Unsupervised learning-based power battery anomaly detection system

Abstract

An unsupervised learning-based power battery anomaly detection system, comprising: a battery parameter association reasoning module, configured to design a dynamic learnable graph neural network algorithm, and explore a potential association of battery parameters; a time sequence feature dependency analysis module, configured to model long-term and short-term time sequence dependency relationships on the basis of a temporal learning network; a coupling temporal-spatial neural architecture module, configured to perform coupling representation on the temporal learning network and a spatial learning network to integrate spatial-temporal association features; and a battery anomaly inference and evaluation module, configured to realize battery state prediction and evaluation by designing an adaptive threshold selection strategy and an experimental stage evaluation mechanism. The present application determines whether a power battery is in a normal operating state by analyzing a change mode of the key parameters of the power battery in a production process, and promptly stops the production process of a battery in an abnormal state, thereby preventing occurrence of accidents, and reducing resource loss in the production process.

Description

A power battery anomaly detection system based on unsupervised learning Technical Field

[0001] The present application relates to the field of battery detection technology, and in particular to a power battery anomaly detection system based on unsupervised learning. Background Art

[0002] Promoting electric vehicles is an effective solution to promoting carbon neutrality and combating the climate crisis. As a key component in advancing electric vehicle development, rechargeable power battery management technology awaits further breakthroughs. Advanced battery management technology not only extends battery life and improves production efficiency, but also effectively reduces the occurrence of accidents. Power battery fault diagnosis and anomaly detection technology are crucial components of power battery management technology, serving battery R&D, production, and vehicle-side use. This allows for the timely detection of battery failures, halting production or use, effectively improving production efficiency and ensuring workshop safety. With the rapid development of computer technology, deep learning-based time series anomaly detection technology has emerged, outperforming other models in both accuracy and robustness. However, deep learning methods are susceptible to data sparsity, and achieving accurate anomaly detection in the absence of labels requires further exploration.

[0003] Chinese patent application publication number CN116224123A discloses a battery anomaly detection system comprising: a communication module configured to acquire a battery image, battery identification information, and identification information of the image acquisition device captured by an image acquisition device; an anomaly detection module configured to acquire a battery image feature dataset from a cloud database based on the battery identification information and identification information of the image acquisition device received from the communication module, and to generate a battery anomaly detection result based at least on the acquired battery image and battery image feature dataset; and a judgment module configured to determine whether the abnormal region is a duplicated abnormal region in response to the battery anomaly detection result indicating the presence of an abnormal region in the battery, and to transmit the battery anomaly detection result in response to determining that the abnormal region is not a duplicated abnormal region. However, this system fails to exploit the association mechanism between battery parameters, fails to effectively address the limitation that battery parameters cannot be directly associated semantically, and fails to adapt to the dynamic scene switching in the battery production environment.

[0004] Application Contents

[0005] In response to the shortcomings of the existing technology, this application proposes a power battery anomaly detection system based on unsupervised learning. By analyzing the change pattern of key parameters of the power battery during the production process, it is determined whether the power battery is in a normal operating state, and the correlation mechanism between battery parameters is explored, which effectively solves the limitation that battery parameters cannot be directly associated semantically, and promptly stops the production process of abnormal batteries to alleviate the occurrence of accidents and reduce resource loss in the production process.

[0006] To implement the above technical solution, the present application provides a power battery anomaly detection system based on unsupervised learning, which specifically includes:

[0007] A battery parameter association reasoning module uses a dynamic, learnable graph neural network to model multi-parameter association mechanisms. In scenarios where it is difficult to directly establish associations based on the semantics of battery parameters and thus determine the graph structure, a learnable graph structure algorithm is constructed to mine potential connections between multiple parameters. Furthermore, in scenarios where parameter association rules may change with battery process steps, a dynamic graph structure is constructed through a dynamic graph neural network to model parameter association mechanisms under different battery process steps.

[0008] A time series feature dependency analysis module is used to construct a long-term and short-term time series network to model the short-term dependency and long-term dependency features of battery parameters, respectively. A temporal convolutional network is used to capture short-term correlations of battery parameters, while a gated recurrent unit is used to capture long-term correlations of battery parameters.

