Battery state prediction device and operation method thereof

Abstract

A battery state prediction device according to one embodiment disclosed in this document may include a generation unit that generates a plurality of machine learning models that learn based on battery data and features included in the battery data and predict the amount of gas generation from the battery, and a controller that applies the battery data to the plurality of machine learning models, obtains a plurality of prediction data that predicts the amount of gas generation from the battery, and predicts the amount of gas generation from the battery based on the plurality of prediction data.

Description

[Technical Field]

[0001] The present invention claims the benefit of priority based on Korean Patent Application Nos. 10-2022-0178733 filed December 19, 2022 and 10-2023-0180528 filed December 13, 2023, and all contents disclosed in the documents of said Korean patent applications are incorporated herein by reference. SUMMARY OF THE INVENTION The embodiments disclosed herein relate to a battery state prediction device and method of operation. [Background technology]

[0002] Electric vehicles receive electricity from an external source to charge their batteries, and then use the voltage charged in the batteries to drive the motor to generate power. Electric vehicle batteries can generate heat due to chemical reactions that occur during the charging and discharging process, and this heat can damage the performance and lifespan of the battery and cause gas to be generated inside the battery.

[0003] Gas generated inside a battery can act as resistance or cause deformation of the battery module and battery pack, increasing the product defect rate. Therefore, a battery management system (BMS) that monitors the battery temperature, voltage, and current can be used to predict whether gas is generated inside the battery and the amount of gas generated.

[0004] The battery management device can predict the amount of gas generation from a battery by training an artificial intelligence model that analyzes the state of the battery with battery data. However, when the same battery data is input to the artificial intelligence model, the battery management device can always derive the same output data, which has a problem that it cannot reflect the actual distribution of gas generation rates that actual batteries generate in the same environment, where the distribution varies from battery to battery. Summary of the Invention [Problem to be solved by the invention]

[0005] One objective of the embodiments disclosed in this document is to provide a battery state prediction device and an operating method thereof that can obtain gas generation amount prediction data in the form of a probability distribution using multiple artificial intelligence models that predict the amount of gas generation in a battery.

[0006] 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]

[0007] A battery state prediction device according to one embodiment disclosed in this document may include a generation unit that generates a plurality of machine learning models that learn based on battery data and features included in the battery data and predict the amount of gas generation from the battery, and a controller that applies the battery data to the plurality of machine learning models, obtains a plurality of prediction data that predicts the amount of gas generation from the battery, and predicts the amount of gas generation from the battery based on the plurality of prediction data.

[0008] According to one embodiment, the plurality of machine learning models include a plurality of deep neural network (DNN) models, and the controller can apply the battery data to the plurality of deep neural network models to obtain the plurality of prediction data.

[0009] According to one embodiment, the battery data may include cumulatively measured temperature, SOC, and SOH of the battery, and the characteristics of the battery data may include at least one of the electrode type, assembly process, and separator type of the battery.

[0010] According to one embodiment, the controller may assign weights to each of the plurality of prediction data based on weights assigned based on characteristics of the battery data during the process of training a deep neural network model used to generate each of the plurality of prediction data, and input the weighted plurality of prediction data into an ensemble learning model to generate the gas generation amount prediction data in the form of a probability distribution.

[0011] According to one embodiment, the controller can calculate the average value of the gas generation amount prediction data in the form of the probability distribution, compare the average value with the already stored gas generation amount measurement value of the battery, and determine the accuracy of the multiple machine learning models.

[0012] An operating method of a battery state prediction device according to one embodiment disclosed in this document may include the steps of generating a plurality of machine learning models that are trained based on battery data and features included in the battery data and that predict the amount of gas generation from the battery, applying the battery data to the plurality of machine learning models to obtain a plurality of prediction data that predicts the amount of gas generation from the battery, and predicting the amount of gas generation from the battery based on the plurality of prediction data.

[0013] According to one embodiment, the plurality of machine learning models include a plurality of deep neural network (DNN) models, and the step of applying the battery data to the plurality of machine learning models to obtain a plurality of prediction data predicting the amount of gas generation from the battery may include the step of applying the battery data to the plurality of deep neural network models to obtain the plurality of prediction data.

[0014] According to one embodiment, the step of generating a plurality of machine learning models for predicting the amount of gas generation of a battery based on the battery data includes generating the plurality of machine learning models based on the battery data including cumulatively measured temperature, SOC, and SOH of the battery, and the characteristics of the battery data may include at least one of an electrode type, an assembly process, and a separator type of the battery.

[0015] According to one embodiment, the step of predicting the amount of gas generation of the battery based on the plurality of prediction data may include a step of assigning weights to each of the plurality of prediction data based on weights assigned based on characteristics of the battery data during the process of training the deep neural network model that generates each of the plurality of prediction data, and a step of inputting the weighted plurality of prediction data into an ensemble learning model to generate the gas generation amount prediction data in the form of a probability distribution.

