Battery cell state determination device and operation method thereof

Abstract

A battery cell state determination device according to one embodiment of the present invention includes a first module that generates first state information about a battery cell based on a time series data set about the battery cell, and a second module that generates second state information about the battery cell based on the time series data set, and the first module is capable of supervised learning based on the second state information.

Description

[Technical Field]

[0001] The present invention claims the benefit of priority based on Korean Patent Application No. 10-2022-0126528, filed on October 4, 2022, and all contents disclosed in the documents of this Korean patent application are incorporated herein by reference.

[0002] SUMMARY OF THE INVENTION The embodiments disclosed herein relate to a battery cell state determination device and its method of operation. [Background technology]

[0003] In recent years, research and development into secondary batteries has been actively conducted. Secondary batteries are batteries that can be charged and discharged, and include conventional Ni / Cd batteries, Ni / MH batteries, and more recent lithium-ion batteries. Lithium-ion batteries have the advantage of having a much higher energy density than conventional Ni / Cd batteries, Ni / MH batteries, etc. Furthermore, because lithium-ion batteries can be manufactured to be small and lightweight, they are used as power sources for mobile devices. In recent years, their range of use has expanded to include power sources for electric vehicles, and they are attracting attention as a next-generation energy storage medium.

[0004] Batteries deteriorate as they are repeatedly charged and discharged. For example, the more a battery is charged and discharged, the more its capacity and resistance deteriorate, and its remaining lifespan may decrease. Furthermore, the degree of deterioration and remaining lifespan of a battery may change depending on the conditions of use.

[0005] If the remaining life of a battery suddenly decreases, safety issues may arise during use of the battery. Therefore, a method for determining the state of a battery and preventing dangers caused by a sudden decrease in battery performance is required. Summary of the Invention [Problem to be solved by the invention]

[0006] It is an object of the embodiments disclosed herein to provide an apparatus and method of operation thereof that includes a first module for determining a state of a battery cell based on a time series data set of the battery cell.

[0007] One objective of the embodiments disclosed herein is to provide an apparatus and method of operation that learns based on an experimental data set collected under pre-defined experimental conditions and determines the state of a battery based on a time series data set.

[0008] The technical problems of the embodiments disclosed in this document are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description. [Means for solving the problem]

[0009] A battery cell state determination device according to one embodiment of the present invention includes a first module that generates first state information about a battery cell based on a time series data set about the battery cell, and a second module that generates second state information about the battery cell based on the time series data set, and the first module is capable of supervised learning based on the second state information.

[0010] According to one embodiment, the time series data set includes a plurality of time series tokens, and the time series tokens may be continuous data having a temporal nature.

[0011] According to one embodiment, the first module may include: an encoder block that extracts context information of the time series dataset from a first time series token included in the time series dataset; a decoder block that generates predicted data based on the context information of the time series dataset and a second time series token included in the time series dataset; and a decision block that generates the first state information based on the predicted data and a third time series token included in the time series dataset.

[0012] In one embodiment, the first module is capable of correcting the forecast data based on the second state information. In one embodiment, the third time-series token may be data collected from the battery cell at the time of determining the state of the battery cell. According to one embodiment, the expected data may be data that is expected to be collected from the battery cell when the battery cell is in a normal state.

[0013] According to one embodiment, the second module may include a long short-term memory block that extracts first features based on the time series data set, a convolution block that extracts second features based on the time series data set, and a combination block that combines the first features and the second features to generate the second state information.

[0014] According to one embodiment, the long short-term memory block may include a cyclic layer that cyclically processes the time series data set, and a storage layer that stores the processing results of the cyclic layer.

[0015] According to one embodiment, the convolution block may include a convolution layer that performs a convolution operation on the time series data set. According to an embodiment, the convolution block may further include a compression layer that compresses a calculation result of the convolution layer, and an activation layer that corrects the calculation result of the convolution layer based on the calculation result of the compression layer.

[0016] According to one embodiment, the second module learns based on an experimental dataset, and the experimental dataset may be a dataset in which data of an experimental battery cell collected under preset experimental conditions is matched with state information of the experimental battery cell.

[0017] According to another embodiment of the present invention, a method for operating a battery cell state determination device may include the steps of collecting a time series data set related to a battery cell, generating first state information related to the battery cell based on the time series data set, generating second state information related to the battery cell based on the time series data set, and performing supervised learning based on the second state information.

[0018] In another embodiment, generating the first state information may include extracting context information of the time series dataset from a first time series token included in the time series dataset; generating forecast data based on the context information of the time series dataset and a second time series token included in the time series dataset; and generating the first state information based on the forecast data and a third time series token included in the time series dataset.

[0019] In another embodiment, generating the second state information may include extracting first features based on the time series data set, extracting second features based on the time series data set, and combining the first features and the second features to generate the second state information.

[0020] In another embodiment, the step of extracting the first feature may include the steps of: cyclically processing the time series data set; storing a result of the cyclic processing for the time series data set; and extracting the first feature by reflecting the stored result of the cyclic processing.

[0021] In another embodiment, the step of extracting the second feature may include the steps of performing a convolution operation on the time-series data set, compressing a result of the convolution operation, and correcting the result of the convolution operation based on the compressed result to extract the second feature.

[0022] In another embodiment, the operating method of the battery cell state determination device further includes a step of learning a method for generating the second state information based on an experimental dataset, and the experimental dataset may be a dataset in which data of an experimental battery cell collected under preset experimental conditions is matched with state information of the experimental battery cell. [Effects of the Invention]

[0023] The battery cell state prediction device and its operating method according to an embodiment disclosed herein can predict the state of a battery cell based on time-series data. The battery cell state prediction device and its operating method according to one embodiment disclosed in this document can verify the prediction results based on the time-series data based on experimental data collected under preset conditions.