[0009] A coupled spatiotemporal neural architecture module is used to integrate a spatial learning network as an embedded component into a temporal learning network system, wherein the spatial learning network is a dynamic learnable graph neural network in the battery parameter association inference module, and the temporal learning network is a long-term and short-term temporal network in the time series feature dependency analysis module. The fusion of temporal and spatial features is achieved by replacing the linear units of the gated recurrent units in the temporal learning network with graph operators in the spatial learning network.

[0010] The battery anomaly inference and evaluation module is based on the unsupervised learning paradigm. The difference between the sequence prediction result and the true label of the sequence is compared with the proposed threshold. If it is greater than the threshold, the point is judged to be abnormal; otherwise, it is judged to be normal.

[0011] Preferably, the dynamic learnable graph neural network in the battery parameter association reasoning module generates a graph structure in the form of creating a learnable battery parameter embedding dictionary, and the graph structure edges are obtained by calculating the similarity of different battery parameter embedding dictionaries. The battery parameter embedding dictionary is weighted initialized in a random manner and adaptively updated during the training process.

[0012] Preferably, the dynamic learnable graph neural network uses feature fusion to generate a dynamic graph structure. For different battery working steps, an embedding layer and a multi-layer perceptron network are used to generate customized features for the battery working steps. The customized features are fused with the learnable graph structure to construct a dynamic graph neural network with work step perception.

[0013] Preferably, the short-term time series network in the time series feature dependency analysis module uses a time convolutional network to capture the short-term correlation of battery parameters, which is formalized as follows:

[0014] in represents the short-term timing dependence of the c-th battery parameter, and represent weight parameters and bias parameters respectively.

[0015] Preferably, the long-term temporal network in the time series feature dependency analysis module uses a gated recurrent unit to capture the long-term correlation of battery parameters, which is formalized as follows:

[0016] z t =sigmoid(W z ·[h t-1 ,E t ]+b z )

[0017] r t =sigmoid(W r ·[h t-1 ,E t ]+b r )

[0018] Among them, z t Represents the update gate, which is used to control the hidden state output h at time t-1 t-1 And the hidden state at time t Enter E t How much flows into the hidden state h at the current time t t In, z t The smaller the value, the more historical event information is input into the network, which is more helpful for the model to learn long-term dependencies; t Represents the reset gate, which is used to control the hidden state output h at time t-1 t-1 What proportion flows into the hidden state at time t In the example, the smaller the value, the more historical information is forgotten, which is more helpful for learning short-term dependencies in time series. Sigmoid(·) and tanh(·) represent the sigmoid and tanh activation functions, respectively. The sigmoid value range is (0, 1), which performs proportional mapping for the update gate and reset gate, and tanh is used to increase the nonlinear expression ability of the model.

[0019] Preferably, in the coupled spatiotemporal neural architecture module, for the short-term time series network, the factor neural architecture model is used to fuse the dynamic learnable graph neural network and the time convolution network. The factor neural architecture model continuously stacks the time learning network and the space learning network, and finally integrates the short-term and long-term spatiotemporal features by constructing a long-term and short-term spatiotemporal fusion layer to model complex spatiotemporal dependency associations. The long-term and short-term spatiotemporal fusion layer is implemented using a multi-layer perceptron network.

[0020] Preferably, the unsupervised learning paradigm in the battery abnormality inference and evaluation module is to predict the battery parameters of the T+1th time step based on the battery parameters of the previous T time steps, rather than directly predicting the battery operating status of the T+1th time step. Batteries in normal status are used as input in both the training and verification stages. The proposed threshold is determined according to the adaptive threshold selection strategy. For different battery parameters, the threshold is the point with the largest prediction error in the verification set. If the absolute value error between the predicted value and the true value of the cth battery parameter in the T+1th time step is greater than the threshold, it is determined that the battery c parameter status is abnormal, otherwise it is normal.

[0021] Preferably, in the battery status prediction stage, when only one battery parameter is predicted to be abnormal, the battery abnormality inference and evaluation module issues a primary warning, which is then subject to secondary confirmation by staff; if two or more battery parameters are predicted to be abnormal, an advanced warning is issued, and a production stop instruction is sent to the battery.