[0016] According to one embodiment, the step of predicting the amount of gas generation of the battery based on the plurality of prediction data may include calculating an average value of the gas generation amount prediction data in the form of the probability distribution, comparing the average value with previously stored measured values ​​of the amount of gas generation of the battery, and determining the accuracy of the plurality of machine learning models. [Effects of the Invention]

[0017] The battery state prediction device and its operating method according to an embodiment disclosed herein can obtain gas generation amount prediction data in the form of a probability distribution using multiple artificial intelligence models. [Brief explanation of the drawings]

[0018] [Figure 1] FIG. 1 illustrates a battery pack according to one embodiment disclosed herein. [Figure 2] 1 is a diagram for generally explaining a battery state prediction device according to an embodiment disclosed herein; [Figure 3] 1 is a block diagram showing the configuration of a battery state prediction device according to an embodiment disclosed in this document. [Figure 4] FIG. 10 is a diagram for explaining the operation of a generating unit according to an embodiment disclosed in this document. [Figure 5] FIG. 10 is a diagram illustrating a method of operation of a controller according to an embodiment disclosed herein. [Figure 6] 1 is a flowchart illustrating a method of operation of a battery state prediction device according to an embodiment disclosed herein. [Figure 7] FIG. 1 is a block diagram showing the hardware configuration of a computing system that realizes a battery state prediction device according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0019] Some embodiments disclosed herein will be described in detail below 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 as long as possible when they appear in other drawings. 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.

[0020] 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 pertain. 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 herein.

[0021] FIG. 1 is a diagram illustrating a battery pack according to one embodiment disclosed herein. Referring to FIG. 1 , a battery pack 1000 according to one embodiment disclosed herein may include a battery module 100 , a battery state prediction device 200 , and a relay 300 .

[0022] The battery module 100 may include a plurality of battery cells 110, 120, 130, and 140. Although the number of battery cells is shown as four in FIG. 1, the number of battery cells is not limited to four, and the battery module 100 may include n (n is a natural number equal to or greater than 2) battery cells.

[0023] The battery module 100 can supply power to a target device (not shown). To this end, the battery module 100 can be electrically connected to the target device. Here, the target device can include an electrical, electronic, or mechanical device that operates by receiving power from a battery pack 1000 including a plurality of battery cells 110, 120, 130, and 140. For example, the target device can be, but is not limited to, an electric vehicle (EV) or an energy storage system (ESS).

[0024] The plurality of battery cells 110, 120, 130, 140 are basic units of a battery that can be used by charging and discharging electrical energy, and may be, but are not limited to, lithium ion (Li-ion) batteries, lithium ion polymer (Li-ion polymer) batteries, nickel cadmium (Ni-Cd) batteries, nickel metal hydride (Ni-MH) batteries, etc. Meanwhile, although Fig. 1 shows one battery module 100, according to an embodiment, a plurality of battery modules 100 may be included.

[0025] The battery state prediction device 200 can predict the amount of gas generated from the multiple battery cells 110, 120, 130, and 140 based on the temperature, current, voltage, SOC (State of Charge), and SOH (State of Health) data of the multiple battery cells 110, 120, 130, and 140. The battery state prediction device 200 can predict the amount of gas generated from the multiple battery cells 110, 120, 130, and 140 based on battery data (A) of the multiple battery cells 110, 120, 130, and 140.

[0026] According to an embodiment, the battery state prediction device 200 may be implemented in the form of a battery management system (BMS). Also, according to an embodiment, the battery state prediction device 200 may be mounted in the battery management system.

[0027] Here, the battery management unit can manage and / or control the state and / or operation of the battery module 100. For example, the battery management unit can manage and / or control the state and / or operation of the plurality of battery cells 110, 120, 130, and 140 included in the battery module 100. The battery management unit can manage charging and / or discharging of the battery module 100.

[0028] The battery management device can also monitor the voltage, current, temperature, etc. of the battery module 100 and / or each of the plurality of battery cells 110, 120, 130, and 140 included in the battery module 100. For monitoring via the battery management device, sensors and various measurement modules (not shown) can be further provided in the battery module 100, a charge / discharge path, or any position on the battery module 100. The battery management device can calculate parameters indicating the state of the battery module 100, such as SOC and SOH, based on the measured values ​​of the monitored voltage, current, temperature, etc.

[0029] The battery management unit can control the operation of the relay 300. For example, the battery management unit can short-circuit the relay 300 to supply power to a target device. In addition, the battery management unit can short-circuit the relay 300 when a charging device is connected to the battery pack 1000.

[0030] The battery management unit can calculate a cell balancing time for each of the multiple battery cells 110, 120, 130, and 140. Here, the cell balancing time can be defined as the time required to balance the battery cells. For example, the battery management unit can calculate the cell balancing time based on the SOC, battery capacity, and balancing efficiency of each of the multiple battery cells 110, 120, 130, and 140.

[0031] FIG. 2 is a diagram for generally explaining a battery state prediction device according to one embodiment disclosed herein. Referring to FIG. 2, the battery state prediction device 200 can extract a portion of the battery data (A) and input it to a plurality of machine learning models 211, 212, 213, and 214.