[0024] The battery cell state prediction device and its operating method according to an embodiment disclosed herein can generate state information corresponding to the state of a battery cell after time-series data is collected.

[0025] The battery cell state determination device and its operating method according to an embodiment disclosed herein can improve the prediction accuracy of state information based on experimental data. In addition, this document can provide various other benefits that can be perceived directly or indirectly. [Brief explanation of the drawings]

[0026] [Figure 1] 1 is a block diagram illustrating a battery cell state determination device according to an embodiment disclosed herein. [Figure 2] FIG. 2 is a block diagram illustrating a first module according to one embodiment disclosed herein. [Figure 3] FIG. 2 is a block diagram illustrating a second module according to one embodiment disclosed herein. [Figure 4]1 is a diagram illustrating a method for determining the state of a battery cell of a first module according to an embodiment disclosed in this document. [Figure 5] 10A and 10B are diagrams for explaining a method for determining the state of a battery cell of a second module according to an embodiment disclosed in this document. [Figure 6] 1 is a flowchart illustrating an operation method of a battery cell state determination device according to an embodiment disclosed herein. [Figure 7] 1 is a flowchart illustrating a method for generating first status information according to an embodiment disclosed in the present document. [Figure 8] 10 is a flowchart illustrating a method for generating second state information according to an embodiment disclosed in the present document. [Figure 9] 10 is a flowchart illustrating a method for generating second status information according to another embodiment disclosed herein. [Figure 10] 1 is a block diagram showing the hardware configuration of a computing system for performing an operation method of a battery cell state determination device according to an embodiment disclosed herein; DETAILED DESCRIPTION OF THE INVENTION

[0027] Hereinafter, the embodiments disclosed herein will be described in detail with reference to exemplary drawings. When assigning reference numerals to components in each drawing, it should be noted that the same reference numerals are assigned to the same components when they appear in other drawings as much as possible. Furthermore, when describing the embodiments disclosed herein, if a detailed description of related known structures or functions is deemed to hinder understanding of the embodiments disclosed herein, such detailed description will be omitted.

[0028] In describing components of the embodiments disclosed herein, terms such as first, second, A, B, (a), (b), etc. may be used. Such terms are merely used to distinguish the component from other components and do not limit the nature, order, or sequence of the components. Furthermore, unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the art to which the embodiments disclosed herein belong. Terms defined in commonly used dictionaries should be interpreted as having a meaning consistent with the context of the relevant art, and should not be interpreted in an idealized or overly formal sense unless expressly defined in this application.

[0029] FIG. 1 is a block diagram illustrating a battery cell state determination device according to an embodiment disclosed herein. Referring to FIG. 1, a battery cell state determination device 1 according to an embodiment disclosed herein may include a first module 10 and a second module 20.

[0030] Furthermore, the first module 10 and the second module 20 can be connected to a battery management system 30 that transmits time series data sets related to the battery cells, and the second module 20 can be connected to an experimental database 40 that transmits experimental data sets.

[0031] The first module 10 is capable of generating first state information about the battery cell based on a time series data set about the battery cell. The first state information regarding the battery cell may be information that predicts the state of the battery cell by the first module 10. The first state information may include information that determines whether the battery cell is in a normal state or an abnormal state.

[0032] The first module 10 can generate predicted data on the state of the battery cell at a preset time point based on the collected time series data set, and determine the state of the battery cell based on the predicted data. A method for determining the state of the battery cells of the first module 10 will be described in detail with reference to FIG.

[0033] The second module 20 can generate second state information about the battery cell based on the time series data set about the battery cell. The second status information regarding the battery cell may be status information of the battery cell corresponding to the collected time series data set, and may include information determining whether the battery cell is in a normal or abnormal state, similar to the first status information. A method for determining the state of the battery cells of the second module 20 will be described in detail with reference to FIG.

[0034] The battery management system 30 can collect information about the battery cells in real time from the battery cells whose states are to be determined. The information about the battery cells collected in real time may be a time-series data set about the battery cells, and can be provided to the first module 10 and the second module 20.

[0035] The time-series data set related to the battery cell may include a plurality of time-series tokens. According to an embodiment, the time-series tokens may be data for each time interval obtained by applying a preset time window to data collected in real time from the battery cell.

[0036] In another embodiment, the time series token may be data for each time interval obtained by calculating feature values ​​of data collected in real time from the battery cells and applying a preset time window to the calculated feature values.

[0037] According to an embodiment, the battery management system 30 is connected to a battery module including battery cells and can collect data related to the battery cells from the battery module in real time. The data collected by the battery management system 30 may be raw data such as the voltage, current, temperature, internal resistance, and impedance of the battery cells.

[0038] The battery management system 30 can also calculate the capacity of the battery cell calculated from the raw data, the state of health (SOH), state of charge (SOC), remaining useful life (RUL), and the change in charge amount relative to the voltage change, etc. The calculation results may be statistically calculated values ​​for the battery cell.

[0039] The battery management system 30 may be installed in a vehicle. The battery management system 30 may apply a time window to data collected during actual vehicle operation and transmit the data to the first module 10 and the second module 20 as a time-series data set. According to another embodiment, the battery management system 30 may calculate feature values ​​of data collected during actual vehicle operation, apply a preset time window to the calculated feature values, and transmit the data to the first module 10 and the second module 20 as a time-series data set. The data to which the time window is applied may be a time-series token, which is data for each time interval.

[0040] The experimental database 40 can store experimental battery cell data collected under preset experimental conditions and can transmit the experimental battery cell data to the second module 20 as an experimental data set.