[0022] Preferably, the battery anomaly inference evaluation module also includes an experimental phase evaluation mechanism, which adaptively adjusts the anomaly prediction results according to the anomaly labels to further evaluate the anomaly algorithm that conforms to the actual battery production scenario.

[0023] The beneficial effects of the power battery anomaly detection system based on unsupervised learning provided by this application are:

[0024] (1) This application designs a dynamic learnable graph network structure for power battery anomaly detection scenarios. It is used to explore the association mechanism between battery parameters, effectively solving the limitation that battery parameters cannot be directly associated semantically. It also generates different graph structures for different battery production processes, making it more suitable for dynamic scene switching in battery production environments. At the same time, the temporal learning network and the spatial learning network are coupled to comprehensively analyze the spatiotemporal characteristics of battery parameters, thereby generating more accurate prediction results.

[0025] (2) This application designs an unsupervised neural network architecture for power battery anomaly detection scenarios, effectively solving the problem of supervised learning being unsuitable due to the lack of labels in battery production scenarios. The proposed battery anomaly inference and evaluation module can further screen anomaly detection algorithms that are more suitable for actual battery production scenarios, thereby improving algorithm development efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] FIG1 is a system module diagram of the present application. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0028] Example: A power battery anomaly detection system based on unsupervised learning

[0029] In particular, the specific task targeted by this embodiment is the abnormal state detection of batteries in the production process of power batteries, and experimental verification is mainly carried out in the formation stage and the capacity division stage. It should be pointed out that the use scenario of this application is not limited to the power battery production process, and any similar tasks (such as multivariate time series anomaly detection, etc.) can use this application as the basic framework. In the formation and capacity division stages, the power battery includes multiple battery steps, each of which represents a process, such as charging and discharging. When the process is switched, the battery parameters may change dramatically, and the trend of each parameter will also change due to the change of the battery step.

[0030] Specifically, in order to better illustrate the detection steps of this application, the embodiment provides a calculation process from raw data to prediction results. First, the raw data records the status / true value X of multiple parameters of the battery at different times during the production process. t,cWhere t represents the time index and c represents the key parameter type index (such as temperature, voltage, current). Furthermore, T is used to represent the total time step and C is the total number of key parameters. There are obvious differences in the index intervals of different types of parameters, so the mean-variance normalization is used to process the original data, which is formalized as follows:

[0031] Represents the normalized battery data, X t,c Represents the original battery data, μ c represents the key parameter c mean, σ c Represents the variance of battery parameter c. In order to capture the high-dimensional hidden correlation of battery data, the parameter matrix For standardized data Generate embedding representation E t,c , where d represents the hidden layer dimension. The process can be formalized as follows:

[0032] As shown in Figure 1, a power battery anomaly detection system based on unsupervised learning builds a deep neural network prediction framework based on unsupervised learning, including: battery parameter association reasoning module; time series feature dependency analysis module; coupled spatiotemporal neural architecture module and battery anomaly inference and evaluation module.

[0033] In particular, the battery parameter association inference module uses a dynamic learnable graph neural network to model a multi-parameter association mechanism. The dynamic learnable graph neural network is the result of the coupling of a learnable graph neural network and a dynamic graph neural network. In scenarios where it is difficult to directly establish associations based on the semantics of the battery parameters and thus impossible to determine the graph structure, the learnable graph neural network constructs a learnable graph structure algorithm to explore potential connections between multiple parameters. In scenarios where the parameter association rules may change with changes in battery processing steps, the dynamic graph neural network constructs a dynamic graph structure to model the parameter association mechanism under different battery processing steps.

[0034] The learnable graph neural network generates a graph structure by creating a learnable battery parameter embedding dictionary. The graph structure edges are obtained by calculating the similarity between different battery parameter embedding dictionaries. The battery parameter embedding dictionary is initialized with random weights and adaptively updated during training.

[0035] Specifically, the learnable battery parameter embedding dictionary is expressed as Each row represents the embedding of a parameter c, and d′ represents the dimension of the embedding of parameter c. The edges of the graph structure are constructed using similarity calculation, which is formalized as follows:

[0036] D -1 / 2 AD-1 / 2 =softmax(σ(N·N T )) (3)

[0037] in is the adjacency matrix of the learnable graph structure, D is the degree matrix of A, which is used to solve the scale inconsistency problem caused by the different number of node edges. softmax represents the softmax activation function, which is used to normalize the adjacency matrix. σ is the LeakyReLU activation function. During the training phase, the learnable battery parameters are embedded in the dictionary The embedded representation of battery parameters can be automatically learned based on the input data, which is achieved through back propagation.