[0032] Although Figure 2 shows that the number of machine learning models is four, this is not limited to four, and the number of machine learning models 211, 212, 213, and 214 can be configured to include n (n is a natural number greater than or equal to 2) machine learning models.

[0033] The battery state prediction device 200 can acquire battery data (A) of the multiple battery cells 110, 120, 130, and 140. The battery management device can acquire battery data (A) including battery measurement values ​​from the voltage value at which the battery's SOC is 0% to the voltage value at which the battery's SOC is 100%, in order to check the amount of gas generated in the multiple battery cells 110, 120, 130, and 140. Therefore, the battery state prediction device 200 can acquire battery data (A) including the voltage, current, temperature, SOC, and SOH of the multiple battery cells 110, 120, 130, and 140 measured cumulatively during a charge / discharge period.

[0034] Here, the battery data (A) may further include at least one feature of the battery data (A) including the separator type, assembly process, electrode type, etc. That is, the battery data (A) may further include at least one data related to the separator type of each battery, the assembly process used to assemble the battery, and / or the electrode type of the battery.

[0035] The battery state prediction device 200 can input battery data (A) into a plurality of machine learning models 211, 212, 213, and 214 to predict the amount of gas generated by the battery. Here, machine learning is a technology that enables a computer to learn and predict a certain result. In general, the results obtained using machine learning include a process of preparing training data for machine learning, training the machine in a manner appropriate for the problem, validating the model with test data, and predicting a result using the model that has passed the validation.

[0036] In machine learning, it is important that the training data closely represents the features that are being generalized through machine learning, so training data is generated using limited training data selected according to certain criteria. If the features to be generalized through machine learning are poorly related to the characteristics of the training data, sampling noise occurs, making it difficult for the machine problem analysis model to find inherent patterns. This increases model error and reduces the model's reliability, regardless of the accuracy of the machine problem analysis model itself. Therefore, machine learning technology requires time to invest in training data evaluation and data processing to select a training dataset.

[0037] The battery state prediction device 200 can generate a learning data set by extracting at least a portion of the battery data (A) in order to generate and train a plurality of machine learning models 211, 212, 213, and 214. Here, the plurality of machine learning models 211, 212, 213, and 214 may mean learning models that can predict the state of a battery, including the amount of gas generated by the battery, based on input battery data.

[0038] That is, the battery state prediction device 200 can generate a plurality of machine learning models 211, 212, 213, and 214 having the same structure based on one learning data set generated by extracting at least a portion of the battery data (A). The battery state prediction device 200 can acquire prediction data that predicts the amount of gas generated from the battery of each of the plurality of machine learning models 211, 212, 213, and 214, using the plurality of machine learning models 211, 212, 213, and 214 generated based on one learning data set generated by extracting at least a portion of the battery data (A).

[0039] The battery state prediction device 200 can combine the gas generation amount prediction data of each of the multiple machine learning models 211, 212, 213, and 214 and finally obtain one gas generation amount prediction data (C) in the form of a probability distribution. Therefore, the battery state prediction device 200 obtains output data (Output) of each of the multiple machine learning models 211, 212, 213, and 214 having the same structure based on one input data (Input), and combines the respective output data to finally obtain one gas generation amount prediction data (C) in the form of a probability distribution, thereby reducing learning errors and increasing reliability.

[0040] The output data generated by an AI model can include random error and main effect. When an AI model repeats an experiment multiple times based on the same input data, it can obtain different output data, which are random error or white noise. Here, if the AI ​​model repeatedly performs the same experiment a sufficient number of times, the average value of the random error will converge to 0, and the AI ​​model can obtain only main effect data.

[0041] The battery state prediction device 200 can obtain the effect of conducting repeated experiments in the same experimental environment by using the same input data to input the same input data to multiple machine learning models 211, 212, 213, and 214 and simultaneously obtaining multiple output data. That is, the battery state prediction device 200 can predict the amount of gas generated by a battery with high accuracy corresponding to the main effect data by inputting at least a portion of the same battery data (A) to multiple machine learning models 211, 212, 213, and 214 having the same structure.

[0042] FIG. 3 is a block diagram showing the configuration of a battery state prediction device according to one embodiment disclosed in this document, and FIG. 4 is a diagram for explaining the operation of a generation unit according to one embodiment disclosed in this document.

[0043] The configuration and operation of the battery state prediction device 200 will be specifically described below with reference to FIGS. Referring to FIG. 3, the battery state prediction device 200 may include a generation unit 210 and a controller 220.

[0044] The generating unit 210 may collect battery data (A). For example, the battery data (A) may be defined as a value recording a change in the state of the battery from a discharged state to a fully charged state or from a fully charged state to a discharged state of the plurality of battery cells 110, 120, 130, and 140. For example, the battery data (A) may include cumulatively measured voltage, current, temperature, SOC, and SOH of the battery. Here, the SOH may include the degree of capacity degradation and the degree of resistance degradation of the battery. In addition, the battery data (A) may further include at least one of data related to the type of separator of the battery, the assembly process used, and / or the electrode type.