[0041] The experimental data set may be a data set in which experimental battery cell data and experimental battery cell state information are matched. For example, the experimental data set may be a data set in which experimental data tokens obtained by applying a time window to data collected from an experimental battery are matched with status information of the experimental battery cells corresponding to each experimental data token. The status information of the experimental battery cells may include information indicating whether the experimental battery cells are in a normal state or an abnormal state.

[0042] The experimental database 40 can receive experimental battery cell data from a plurality of experimental battery cells operating under preset experimental conditions. The status information of the experimental battery cells can be calculated by an experimental battery management system including the experimental battery cells, or by the experimental database 40. The experimental database 40 can send the experimental data set with the matching state information to the second module 20 .

[0043] The second module 20 can learn, based on the received experimental data set, how to generate state information based on the time series data set of the battery cells. In other words, the second module 20 can learn a method for generating status information based on a data set of experimental battery cells whose status information has been matched, and generate second status information from the time series data set using the learned method for generating status information.

[0044] The second state information generated by the second module 20 can be the basis for teacher forcing of the first module 10 . Supervised learning may be a method of improving learning accuracy by presenting correct values ​​(ground truth) for input values ​​when learning a module.

[0045] According to the embodiment, second state information for a time series dataset is generated via a second module 20 that has completed learning based on an experimental dataset, and the second state information is input to the first module 10 as a correct answer value for the time series dataset, thereby improving the accuracy of the first state information generated by the first module 10.

[0046] The first module 10 receives the time series data set and the second state information generated from the time series data set, thereby improving the accuracy of generating prediction data for generating the first state information. The first module 10 can receive the second state information from the second module 20 and learn by reflecting the received second state information.

[0047] According to the embodiment, the first module 10 can improve the accuracy of the predicted data by reflecting the second state information in the generation of the predicted data. The first module 10 can improve the accuracy of the predicted data generated based on the time-series data set by learning a method for generating predicted data so that first state information that matches the input second state information can be generated. In other words, the first module 10 can perform supervised learning by reflecting the second state information in the first state information.

[0048] The training of the second module 20 based on the experimental data set can be performed before the generation of the second state information based on the real-time data set, and the training of the second module 20 based on the experimental data set can be performed before the training of the first module 10.

[0049] FIG. 2 is a block diagram illustrating a first module according to one embodiment disclosed herein. The first module 10 may include an encoder block 100 that extracts context information of a time series dataset, a decoder block 200 that generates prediction data based on the context information of the time series dataset, and a decision block 300 that generates first state information based on the prediction data.

[0050] The first module 10 may be a transformer model that learns and predicts continuous data based on attention, which may be a data processing technique that improves the accuracy of data processing by assigning weights based on the relevance of input tokens.

[0051] The first module 10 is capable of receiving a time series data set comprising a plurality of time series tokens. According to an embodiment, the time-series data token may be data for each time interval generated by applying a preset time window to the time-series data collected by the battery management system 30 from the battery cells.

[0052] The first module 10 can generate forecast data using the plurality of time-series data tokens, and generate first status information regarding the battery cell based on the generated forecast data.

[0053] According to an embodiment, the first module 10 can generate first state information based on the first time series token, the second time series token, and the third time series token included in the time series data set.

[0054] The first module 10 includes an encoder block 100 that can extract context information of the time series data set from a first time series token included in the time series data set.

[0055] The first time series token may include multiple tokens for a predetermined interval so as to extract context information for the entire time series data set.

[0056] According to an embodiment, the first time-series tokens may be set differently depending on the learning method of the first module 10. The first time-series tokens may share tokens for any time interval with the second time-series tokens input to the decoder block 200.

[0057] According to other embodiments, the first time series token may include a token for a time interval earlier than the second time series token. The first time-series token may include consecutive tokens having a temporality or order so as to extract context information from the time-series data set. For example, the first time-series token may be obtained by applying a preset time window to consecutive data collected from a battery cell whose state is to be determined. Thus, the first time-series token may be consecutive sequence data related to the battery cell whose state is to be determined.

[0058] The encoder block 100 may include an encoder input layer 110 that converts input tokens into dimensions that can be learned or processed, a position encoder layer 120 that reflects the relative positions of the converted tokens in the encoder input layer 110, and an encoder layer 130 that extracts contextual information from the tokens.

[0059] The encoder input layer 110 can digitize the multiple tokens included in the input first time-series tokens by feature extraction. More specifically, the encoder input layer 110 can vectorize the multiple tokens and map each token onto a geometric space.

[0060] According to an embodiment, the encoder input layer 110 may be an embedding layer, and the tokens processed by the encoder input layer 110 may be referred to as embedding vectors.

[0061] The position encoder layer 120 can add position information to the embedding vector so that the position information of the multiple tokens included in the first time-series token can be reflected in learning or data processing.

[0062] The first module 10 receives the input of time-series data that is the basis for learning or processing at once, and can therefore reflect position information between tokens via the position encoder layer 120. In other words, the position encoder layer 120 can reflect the association between tokens in learning and prediction.

[0063] The encoder block 100 may include multiple encoder layers 130, and the number of encoder layers 130 may be a hyperparameter and a tuning value for optimization.

[0064] The encoder layer 130 can perform self-attention on the vector corresponding to the input first time-series token and transmit the context information obtained by the self-attention to the decoder layer 220.

[0065] The encoder layer 130 may include a self-attention layer 131, a normalization layer 132, a feedforward layer 133, and a normalization layer 134. The self-attention layer 131 can linearly transform vectors for tokens included in the input first time-series tokens, and generate a query vector, a key vector, and a value vector.

[0066] The self-attention layer 131 can perform attention based on the tokens contained in the first time series of tokens. The self-attention layer 131 can pay attention to all tokens included in the first time-series tokens.