[0038] Furthermore, the dynamic graph neural network structure uses feature fusion to generate a dynamic graph structure. For different battery process steps, an embedding layer and a multi-layer perceptron network are used to generate customized features for each battery process step. These customized features are then fused with the learnable graph structure to construct a process-step-aware dynamic graph neural network.

[0039] Specifically, we first generate embedding representations for different battery steps The process is similar to formula (2). In order to improve the representation ability of battery process step embedding, a multi-layer perceptron network is used to enhance the embedding representation. The formalization is as follows:

[0040] in, represents the embedded representation of the battery step at time step t after enhancement. MLP(·) represents the multi-layer perceptron network, represents the weight parameter, Represents the bias parameter. The dynamic learning graph neural network is generated by integrating the enhanced embedded representation of the battery process step and the learnable battery parameter embedding dictionary. The dynamic learning graph neural network embeds the learnable battery parameter into the dictionary. Updated to The dynamic learnable graph structure calculation process can be formalized as follows:

[0041] in, is the dynamic learnable graph structure adjacency matrix, which represents the battery parameter association relationship at the t-th time step, for Furthermore, the dynamic learning graph structure adjacency matrix can be expressed as The dynamic learnable graph neural network calculation process can be formalized as follows:

[0042] in represents the feature embedding encoded by the dynamic learnable graph neural network at time t, W g and b g Represent the weight parameters and bias parameters respectively. Considering each battery parameter as a node on the graph, Equation (6) transfers the neighbor node information to the node itself in the form of message passing. According to the dynamic learnable graph structure adjacency matrix created by Equation (5), the neighbor node information is automatically mined from the original data representation in a data-driven manner and changes with the evolution of time steps.

[0043] The time series feature dependency analysis module constructs a long-term and short-term time series network to model the short-term dependency and long-term dependency features of battery parameters respectively. The long-term and short-term time series network includes a short-term time series network and a long-term time series network. The short-term time series network uses a time convolutional network to capture the short-term correlation of battery parameters, which is formalized as follows:

[0044] in represents the short-term timing dependence of the c-th battery parameter, and denote the weight parameters and bias parameters respectively. In addition, formula (7) adopts residual connection to prevent the gradient vanishing problem of neural network.

[0045] Furthermore, the long-term temporal network uses gated recurrent units to capture the long-term association of battery parameters, which is formalized as follows:

[0046] z t =sigmoid(W z ·[h t-1 ,E t ]+b z ) (8)

[0047] r t =sigmoid(W r ·[h t-1 ,E t ]+b r ) (9)

[0048] Among them, z t Represents the update gate, which is used to control the hidden state output h at time t-1 t-1 And the hidden state at time t Enter E t How much flows into the hidden state h at the current time tt In, z t The smaller the value, the more historical event information is input into the network, which is more helpful for the model to learn long-term dependencies. t Represents the reset gate, which is used to control the hidden state output h at time t-1 t-1 What proportion flows into the hidden state at time t In [ ], smaller values ​​indicate more historical information is forgotten, which is more helpful for learning short-term dependencies in time series. Sigmoid(·) and tanh(·) represent the sigmoid and tanh activation functions, respectively. Sigmoid has a value range of (0, 1) and provides proportional mapping for the update and reset gates, while tanh is used to increase the model's nonlinear expressiveness. Combining the update and reset gate mechanisms allows for adaptive acquisition of long- and short-term historical information, further facilitating the model's outlier detection.

[0049] The coupled spatiotemporal neural architecture module integrates the spatial learning network as an embedded component into the temporal learning network system. The spatial learning network is the dynamic learnable graph neural network, and the temporal learning network is the long-term and short-term time series network. The fusion of temporal and spatial features is achieved by replacing the linear units of the gated recurrent units in the temporal learning network with graph operators in the spatial learning network.