[0045] The generation unit 210 can extract at least a portion of the battery data (A) as a training data set and generate multiple machine learning models 211, 212, 213, and 214. For example, the generation unit 210 can extract 80% of the battery data (A) as a training data set and generate multiple machine learning models 211, 212, 213, and 214.

[0046] According to an embodiment, the multiple machine learning models 211, 212, 213, and 214 may include multiple deep neural network (DNN) models. A deep neural network model is an artificial neural network technology that includes multiple hidden layers between an input layer and an output layer. A deep neural network model includes multiple hidden layers and is capable of learning various complex nonlinear relationships. The generation unit 210 can generate multiple machine learning models 211, 212, 213, and 214 that can predict the amount of gas generation in a battery using at least a portion of the battery data (A) as a training dataset. The amount of gas generation in a battery may increase as the SOC and temperature of the battery increase. The amount of gas generation in a battery may also vary depending on the type of battery separator, the assembly process used to assemble the battery, and / or the type of battery electrodes. Therefore, the multiple machine learning models 211, 212, 213, and 214 can predict whether gas will be generated inside the battery and the amount of gas generated based on the battery voltage, current, temperature, SOC, SOH, as well as the assembly process, electrode type, and separator type included in the battery data (A).

[0047] 4, the generation unit 210 can generate a plurality of machine learning models 211, 212, 213, and 214 trained based on at least one feature of the battery data (A). For example, when considering the separator type and electrode type among the features of the battery data (A), the generation unit 210 can extract, as a training dataset, battery data (A) including the separator type and electrode type of the battery in addition to the voltage, current, temperature, SOC, and SOH of the battery, and generate a plurality of machine learning models 211, 212, 213, and 214.

[0048] Here, the generation unit 210 assigns various weights to the characteristics of the battery data (A), including the battery separator type, assembly process, and electrode type, among the various characteristics included in the battery data (A), and allows multiple machine learning models 211, 212, 213, and 214 to learn the battery data (A).

[0049] For example, the generation unit 210 may generate the plurality of machine learning models 211, 212, 213, and 214 by minimizing the intervention of relatively unimportant data among the separator type, assembly process, and electrode type of the battery based on a drop-out method (a) and utilizing a regularization method for eliminating overfitting of the plurality of machine learning models 211, 212, 213, and 214. That is, the generation unit 210 may generate the plurality of machine learning models 211, 212, 213, and 214 by training battery data (A) from which node connections associated with characteristics for which intervention is not desired or minimized among the separator type, assembly process, and electrode type of the battery are removed.

[0050] As another example, the generation unit 210 may generate a plurality of machine learning models 211, 212, 213, and 214 by training the battery data (A) while fixing weights (b) assigned to characteristics of the battery data (A), including the separator type, assembly process, and electrode type of the battery, among various characteristics included in the battery data (A). For example, when attempting to generate a machine learning model that is significantly affected by the separator type of the battery, the generation unit 210 may train the battery data (A) by assigning a high weight (e.g., *a) to the separator type among the characteristics of the battery data (A). Furthermore, to generate a machine learning model that is less affected by the electrode type of the battery, the generation unit 210 may train the battery data (A) by assigning a low weight (e.g., *b) to the electrode type among the characteristics of the battery data (A).

[0051] As another example, the generation unit 210 may generate a plurality of machine learning models 211, 212, 213, and 214 that are trained on the battery data (A) by applying (c) biases (e.g., +a', +b') to nodes connected to the battery data (A), including the battery separator type, assembly process, and electrode type among various characteristics included in the battery data (A).

[0052] In addition, the generation unit 210 can generate multiple machine learning models 211, 212, 213, and 214 that are trained on the battery data (A) based on weight constraints including L1 constraints (Lasso) and L2 constraints (Ridge), and an embodiment disclosed in this document is not limited to such examples.

[0053] The generating unit 210 may perform min-max scaling on at least a portion of the battery data (A). Here, max-min scaling is a method of adjusting the range of all variables because if the size or unit of a numeric variable differs for each variable, the effect on the dependent variable is not properly reflected. The generating unit 210 may perform min-max scaling on at least a portion of the battery data (A) to convert at least a portion of the battery data (A) into a value between 0 and 1.

[0054] The generation unit 210 can perform K-fold cross validation on the multiple machine learning models 211, 212, 213, and 214 to determine the accuracy of the multiple machine learning models 211, 212, 213, and 214. K-fold cross validation is a method in which a preprocessed dataset is divided into a training dataset and a test set, the training dataset is divided into "K" folds, one fold is used for validation, and (K-1) folds are used for model training, allowing all data to be used in the training and validation processes. For example, the generation unit 210 can perform 5-fold cross validation on the multiple machine learning models 211, 212, 213, and 214 to determine the accuracy of the multiple machine learning models 211, 212, 213, and 214.