[0067] Attention may be an operation of generating a query vector, a key vector, and a value vector for each token contained in the first time series token using a weight matrix, and using the query vectors, key vectors, and value vectors for all tokens to generate an output vector having attention values ​​for all tokens contained in the input first time series token.

[0068] The self-attention layer 131 can calculate the weight matrix and the vector of input tokens to generate a query vector, a key vector, and a value vector. The query vector, the key vector, and the weight matrix for generating the value vector may be parameters that are updated during the learning process.

[0069] The self-attention layer 131 calculates the similarity between the query vector corresponding to each token and all key vectors corresponding to each token, and reflects the similarity as a weight in each value vector mapped to the key vector. The sum of the value vectors reflecting the weight can be an output vector having the attention value for all tokens.

[0070] According to an embodiment, the self-attention layer 131 can perform multiple attention processes in parallel on the first time-series tokens to generate multiple output vectors. Each attention process can be performed using a different weight matrix. The operation of performing multiple attention processes in parallel can be referred to as multi-head attention. The number of attention processes performed in parallel can be a tuning value for optimization. The self-attention layer 131 can concatenate the output vectors generated as a result of the parallel attention and multiply them by an additional weight matrix.

[0071] The normalization layer 132 can prevent information loss by adding the vector input to the self-attention layer 131 to the vector output from the self-attention layer 131 and normalizing the summed vector. In other words, the normalization layer 132 can perform residual connection and layer normalization on the output of the self-attention layer 131.

[0072] The feedforward layer 133 may be a fully connected layer including multiple hidden layers. A vector input to the feedforward layer 133 may be output after being influenced by weights applied between the hidden layers included in the feedforward layer 133. The size of the hidden layers included in the feedforward layer 133 may be a tuning value for optimization.

[0073] The normalization layer 134 can prevent information loss by adding the vector input to the feedforward layer 133 to the vector output from the feedforward layer 133 and normalizing the summed vector. In other words, the normalization layer 134 can perform residual connection and layer normalization on the output of the feedforward layer 133.

[0074] The above operation can be repeated as many times as the number of encoder layers 130. In addition, the vector output from the encoder layer 130 can have the same magnitude as the vector input from the position encoder layer 120 to the encoder layer 130.

[0075] The encoder block 100 can transmit the output value of the encoder layer 130 to the decoder block 200 as context information. The encoder block 100 performs the above operation every time a time series data set is input from the battery management system 30 to the first module 10, and can update the context information. According to an embodiment, the context information may be a collection of key vectors and value vectors obtained from the input time series dataset.

[0076] The decoder block 200 can generate an output sequence corresponding to the expected data based on the second time series tokens included in the time series data set and the context information output from the encoder block 100 .

[0077] The decoder block 200 may include a decoder input layer 210 that converts input tokens into dimensions that can be learned or processed, a decoder layer 220 that performs decoding based on the converted tokens of the decoder input layer 210, and a linear mapping layer 230 that generates an output sequence.

[0078] Similar to the encoder input layer 110, the decoder input layer 210 can digitize, by feature extraction, multiple tokens included in the second time-series tokens input to the decoder block 200. More specifically, the decoder input layer 210 can vectorize multiple tokens and map each token onto a geometric space.

[0079] The second time series token may include a plurality of tokens for a predetermined entire interval. According to an embodiment, the second time-series tokens may be set differently depending on the learning method of the first module 10. The second time-series tokens may share tokens for any time interval with the first time-series tokens input to the encoder block 100.

[0080] According to other embodiments, the second time series tokens may include tokens for time intervals following the first time series token and may include tokens for time intervals preceding the third time series token.

[0081] According to an embodiment, the decoder input layer 210 may be an embedding layer, and the tokens processed by the decoder input layer 210 may be referred to as embedding vectors.

[0082] The decoder layer 220 can decode the context information output from the encoder block 100 and learn features of the time series dataset input to the first module 10 based on the decoded context information and the embedded second time series tokens.

[0083] The decoder block 200 may include multiple decoder layers 220, and the number of decoder layers 220 may be a hyper parameter and a tuning value for optimization.

[0084] Furthermore, the decoder layer 220 can infer predicted data, which is data of the battery cells for the time intervals from the second time series token onwards, based on the features of the learned time series data set.

[0085] The decoder layer 220 may include a masked self-attention layer 221, a normalization layer 222, an encoder-decoder attention layer 223, a normalization layer 224, a feedforward layer 225, and a normalization layer 226.

[0086] The masked self-attention layer 221 can linearly transform the vectors for the tokens included in the second time-series tokens to generate a query vector, a key vector, and a value vector.

[0087] The masked self-attention layer 221 can perform attention based on the tokens included in the second time-series tokens. However, when performing attention, the masked self-attention layer 221 can perform attention by masking the tokens corresponding to the time interval after each token is obtained.

[0088] Masking may be an operation that prevents the decoder block 200 from referencing tokens for time intervals after the time interval to be learned or inferred during attention calculation. Attention without masking may be substantially the same as the attention of the self-attention layer 131 included in the encoder block 100 described above.

[0089] For example, the masked self-attention layer 221 can perform attention based on the tokens contained in the second time series of tokens input to the decoder block 200, and the attention performed by the masked self-attention layer 221 may be multi-head attention.

[0090] The normalization layer 222 can prevent information loss by adding the vector input to the masked self-attention layer 221 to the vector output from the masked self-attention layer 221 and normalizing the summed vector. In other words, the normalization layer 222 can perform residual connection and layer normalization on the output of the masked self-attention layer 221.

