristics of the abnormal data are obtained and the characteristics are classified. After an anomaly is triggered, a supervised classification model is called to give the probability of the fault type.
2. The fault detection and classification method for energy storage batteries according to claim 1, characterized in that, step include: Step S1: Collect data on battery cell / module voltage, current, temperature, and SoC during actual operation. Collect normal operation data (D) and data when a fault occurs, and perform fault classification and labeling to obtain a fault label dataset. ; Step S2: For normal operation data The window data is channel normalized and divided into training set, test set and validation set; the training set is used to train the MC-LSTM-AE autoencoder model, the validation set is used to adjust the model parameters, and the test set is used to evaluate the model performance. Step S3: Train the MC-LSTM-AE autoencoder model using the training set from the normal operating data, and adjust the model parameters through training on the validation set to obtain the trained autoencoder model and the reconstruction matrix with the same dimensions as the training set; Step S4: Train the SVDD model using the error between the training dataset and the obtained reconstruction matrix to obtain the minimum hypersphere radius. ; Step S5: Transfer the fault label dataset The dataset is divided into a training set, a validation set, and a test set. The DRSN classification model is trained to perform fault diagnosis and classification on abnormal data. Step S6: Deploy the trained autoencoder model and classification model to achieve two-stage online fault diagnosis. For the data to be detected, the anomaly diagnosis module first determines whether the data is abnormal. If it is abnormal, the data to be detected is input into the fault diagnosis classification model to determine which type of fault it belongs to.
3. The fault detection and classification method for energy storage batteries according to claim 2, characterized in that, Fault types include internal short circuit, electrolyte leakage, thermal runaway, and excessive battery aging.
4. The fault detection and classification method for energy storage batteries according to claim 1, characterized in that, The unsupervised anomaly detection employs a multi-channel LSTM autoencoder model. This model is used to extract multi-dimensional time series features from battery normal operating state data, enhance information extraction through an attention layer, learn to capture the core features of battery normal operating data, and simultaneously use a support vector data description algorithm to calculate the reconstruction error threshold of the MC-LSTM-AE model. The threshold is then used to determine the anomaly data.
5. The fault detection and classification method for energy storage batteries according to claim 4, characterized in that, The LSTM structure based on the multi-channel LSTM autoencoder model controls the updating and utilization of historical sequence data through input gates, output gates, and forget gates. It uses a multi-channel LSTM as the encoder and combines an attention mechanism to realize the connection between different feature channels of long-term sequences.
6. The fault detection and classification method for energy storage batteries according to claim 5, characterized in that, The data input to the encoder in the multi-channel LSTM autoencoder model includes the battery cells within the time window. The characteristic data, including voltage, current, temperature, and SoC, are shown below: In the formula, , , , Representing the first The time range of individual battery cells Voltage, current, temperature, and SoC characteristic data between them; In the encoder, for each feature data, an intermediate feature representation is obtained through LSTM and a linear layer; this includes voltage data features. ,in and These are the weights and biases of the linear layer, respectively. The same operation was performed on current, temperature, and SoC feature channel data to obtain... , , Then, the data from all feature channels are concatenated to obtain... This means that the input for the multi-head attention mechanism is obtained.
7. The fault detection and classification method for energy storage batteries according to claim 6, characterized in that, In multi-head attention mechanisms, for each attention head, the input... Transform into three different matrices: , , ,in, If the weights are trainable, then the attention is calculated as follows: in, yes The dimension; Feature fusion is performed using a multi-head attention mechanism, expressed as follows: in, , These are the trainable weights of the attention head.
8. The fault detection and classification method for energy storage batteries according to claim 7, characterized in that, In the training process of deep networks, a residual normalization layer is used, which is represented as: in, It is the characteristic normalization function. It is the output of the attention layer; The output of the attention layer The battery characteristics will be further optimized through a linear residual normalization layer; these characteristics are calculated using the following formula: Among them, This represents the output of the linear residual normalization layer. and These represent the training weights and biases of that layer, respectively. at last, Through a name The flattening function transforms the input features into a low-dimensional space, and then the encoder output is obtained through a linear layer. Its expression is: in, and These represent the training weights and biases of the last linear layer, respectively.
9. The fault detection and classification method for energy storage batteries according to claim 8, characterized in that, The supervised classification model is used to obtain fault features of the acquired abnormal data using a deep residual shrinking network (DRSN) and then classify the features.
10. A storage medium capable of being read by a computer or processor, characterized in that, The storage medium stores a computer program for executing any one of the methods of claims 1 to 9.