[0055] Furthermore, for example, the generation unit 210 can evaluate the performance of the multiple machine learning models 211, 212, 213, and 214 based on the MAE (Mean Absolute Error), which is obtained by converting the difference (Error) between the actual value and the predicted value into an absolute value and averaging it.

[0056] The controller 220 can use at least a portion of the battery data (A) including voltage, current, temperature, SOC, and SOH changes of the multiple battery cells 110, 120, 130, 140 as a test data set to input to multiple machine learning models 211, 212, 213, 214.

[0057] For example, after the generation unit 210 extracts 80% of the battery data (A) as a learning dataset and generates multiple machine learning models 211, 212, 213, and 214, the controller 220 can extract the remaining 20% ​​of the battery data (A) as a test dataset and input it to the multiple machine learning models 211, 212, 213, and 214.

[0058] FIG. 5 is a diagram illustrating a method of operation of a controller according to an embodiment disclosed herein. 5, the controller 220 applies at least a portion of the battery data (A) to a plurality of machine learning models 211, 212, 213, and 214, and can obtain a plurality of pieces of prediction data B1, B2, B3, and B4 that predict the amount of gas generated from the battery from each of the plurality of machine learning models 211, 212, 213, and 214. That is, the controller 220 can input at least a portion of the battery data (A), which is one piece of input data (Input), to each of the separate machine learning models and obtain a plurality of pieces of output data (Output). The controller 220 can generate gas generation amount prediction data (C) that predicts the amount of gas generated from the battery based on the plurality of prediction data B1, B2, B3, and B4.

[0059] The controller 220 can predict the state of the battery based on the plurality of prediction data B1, B2, B3, and B4. According to an embodiment, the controller 220 can assign weights x1, x2, x3, and x4 to the plurality of prediction data B1, B2, B3, and B4 generated by the plurality of machine learning models 211, 212, 213, and 214, respectively, based on at least one of the characteristics of the battery data (A), such as the separator type, the assembly process, and the electrode type of the battery. For example, when assigning weight x1 to the first prediction data B1 based on one of the characteristics (features) of the battery data (A), the size of the first prediction data B1 can be increased, thereby increasing the proportion of the first prediction data B1 in the total prediction data B1, B2, B3, and B4. Here, the weights x1, x2, x3, and x4 assigned to each of the multiple prediction data B1, B2, B3, and B4 can be associated with the weights assigned to the characteristics of the battery data (A) when training each of the multiple machine learning models 211, 212, 213, and 214.

[0060] The controller 220 can assign weights to the plurality of prediction data B1, B2, B3, and B4 according to weights assigned based on the characteristics of the battery data (A) during the process of training the deep neural network model used to generate each of the plurality of prediction data B1, B2, B3, and B4. That is, when training the plurality of machine learning models 211, 212, 213, and 214, the weights assigned to the characteristics of the battery data (A) are distinguished from the weights assigned to each of the plurality of prediction data B1, B2, B3, and B4. For example, in order to determine the weight to be assigned to the prediction data B1, the controller 220 can take into account the weights assigned based on the characteristics of the battery data (A) during training of the deep neural network model included in the machine learning model 211 that generates the prediction data B1.

[0061] Similarly, in order to determine the weight to be assigned to the prediction data B2, the controller 220 can take into consideration the weight assigned based on the characteristics of the battery data (A) when training the deep neural network model included in the machine learning model 212 that generates the prediction data B2, take into consideration the weight assigned based on the characteristics of the battery data (A) when training the deep neural network model included in the machine learning model 213 that generates the prediction data B3, and take into consideration the weight assigned based on the characteristics of the battery data (A) when training the deep neural network model included in the machine learning model 214 that generates the prediction data B4.

[0062] According to an embodiment, weights assigned to the plurality of prediction data B1, B2, B3, and B4 may be proportional to weights assigned based on the characteristics of the battery data (A) during training of the respective machine learning models 211, 212, 213, and 214. For example, if characteristics of the battery data (A) related to the separator type and the assembly process are taken into account during training of the machine learning model 211 that generates the prediction data B1, the controller 220 may assign to the prediction data B1 a weight corresponding to the separator type and a weight corresponding to the assembly process type, and the weights corresponding to the separator type and the assembly process may be weights proportional to the weights assigned to the characteristics of the battery data related to the separator type and the weights assigned to the characteristics of the battery data related to the assembly process during training of the machine learning model 211. Here, the weight x1 may be, but is not limited to, the sum of the weight corresponding to the separator type and the weight corresponding to the assembly process type.

[0063] That is, the controller 220 can calculate the final gas generation amount prediction data (C) by applying a fuzzy algorithm that can reflect the importance and characteristics of specific variables, rather than simply calculating the average of the multiple prediction data B1, B2, B3, B4 by assigning weights x1, x2, x3, x4 to each of the multiple prediction data B1, B2, B3, B4.