[0091] The encoder-decoder attention layer 223 may perform multi-head attention like the aforementioned masked self-attention layer 221 or self-attention layer 131, or may be a layer that performs attention by utilizing all of the output values ​​of the encoder block 100 and the masked self-attention layer 221.

[0092] More specifically, the encoder-decoder attention layer 223 can perform attention using a vector based on the context information output by the encoder block 100 and the tokens included in the second time-series tokens.

[0093] The encoder block 100 can send context information to the encoder-decoder attention layer 223, and the context information can include key vectors and value vectors for the output values ​​of the encoder layer 130.

[0094] The encoder-decoder attention layer 223 can generate an output vector having attention values ​​for all tokens contained in the second time series token based on a query vector generated from the vector output through the mask self-attention layer 221 and the normalization layer 222 and a key vector and value vector for the output value of the encoder layer 130.

[0095] According to an embodiment, the encoder-decoder attention layer 223 performs multi-head attention, but performs parallel attention on all tokens contained in the second time series token, and can concatenate the generated output vectors and multiply them with an additional weight matrix.

[0096] The normalization layer 224 can prevent information loss by adding the vector input to the encoder-decoder attention layer 223 to the vector output from the encoder-decoder attention layer 223 and normalizing the summed vector. In other words, the normalization layer 224 can perform residual connection and layer normalization on the output of the encoder-decoder attention layer 223.

[0097] The feedforward layer 225 may be fully connected and include multiple hidden layers. A vector input to the feedforward layer 225 may be output after being influenced by weights applied between the hidden layers included in the feedforward layer 225. The size of the hidden layers included in the feedforward layer 225 may be a tuning value for optimization.

[0098] The normalization layer 226 can prevent information loss by adding the vector input to the feedforward layer 225 to the vector output from the feedforward layer 225 and normalizing the summed vector. In other words, the normalization layer 226 can perform residual connection and layer normalization on the output of the feedforward layer 225.

[0099] The above operation can be repeated for the number of decoder layers 220. In addition, the vector output from the decoder layer 220 can have the same size as the vector input from the decoder input layer 210 to the decoder layer 220.

[0100] The linear mapping layer 230 may be a fully connected layer that determines expected data, which is data that the battery management system 30 is expected to collect from the battery cells when the battery cells are in a normal state.

[0101] In other words, the predicted data may be data that the battery management system 30 is expected to collect if the battery cell is in a normal state at the time when the first module 10 generates the state information of the battery cell. The forecast data may be data having the same dimensions as the tokens contained in the time series data set input to the first module 10 .

[0102] The decision block 300 can generate first status information about the battery cell based on the forecast data and the third time series token. The third time-series token may be time-series data that the battery management system 30 actually collects from the battery cell at the time the state of the battery cell is determined.

[0103] The decision block 300 may generate first status information by comparing the expected data with the third time-series token. For example, the decision block 300 may calculate a difference between the expected data and the third time-series token, and determine that the battery cell is in an abnormal state if the calculated difference exceeds a preset value.

[0104] According to an embodiment, the decision block 300 can improve the accuracy of the generated predicted data by training the encoder block 100 and the decoder block 200 based on the second state information received from the second module 20.

[0105] The decision block 300 compares the first state information with the second state information and corrects the learning parameters of the encoder block 100 and the decoder block 200 for generating predicted data based on the comparison value. Correcting the learning parameters can improve the accuracy of generating predicted data.

[0106] FIG. 3 is a block diagram illustrating a second module according to one embodiment disclosed herein. The second module 20 can learn based on an experimental data set, which is data of experimental battery cells collected under preset experimental conditions, and generate second state information based on a time-series data set related to the battery cells.

[0107] The second module 20 may include a long short-term memory block 400 , a convolution block 500 , and a combination block 600 . The long short-term memory block 400 may include a shuffle layer 410 , a long short-term memory layer 420 , and a dropout layer 430 .

[0108] The convolution block 500 may include a convolution layer 510 and a pooling layer 520 . The experimental data set may be a data set in which experimental data tokens for preset experimental conditions are matched with state information of experimental battery cells.

[0109] The long short-term memory block 400 can output a first feature based on the received time series data set, and the convolution block 500 can output a second feature based on the time series data set. The combining block 600 can combine the first feature and the second feature to generate second state information.

[0110] The long short-term memory block 400 may include a shuffle layer 410 , a long short-term memory layer 420 , and a dropout layer 430 . The shuffle layer 410 is a layer that facilitates processing of multivariate time series data, and can improve processing speed and prevent overfitting. The long short-term memory layer 420 is a network for processing sequentially input time series data, and can include a cyclic layer 421 that cyclically processes the time series data set by using the output for the input token as input again, and a memory layer 422 that stores the processing results of the cyclic layer for the previous input token. In other words, the long short-term memory layer 420 may be a long short-term memory (LSTM) layer. According to an embodiment, the long short-term memory layer 420 may be an attention LSTM layer that performs attention. The dropout layer 430 is a layer for preventing overfitting, and can stochastically remove some connections from the fully connected layer.

[0111] The convolution block 500 may include a convolution layer 510 and a pooling layer 520. The convolution block 500 may include multiple convolution layers 510.

[0112] The convolution layer 510 can include a convolution layer 511 that performs a convolution operation on a time-series data set, a compression layer 512 that reduces the dimension of the output of the convolution layer 511, and an activation layer 513 that normalizes the output of the compression layer 512 and assigns weights to it.

[0113] The pooling layer 520 may be a layer that downsamples the operation result of the convolution layer 510 to reduce the size, and the second feature may be output through the pooling layer 520.

[0114] The combining block 600 can combine the first feature and the second feature to determine the state of the battery cell from the input time-series data set. The state of the battery cell determined by the combining block 600 may be second state information.