[0050] Specifically, the coupled spatiotemporal neural architecture module updates equations (8) to (10) to equations (12) to (14):

[0051] By integrating a dynamic learnable graph structure into a long-term temporal learning network, the coupling of temporal and spatial feature correlations is achieved, thereby enhancing the model's representational capabilities. Specifically, for short-term temporal networks, a factored neural architecture is employed to fuse the dynamic learnable graph neural network with a temporal convolutional network. This factored neural architecture sequentially stacks the temporal learning network and the spatial learning network. Finally, a long-term and short-term spatiotemporal fusion layer is constructed to integrate short-term and long-term spatiotemporal features to model complex spatiotemporal dependencies. This long-term and short-term spatiotemporal fusion layer is implemented using a multilayer perceptron network.

[0052] The battery anomaly inference and evaluation module is based on an unsupervised learning paradigm and compares the difference between the sequence prediction result and the true label of the sequence with a proposed threshold. If it is greater than the threshold, the point is judged to be abnormal; otherwise, it is judged to be normal.

[0053] The unsupervised learning paradigm predicts the battery parameters at the T+1th time step based on the battery parameters at the previous T time steps, rather than directly predicting the battery's operating status at the T+1th time step. Specifically, normal batteries are used as input during both the training and validation phases. The proposed threshold is determined based on an adaptive threshold selection strategy. For different battery parameters, the threshold is the point in the validation set with the largest prediction error. That is, if the absolute difference between the predicted value and the true value of the cth battery parameter at the T+1th time step is greater than the threshold, the battery's c parameter status is considered abnormal; otherwise, it is considered normal.

[0054] In particular, during the battery status prediction stage, if only one battery parameter is predicted to be abnormal, a primary warning will be issued and the staff will conduct a secondary confirmation; if two or more battery parameters are predicted to be abnormal, an advanced warning will be issued and a stop production instruction will be sent to the battery.

[0055] Furthermore, the battery anomaly inference and evaluation module also includes an experimental phase evaluation mechanism that adaptively adjusts anomaly prediction results based on anomaly labels to further evaluate anomaly algorithms that conform to real battery production scenarios.

[0056] Specifically, in actual battery production applications, point-level metrics are unnecessary. If the algorithm can detect an anomaly within t1 of its occurrence, the algorithm is considered to meet the requirements. That is, if an anomaly is detected within an acceptable latency range (within t1), the abnormal segment (or battery) is considered correctly detected. This evaluation mechanism during the experimental phase can effectively help experimenters screen algorithms that meet actual application scenarios and improve algorithm development efficiency.

[0057] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications are also considered to be within the scope of protection of the present application.

Claims

1. A power battery anomaly detection system based on unsupervised learning, characterized in that: include: A battery parameter association reasoning module, which uses a dynamic learnable graph neural network to model a multi-parameter association mechanism. In a scenario where it is difficult to directly establish associations in the semantics of battery parameters and thus it is impossible to determine the graph structure, a learnable graph structure algorithm is constructed to mine potential connections between multiple parameters. In a scenario where the parameter association rules may change with the change of the battery process steps, a dynamic graph structure is constructed through a dynamic graph neural network to model the parameter association mechanism under different battery process steps; A time series feature dependency analysis module, which is used to construct a long-term and short-term time series network, model the short-term dependency and long-term dependency features of battery parameters respectively, use a time convolutional network to capture the short-term correlation of battery parameters, and use a gated recurrent unit to capture the long-term correlation of battery parameters; A coupled spatiotemporal neural architecture module, wherein the coupled spatiotemporal neural architecture module is used to integrate a spatial learning network as an embedded component into a temporal learning network system, wherein the spatial learning network is a dynamically learnable graph neural network in a battery parameter association reasoning module, and the temporal learning network is a long-term and short-term temporal network in a time series feature dependency analysis module, and the fusion of temporal and spatial features is achieved by replacing the linear unit of the gated recurrent unit in the temporal learning network with the graph operator in the spatial learning network; A battery anomaly inference and evaluation module, which is based on an unsupervised learning paradigm and compares the difference between the sequence prediction result and the sequence true label with a proposed threshold. If the difference is greater than the threshold, the point is determined to be abnormal; On the contrary, it is judged as normal.