[0064] For example, the controller 220 can input multiple pieces of prediction data B1, B2, B3, and B4, each assigned with weights x1, x2, x3, and x4, into an ensemble learning model to generate gas generation amount prediction data (C) in the form of a probability distribution. Here, the ensemble learning model is a machine learning method that combines two or more learning models to perform better than a single learning model. When the reliability of each model differs, the ensemble learning model can calculate a weighted sum by assigning weights to the output data of each model rather than averaging the output data of each model. Here, the weighted sum can be defined as an average value calculated by reflecting weights corresponding to the importance or influence of the data values ​​when calculating the average of the data.

[0065] The controller 220 can calculate the mean value (Mean) of the probability distribution type gas generation amount prediction data (C). Specifically, the controller 220 can calculate the 95% prediction interval (Prediction Interval) of the mean value (Mean) of the probability distribution type gas generation amount prediction data (C). For example, the controller 220 can calculate the mean value (Mean) and standard deviation of the probability distribution type gas generation amount prediction data (C), and then calculate "mean value ± standard deviation * 1.96" as the 95% prediction interval of the mean value of the probability distribution type gas generation amount prediction data (C).

[0066] The controller 220 can compare the prediction interval of the probability distribution type gas generation amount prediction data (C) with the already stored measured gas generation amount of the battery to determine the accuracy of the multiple machine learning models 211, 212, 213, 214.

[0067] As described above, the battery state prediction device according to one embodiment disclosed in this document can obtain gas generation amount prediction data in the form of a probability distribution using multiple artificial intelligence models that predict the amount of gas generated in a battery.

[0068] The battery state prediction device inputs a small amount of input data into multiple artificial intelligence models to calculate highly accurate final prediction data, thereby reducing the time and cost required for data collection and management.

[0069] Furthermore, the battery state prediction device can set weights according to the characteristics of the data, and can reflect the inherent characteristics of the actual battery data and the characteristics of the actual battery usage environment.

[0070] FIG. 6 is a flowchart illustrating a method of operation of a battery state prediction device according to one embodiment disclosed herein. The battery state predicting device 200 is substantially similar to the battery state predicting device 200 described with reference to FIGS. 1 to 5, and therefore will be described briefly below to avoid duplication of description.

[0071] Referring to Figure 6, the operating method of the battery state prediction device may include a step (S101) of generating a plurality of machine learning models that predict the amount of gas generated in the battery based on battery data, a step (S102) of applying the battery data to the plurality of machine learning models to obtain a plurality of prediction data that predicts the amount of gas generated in the battery, and a step (S103) of predicting the amount of gas generated in the battery based on the plurality of prediction data.

[0072] Steps S101 to S103 will be specifically described below. In step S101, the generating unit 210 may collect battery data (A). For example, the battery data (A) may be defined as a value recording a change in the state of the battery from a discharged state to a fully charged state or from a fully charged state to a discharged state of the plurality of battery cells 110, 120, 130, and 140. For example, the battery data (A) may include cumulatively measured voltage, current, temperature, SOC, and SOH of the battery. Here, the SOH may include the degree of capacity degradation and the degree of resistance degradation of the battery. In addition, the battery data (A) may further include at least one of data related to the type of separator of the battery, the assembly process used, and / or the electrode type.

[0073] In step S101, the generation unit 210 can extract at least a portion of the battery data (A) as a training data set and generate multiple machine learning models 211, 212, 213, and 214. In step S101, the generation unit 210 can extract, for example, 80% of the battery data (A) as a training data set and generate multiple machine learning models 211, 212, 213, and 214.

[0074] In step S101, the generation unit 210 can generate a plurality of machine learning models 211, 212, 213, and 214 trained based on at least one or more characteristics (features) of the battery data (A). For example, when the separator type and electrode type of the characteristics (features) of the battery data (A) are taken into consideration, the generation unit 210 can extract, as a training dataset, battery data (A) including the separator type and electrode type of the battery in addition to the voltage, current, temperature, SOC, and SOH of the battery, and generate a plurality of machine learning models 211, 212, 213, and 214.

[0075] Here, the generation unit 210 assigns various weights to the characteristics of the battery data (A), including the battery separator type, assembly process, and electrode type, among the various characteristics included in the battery data (A), and allows multiple machine learning models 211, 212, 213, and 214 to learn the battery data (A).

[0076] For example, the generation unit 210 may generate the plurality of machine learning models 211, 212, 213, and 214 by minimizing the intervention of relatively unimportant data among the separator type, assembly process, and electrode type of the battery based on a drop-out method (a) and utilizing a regularization method for eliminating overfitting of the plurality of machine learning models 211, 212, 213, and 214. That is, the generation unit 210 may generate the plurality of machine learning models 211, 212, 213, and 214 by training battery data (A) from which node connections associated with characteristics for which intervention is not desired or minimized among the separator type, assembly process, and electrode type of the battery are removed.

[0077] As another example, the generation unit 210 may generate a plurality of machine learning models 211, 212, 213, and 214 by training the battery data (A) while fixing weights (b) assigned to characteristics of the battery data (A), including the separator type, assembly process, and electrode type of the battery, among various characteristics included in the battery data (A). For example, when attempting to generate a machine learning model that is significantly affected by the separator type of the battery, the generation unit 210 may train the battery data (A) by assigning a high weight (e.g., *a) to the separator type among the characteristics of the battery data (A). Furthermore, to generate a machine learning model that is less affected by the electrode type of the battery, the generation unit 210 may train the battery data (A) by assigning a low weight (e.g., *b) to the electrode type among the characteristics of the battery data (A).