[0115] For example, the second state information may include information determining whether the battery cell is in a normal state or an abnormal state. The second state information may be state information of the battery cell at the time when the time series data set is input, or more specifically, may be state information of the battery cell at the time when the third time series token included in the time series data set is collected.

[0116] The second module 20 can be trained based on an experimental data set. The experimental data set may be a data set in which experimental data tokens for preset experimental conditions are matched with state information of experimental battery cells. The second module 20 can input the experimental data tokens included in the experimental data set into the long short-term memory block 400 and the convolution block 500, learn to extract features corresponding to the experimental data tokens from each block, and train the combination block 600 to combine the features to predict the state of the experimental battery cells.

[0117] More specifically, the second module 20 can improve learning accuracy by comparing the state information of the experimental battery cell matched to the experimental data token with the information predicted through the long short-term memory block 400, the convolution block 500, and the combination block 600.

[0118] The second module 20 can receive the time series dataset after learning based on the experimental dataset, and predict the state of the battery cell based on the received time series dataset.

[0119] FIG. 4 is a diagram for explaining a method for determining the state of a battery cell of a first module according to an embodiment disclosed in this document. For convenience of explanation, a case will be described as an example in which the time series data set includes first to sixth tokens (T1, T2, T3, T4, T5, and T6), and the state of the battery cell for the time interval in which the sixth token (T6) was collected is determined.

[0120] The first module 10 may select a first time-series token from the time-series data set for extracting context information of the time-series data set. Exemplarily, the first time-series token may include first to fourth tokens (T1, T2, T3, T4) to determine the state of the battery cell for the time interval in which the sixth token (T6) was collected.

[0121] The first time series token may include tokens (for example, T1, T2, T3, etc.) collected chronologically earlier than the token (T6) collected in the time interval that is the basis for determining the state of the battery cell.

[0122] The encoder block 100 can receive the first time series tokens and extract context information for the time series data set from the received first time series tokens.

[0123] The encoder block 100 can update the extracted context information each time a time series data set is input from the battery management system 30 to the first module 10 . The method for extracting context information has been explained above with reference to FIG. 2, so a duplicate explanation will be omitted.

[0124] The decoder block 200 receives the output value of the encoder block 100 as context information, and can perform attention by reflecting the context information in the encoder-decoder attention layer 223 included in the decoder layer 220.

[0125] The decoder block 200 can select a second time series token for generating prediction data from the time series data set. The second time series token may include a token (T5) collected immediately before collecting a token (T6) collected in a time interval that is a reference for determining the state of the battery cell.

[0126] According to an embodiment, the second time-series token may include multiple tokens (T4, T5). The tokens (T4, T5) included in the second time-series token may be consecutive tokens having a temporal relationship. In other words, the second time-series token may include tokens collected in consecutive time intervals, such as the fourth token and the fifth token.

[0127] The decoder block 200 can generate predicted data (T5', T6') based on the context information and the second time series token. According to an embodiment, the decision block 300 can compare the predicted data (T5', T6') with the tokens (T5, T6) corresponding to the same time interval to improve the accuracy of the decoding algorithm of the decoder block 200. The predicted data generated via the decoder block 200 may be consecutive data corresponding to the tokens input to the decoder block 200 .

[0128] According to another embodiment, the second time series token input to the decoder block 200 may include only the fifth token (T5), which is the token collected immediately before the time interval that is the basis for determining the state of the battery cell, and the decoder block 200 may generate predicted data (T6') for one time interval based on the fifth token (T5).

[0129] The predicted data (T5', T6') output from the decoder block 200 may be data that a battery cell in a normal state is predicted to have for the time period.

[0130] The decision block 300 may generate first status information (S1) by comparing the token (T6) collected from the battery cell during a time interval that is a reference for determining the actual status with the predicted data (T6').

[0131] The decision block 300 can receive the second state information (S2) from the second module 20. The decision block 300 can adjust the tuning values ​​of the encoder block 100 and the decoder block 200 so as to treat the received second state information (S2) as a correct value for the input time series data set (T1, T2, T3, T4, T5, and T6) and improve the accuracy of the first state information (S2) generation algorithm.

[0132] The encoder block 100 and the decoder block 200 can tune the learning algorithm by reflecting the second state information (S2), thereby improving the accuracy of the generated predicted data (T5', T6').

[0133] The first module 10 can improve the accuracy of the predicted data (T5', T6') generated based on the time series data set by learning how to generate predicted data so that first state information (S1) that matches the input second state information (S2) can be generated.

[0134] FIG. 5 is a diagram for explaining a method for determining the state of the battery cells of the second module according to an embodiment disclosed in this document. For convenience of explanation, a case will be described as an example in which the time series data set includes first to sixth tokens (T1, T2, T3, T4, T5, and T6), and the state of the battery cell for the time interval in which the sixth token (T6) was collected is determined.

[0135] The time series data set can be input to the long short-term memory block 400 and the convolution block 500 in parallel. The long short-term memory block 400 can extract a first feature (F1) based on consecutive tokens (T1 to T6) contained in the time series data set.

[0136] The convolution block 500 can extract a second feature (F2) based on the consecutive tokens (T1 to T6) included in the time series data set.

[0137] The operation of the long short-term memory block 400 and the convolution block 500 has been explained above with reference to FIG. 3, so a duplicate explanation will be omitted. The combining block 600 can combine the first feature (F1) and the second feature (F2) to generate second state information (S2).

[0138] The second module 20 can determine the state information with higher accuracy than the first module 10 because the second module 20 completes learning based on the experimental data set and receives the time-series data set.

[0139] In addition, since the experimental data set is a data set in which experimental data tokens for preset experimental conditions are matched with status information of the experimental battery cells, the second module 20 can improve the accuracy of determining status information based on the status information of the experimental battery cells. Therefore, the first module 10 can perform supervised learning using the second state information (S2) generated by the second module 20 as a correct answer value.