2. The power battery anomaly detection system based on unsupervised learning according to claim 1, characterized in that: The dynamic learnable graph neural network in the battery parameter association reasoning module generates a graph structure in the form of creating a learnable battery parameter embedding dictionary, and the graph structure edges are obtained by calculating the similarity of different battery parameter embedding dictionaries. The battery parameter embedding dictionary is weighted initialized in a random manner and adaptively updated during the training process.

3. The power battery anomaly detection system based on unsupervised learning as claimed in claim 2, characterized in that: The dynamic learnable graph neural network uses feature fusion to generate a dynamic graph structure. For different battery process steps, an embedding layer and a multi-layer perceptron network are used to generate customized features for the battery process steps. The customized features are fused with the learnable graph structure to construct a dynamic graph neural network with process step perception.

4. The power battery anomaly detection system based on unsupervised learning according to claim 1, characterized in that: The short-term time series network in the time series feature dependency analysis module uses a time convolutional network to capture the short-term correlation of battery parameters, which is formalized as follows: in represents the short-term timing dependence of the c-th battery parameter, and Represent the weight parameters and bias parameters.

5. The power battery anomaly detection system based on unsupervised learning according to claim 1, characterized in that: The long-term time series network in the time series feature dependency analysis module uses gated recurrent units to capture the long-term correlation of battery parameters, which is formalized as follows: t =sigmoid(W z ·[h t-1 ,E t ]+b z ) t =sigmoid(W r ·[h t-1 ,E t ]+b r ) Among them, z t represents the update gate, which is used to control the hidden state output h at time t-1 t-1 And the hidden state at time t Enter E t How much flows into the hidden state h at the current time t t In, z t The smaller the value, the more historical event information is input into the network, which is more helpful for the model to learn long-term dependencies; t Represents the reset gate, which is used to control the hidden state output h at time t-1 t-1 What proportion flows into the hidden state at time t In , the smaller the value is, the more historical information is forgotten, which is more helpful for learning short-term dependencies in time series; sigmoid(·) and tanh(·) represent sigmoid and tanh activation functions respectively. The value range of sigmoid is (0,1), which is used to perform proportional mapping for update gate and reset gate, and tanh is used to increase the nonlinear expression ability of the model.

6. The power battery anomaly detection system based on unsupervised learning according to claim 1, characterized in that: In the coupled spatiotemporal neural architecture module, for the short-term time series network, the factor neural architecture model is used to fuse the dynamic learnable graph neural network and the temporal convolutional network. The factor neural architecture model continuously stacks the temporal learning network and the spatial learning network, and finally integrates the short-term and long-term spatiotemporal features by constructing a long-term and short-term spatiotemporal fusion layer to model complex spatiotemporal dependency associations. The long-term and short-term spatiotemporal fusion layer is implemented using a multi-layer perceptron network.

7. The power battery anomaly detection system based on unsupervised learning as claimed in claim 1, characterized in that: The unsupervised learning paradigm in the battery abnormality inference evaluation module predicts the battery parameters of the T+1th time step based on the battery parameters of the previous T time steps, rather than directly predicting the battery operating status of the T+1th time step. Batteries in normal status are used as input in both the training stage and the verification stage. The proposed threshold is determined according to the adaptive threshold selection strategy. For different battery parameters, the threshold is the point with the largest prediction error in the verification set. If the absolute value error between the predicted value and the true value of the cth battery parameter in the T+1th time step is greater than the threshold, it is determined that the c parameter status of the battery is abnormal, otherwise it is normal.

8. The power battery anomaly detection system based on unsupervised learning according to claim 7, characterized in that: During the battery status prediction stage, if only one battery parameter is predicted to be abnormal, the battery abnormality inference and evaluation module will issue a primary warning, which will be confirmed by the staff for a second time; if two or more battery parameters are predicted to be abnormal, an advanced warning will be issued and a stop production instruction will be sent to the battery.

9. The power battery anomaly detection system based on unsupervised learning according to claim 1, characterized in that: The battery anomaly inference evaluation module also includes an experimental phase evaluation mechanism, which adaptively adjusts the anomaly prediction results according to the anomaly labels to further evaluate the anomaly algorithm that conforms to the actual battery production scenario.

Images