[0078] As another example, the generation unit 210 may generate a plurality of machine learning models 211, 212, 213, and 214 that are trained on the battery data (A) by applying (c) biases (e.g., +a', +b') to nodes connected to the battery data (A), including the battery separator type, assembly process, and electrode type among various characteristics included in the battery data (A).

[0079] In addition, the generation unit 210 can generate multiple machine learning models 211, 212, 213, and 214 that are trained on the battery data (A) based on weight constraints including L1 constraints (Lasso) and L2 constraints (Ridge), and an embodiment disclosed in this document is not limited to such examples.

[0080] In step S101, according to an embodiment, the multiple machine learning models 211, 212, 213, and 214 may include multiple deep neural network (DNN) models. A deep neural network model is an artificial neural network technology that includes multiple hidden layers between an input layer and an output layer. A deep neural network model includes multiple hidden layers and is capable of learning various complex nonlinear relationships. In step S101, the generation unit 210 may use at least a portion of the battery data (A) as a training dataset to generate the multiple machine learning models 211, 212, 213, and 214 that can predict the amount of gas generation from a battery.

[0081] In step S101, the generating unit 210 may perform min-max scaling on at least a portion of the battery data (A). Here, max-min scaling is a method of adjusting the range of all variables because if the size or unit of a numeric variable differs for each variable, the effect on the dependent variable is not properly reflected. The generating unit 210 may perform min-max scaling on at least a portion of the battery data (A) to convert at least a portion of the battery data (A) into a value between 0 and 1.

[0082] In step S101, the generation unit 210 can perform K-fold cross validation on the multiple machine learning models 211, 212, 213, and 214 to determine the accuracy of the multiple machine learning models 211, 212, 213, and 214. K-fold cross validation is a method in which a preprocessed dataset is divided into a training dataset and a test set, the training dataset is divided into "K" folds, one fold is used for validation, and (K-1) folds are used for model training, allowing all data to be used in the training and validation processes. For example, the generation unit 210 can perform 5-fold cross validation on the multiple machine learning models 211, 212, 213, and 214 to determine the accuracy of the multiple machine learning models 211, 212, 213, and 214.

[0083] In step S101, for example, the generation unit 210 can also evaluate the performance of multiple machine learning models 211, 212, 213, and 214 based on the MAE (Mean Absolute Error), which is obtained by converting the difference (Error) between the actual value and the predicted value into an absolute value and averaging it.

[0084] In step S102, the controller 220 can use at least a portion of the battery data (A) including the voltage, current, temperature, SOC, SOH change, battery separator type, electrode type, and assembly process of the multiple battery cells 110, 120, 130, 140 as a test data set and input it into multiple machine learning models 211, 212, 213, 214.

[0085] In step S102, for example, after the generation unit 210 extracts 80% of the battery data (A) as a learning dataset and generates multiple machine learning models 211, 212, 213, and 214, the controller 220 can extract the remaining 20% ​​of the battery data (A) as a test dataset and input it to the multiple machine learning models 211, 212, 213, and 214.

[0086] In step S102, the controller 220 applies at least a portion of the battery data (A) to multiple machine learning models 211, 212, 213, and 214, and can obtain multiple prediction data B1, B2, B3, and B4 that predict the amount of gas generated by the battery from each of the multiple machine learning models 211, 212, 213, and 214.

[0087] In step S102, the controller 220 inputs at least a portion of the battery data (A), which is one piece of input data (Input), into separate machine learning models, and can obtain multiple pieces of output data (Output). The controller 220 can generate gas generation amount prediction data (C) that predicts the amount of gas generation from the battery based on the multiple pieces of prediction data B1, B2, B3, and B4.

[0088] In step S102, the controller 220 can apply at least a portion of the battery data (A) to a plurality of deep neural network models (DNN) included in a plurality of machine learning models 211, 212, 213, and 214.

[0089] In step S103, the controller 220 can predict the state of the battery based on the plurality of prediction data B1, B2, B3, and B4. In step S103, according to the embodiment, the controller 220 can assign weights x1, x2, x3, and x4 to each of the plurality of prediction data B1, B2, B3, and B4 generated by the plurality of machine learning models 211, 212, 213, and 214 based on the characteristics of the acquired battery data (A). Here, the weights x1, x2, x3, and x4 assigned to each of the plurality of prediction data B1, B2, B3, and B4 can be associated with weights assigned to the characteristics of the battery data (A) when training each of the plurality of machine learning models 211, 212, 213, and 214.