[0140] FIG. 6 is a flowchart illustrating a method of operation of a battery cell state determination device according to an embodiment disclosed herein. The second module 20 can learn how to generate the second state information based on the experimental data set (S100).

[0141] The experimental data set may be data that matches experimental data tokens with status information about experimental battery cells. The second module 20 can generate the second state information based on the time-series data set related to the battery cell by learning in advance how to generate the state information.

[0142] The battery management system 30 can collect a time series data set related to the battery cells (S200). The time series data set may be time series data collected by the battery management system 30 up to the point in time when a state determination of the battery cells is required.

[0143] The time-series data set may also include a plurality of tokens obtained by applying a preset time window to time-series data obtained from the battery cell. The tokens contained in the time-series data set may be continuous data having a temporal property.

[0144] The first module 10 may generate first state information based on the time series data set (S300). The first module 10 may include an encoder block 100, a decoder block 200, and a determination block 300. The first state information may be state information of the battery cell at a time when the state of the battery cell needs to be determined. For example, the state information of the battery cell may include information regarding whether the battery cell is normal or abnormal.

[0145] The second module 20 may include a long-short-term memory block 400, a convolution block 500, and a combination block 600, and the second status information may be, similar to the first status information, status information of the battery cell for a time when the status of the battery cell needs to be determined. The second module 20 may generate the battery cell status information in a different manner than the first module 10.

[0146] The first module 10 can perform supervised learning of the first state information generation algorithm based on the second state information (S500). The first module 10 reflects the second state information as a correct answer value, and can improve the accuracy of the prediction data generation algorithm.

[0147] FIG. 7 is a flowchart illustrating a method for generating first status information according to an embodiment disclosed herein. The encoder block 100 included in the first module 10 can extract context information of the time series data set from a first time series token included in the time series data set (S310).

[0148] The first time series token may include multiple tokens for a predetermined interval so as to extract context information for the entire time series data set. According to an embodiment, the context information may include key vectors and value vectors for the output values ​​of the encoder block 100 .

[0149] The decoder block 200 included in the first module 10 can generate predicted data based on the context information of the time series data set and the second time series tokens included in the time series data set (S320).

[0150] The second time-series tokens may include tokens collected immediately before the time when the state of the battery cell needs to be determined, and may include tokens having temporal continuity.

[0151] The predicted data may be data that is predicted to be collected from the battery cell if the battery cell is in a normal state at the time when the state of the battery cell needs to be determined.

[0152] The decoder block 200 can generate predicted data using the query vector obtained from the second time series token and the key vector and value vector for the output value of the encoder block 100.

[0153] The decision block 300 may generate first state information based on the forecast data and the third time series tokens included in the time series data set (S330). The third time-series token may be actual data of the battery cell obtained from the battery cell at the time when the state of the battery cell needs to be determined.

[0154] The decision block 300 can compare the predicted data with the third time-series token to determine whether the state of the battery cell is normal or not, and generate first state information.

[0155] In other words, when the condition of the battery cell is normal, the judgment block 300 can judge the condition of the battery cell by comparing the predicted data, which is the data that the battery cell is expected to have, with the third time series token, which is the data of the actual battery cell.

[0156] FIG. 8 is a flowchart illustrating a method for generating second status information according to an embodiment disclosed herein. The long short-term memory block 400 included in the second module 20 can extract a first feature based on the time series data set (S410). The long short-term memory block 400 may be an LSTM block including a recursive layer 421 and a memory layer 422 .

[0157] The convolution block 500 included in the second module 20 can extract second features based on the time series data set (S420). The convolution block 500 may include a convolution layer 511 that performs a convolution operation.

[0158] The combination block 600 included in the second module 20 can combine the first feature and the second feature to generate second state information (S430). The first module 10 can perform supervised learning based on the second state information.

[0159] FIG. 9 is a flowchart illustrating a method for generating second status information according to another embodiment disclosed in the present document. The long short-term memory block 400 included in the second module 20 can process the time series data set in a cyclical manner (S411). The cyclic processing of the time series data set can be performed by the cyclic layer 421 included in the long short-term memory block 400.

[0160] The long short-term memory block 400 can store the results of the cyclic processing on the time series data set (S412). The results of cyclic processing of the time series data set can be stored by the memory layer 422 included in the long short-term memory block 400.

[0161] The long short-term memory block 400 can extract the first feature by reflecting the stored cyclic processing result (S413). The convolution block 500 included in the second module 20 can perform a convolution operation on the time series data set (S421). The convolution operation can be performed by a convolution layer 511 that includes multiple hidden layers.

[0162] The convolution block 500 can compress the result of the convolution operation (S422). The compression can be performed by a compression layer 512 that the convolution block 500 includes.

[0163] The convolution block 500 can correct the result of the convolution operation based on the compression result to extract the second feature (S423). The combining block 600 can generate second state information based on the extracted first and second features.

[0164] FIG. 10 is a block diagram showing the hardware configuration of a computing system for performing the method of operating a battery cell state determination device according to an embodiment disclosed herein.

[0165] Referring to FIG. 10, a computing system 1000 according to one embodiment disclosed herein may include an MCU 1010, a memory 1020, an input / output I / F 1030, and a communication I / F 1040.

[0166] The MCU 1010 may be a processor that executes various programs stored in the memory 1020 (e.g., a battery cell voltage or current collection program, a relay control program included in the battery pack, a battery cell remaining life calculation program, a battery cell capacity degradation diagnosis program, a battery cell resistance degradation determination program, etc.), processes various information including battery cell remaining life information, battery cell capacity degradation information, and battery cell resistance degradation information through such programs, and performs the operations of the battery cell state determination device shown in FIG. 1 described above.

[0167] The memory 1020 can store various programs related to the collection and diagnosis of battery log information. The memory 1020 can also store various information related to the battery cells as time-series data, such as battery current, voltage, charge / discharge condition information, battery cell voltage information within a set number of charge / discharge cycles, and battery cell dQ / dV information within a set number of charge / discharge cycles. The memory 1020 can also include the operation algorithms of the modules 10 and 20 shown in FIG. 1.

[0168] A plurality of such memories 1020 may be provided as necessary. The memories 1020 may be volatile memories or nonvolatile memories. As the volatile memories 1020, RAM, DRAM, SRAM, etc. may be used. As the nonvolatile memories 1020, ROM, PROM, EAROM, EPROM, EEPROM, flash memory, etc. may be used. The examples of the memories 1020 listed above are merely illustrative and are not limited to these examples.

[0169] The input / output I / F 1030 can provide an interface that connects input devices (not shown) such as a keyboard, mouse, or touch panel, and output devices such as a display (not shown), to the MCU 1010, enabling data to be sent and received.

[0170] The communication I / F 1040 is configured to be able to transmit and receive various data to and from a server, and may be any device capable of supporting wired or wireless communication. For example, the battery cell state determination device may receive an experimental data set related to experimental battery cells from a separately provided external server via the communication I / F 1040. The battery cell determination device may store the received experimental data set in an experimental database. According to an embodiment, the experimental database may be provided outside the battery state determination device.

[0171] In this way, a computer program according to one embodiment disclosed in this document may be recorded in memory 1020 and processed by MCU 1010 to be realized, for example, as a module that performs each function shown in FIG. 1.

[0172] The above description is merely an illustrative example of the technical ideas disclosed in this document, and various modifications and variations are possible within the scope of those skilled in the art to which the embodiments disclosed in this document pertain without departing from the essential characteristics of the embodiments disclosed in this document.

[0173] Therefore, the embodiments disclosed in this document are intended to illustrate, not limit, the technical ideas disclosed in this document, and such embodiments do not limit the scope of the technical ideas disclosed in this document. The scope of protection of the technical ideas disclosed in this document should be interpreted according to the claims below, and all technical ideas within the scope equivalent thereto should be interpreted as being included in the scope of rights of this document.

Claims

1. a first module that generates first status information for the battery cell based on a time series data set for the battery cell; a second module that generates second status information about the battery cell based on the time series data set; Including, The first module performs supervised learning based on the second state information.

2. The battery cell state determination device according to claim 1 , wherein the time-series data set includes a plurality of time-series tokens, and the plurality of time-series tokens are continuous data having a temporal property.

3. The first module is an encoder block that extracts context information of the time series data set from a first time series token included in the time series data set; a decoder block that generates prediction data based on context information of the time series data set and second time series tokens included in the time series data set; a judgment block that generates the first state information based on the forecast data and a third time-series token included in the time-series data set.

4. The battery cell state determination device according to claim 3 , wherein the first module corrects the prediction data based on the second state information.

5. The battery cell state determination device according to claim 3 , wherein the third time-series token is data collected from the battery cell at a time when the state of the battery cell is determined.

6. The battery cell state determination device according to claim 3 , wherein the predicted data is data that is predicted to be collected from the battery cell when the battery cell is in a normal state.

7. The second module is a long short-term memory block for extracting first features based on the time series data set; a convolution block for extracting second features based on the time series data set; The battery cell state determination device according to claim 1 , further comprising: a combination block that combines the first feature and the second feature to generate the second state information.

8. The long short-term memory block includes a cyclic layer that cyclically processes the time series data set; The battery cell state determination device according to claim 7 , further comprising: a memory layer that stores a processing result of the circulation layer.

9. The battery cell state determination device according to claim 7 , wherein the convolution block includes a convolution layer that performs a convolution operation on the time-series data set.

10. The convolution block includes a compression layer that compresses the calculation result of the convolution layer; The battery cell state determination device according to claim 9 , further comprising: an activation layer that corrects the calculation result of the convolution layer based on the calculation result of the compression layer.

11. the second module learns based on an experimental data set; The battery cell state determination device according to claim 1 , wherein the experimental data set is a data set in which data of an experimental battery cell collected under preset experimental conditions is matched with state information of the experimental battery cell.

12. collecting a time series data set for a battery cell; generating first state information about the battery cell based on the time series data set; generating second state information about the battery cell based on the time series data set; performing supervised learning based on the second state information; A method for operating a battery cell state determination device, comprising:

13. The step of generating first status information includes: extracting context information of the time series dataset from a first time series token included in the time series dataset; generating forecast data based on context information of the time series dataset and second time series tokens included in the time series dataset; and generating the first state information based on the forecast data and a third time-series token included in the time-series data set.

14. The step of generating second state information includes: extracting a first feature based on the time series data set; extracting second features based on the time series data set; and combining the first feature and the second feature to generate the second status information.

15. The step of extracting the first feature includes: cyclically processing the time series data set; storing the results of the cyclic processing on the time series data set; The method of claim 14 , further comprising: extracting the first feature by reflecting the stored result of the circulation process.

16. The step of extracting the second feature includes: performing a convolution operation on the time series data set; compressing the result of the convolution operation; The method of claim 14 , further comprising: correcting the result of the convolution operation based on the result of the compressed operation to extract the second feature.

17. further comprising learning how to generate the second state information based on an experimental data set; The method of claim 12 , wherein the experimental data set is a data set in which data of an experimental battery cell collected under preset experimental conditions is matched with state information of the experimental battery cell.

Images