[0090] According to the embodiment, the controller 220 can assign weights to the plurality of prediction data B1, B2, B3, and B4 according to weights assigned based on the characteristics of the battery data (A) during training of the deep neural network model used to generate each of the plurality of prediction data B1, B2, B3, and B4. That is, when training the plurality of machine learning models 211, 212, 213, and 214, the weights assigned to the characteristics of the battery data (A) are distinguished from the weights assigned to each of the plurality of prediction data B1, B2, B3, and B4. For example, in order to determine the weight to be assigned to the prediction data B1, the controller 220 can take into account the weights assigned based on the characteristics of the battery data (A) during training of the deep neural network model included in the machine learning model 211 that generates the prediction data B1.

[0091] Similarly, in order to determine the weight to be assigned to the prediction data B2, the controller 220 can take into consideration the weight assigned based on the characteristics of the battery data (A) when training the deep neural network model included in the machine learning model 212 that generates the prediction data B2, take into consideration the weight assigned based on the characteristics of the battery data (A) when training the deep neural network model included in the machine learning model 213 that generates the prediction data B3, and take into consideration the weight assigned based on the characteristics of the battery data (A) when training the deep neural network model included in the machine learning model 214 that generates the prediction data B4.

[0092] In step S103, for example, the controller 220 can input multiple pieces of prediction data B1, B2, B3, and B4, each assigned with weights x1, x2, x3, and x4, into an ensemble learning model to generate gas generation amount prediction data (C) in the form of a probability distribution. Here, the ensemble learning model is a machine learning method that combines two or more learning models and exhibits better performance than a single learning model.

[0093] In step S103, the controller 220 can calculate the mean value (Mean) of the probability distribution type gas generation amount prediction data (C). Specifically, in step S103, the controller 220 can calculate the 95% prediction interval (Prediction Interval) of the mean value (Mean) of the probability distribution type gas generation amount prediction data (C). In step S103, for example, the controller 220 can calculate the mean value (Mean) and standard deviation of the probability distribution type gas generation amount prediction data (C), and then calculate "mean value ± standard deviation * 1.96" as the 95% prediction interval of the mean value of the probability distribution type gas generation amount prediction data (C).

[0094] In step S103, the controller 220 can compare the prediction interval of the probability distribution form of gas generation amount prediction data (C) with the already stored measured gas generation amount of the battery to determine the accuracy of the multiple machine learning models 211, 212, 213, and 214.

[0095] Referring to FIG. 7, a computing system 2000 according to one embodiment disclosed herein may include an MCU 2100, a memory 2200, an input / output I / F 2300, and a communication I / F 2400.

[0096] The MCU 2100 may be a processor that executes various programs (e.g., a battery gas generation amount prediction program) stored in the memory 2200, processes various data through such programs, and performs the functions of the battery state prediction device 200 shown in Figure 1 described above.

[0097] The memory 2200 can store various programs related to the operation of the battery state prediction device 200. The memory 2200 can also store operation data of the battery state prediction device 200.

[0098] A plurality of such memories 2200 may be provided as necessary. The memories 2200 may be volatile memories or nonvolatile memories. The volatile memories 2200 may be RAM, DRAM, SRAM, etc. The nonvolatile memories 2200 may be ROM, PROM, EAROM, EPROM, EEPROM, flash memory, etc. The examples of the memories 2200 listed above are merely illustrative and are not limited to these examples.

[0099] The input / output I / F 2300 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 2100, enabling data to be sent and received.

[0100] The communication I / F 2400 is configured to be able to send and receive various data to and from a server, and may be any device that supports wired or wireless communication. For example, programs for measuring resistance and diagnosing abnormalities and various data can be sent and received from a separately provided external server via the communication I / F 2400.

[0101] The above description merely exemplifies the technical ideas of the present disclosure, and various modifications and variations are possible by a person having ordinary knowledge in the technical field to which the present disclosure pertains, without departing from the essential characteristics of the present disclosure.

[0102] Therefore, the embodiments disclosed in this disclosure are intended to illustrate, not limit, the technical idea of ​​the disclosure, and the scope of the technical idea of ​​the disclosure is not limited by such embodiments. The scope of protection of the disclosure should be interpreted by the claims below, and all technical ideas within the equivalent range should be interpreted as being included in the scope of rights of the disclosure.

Claims

1. a generation unit that generates a plurality of machine learning models that are trained based on battery data and characteristics included in the battery data and that predict the amount of gas generation from the battery; a controller that applies the battery data to the plurality of machine learning models to obtain a plurality of prediction data that predicts the amount of gas generated from the battery, and predicts the amount of gas generated from the battery based on the plurality of prediction data; A battery state prediction device comprising:

2. the plurality of machine learning models includes a plurality of deep neural network models; The controller The battery state prediction device according to claim 1 , further comprising: applying the battery data to the plurality of deep neural network models to obtain the plurality of prediction data.

3. the battery data includes cumulatively measured temperature, SOC, and SOH of the battery; The battery state prediction device according to claim 2 , wherein the battery data characteristics include at least one of an electrode type, an assembly process, and a separator type of the battery.

4. The controller 4. The battery state prediction device according to claim 2 or 3, wherein, in the process of training each of the plurality of deep neural network models that generate each of the plurality of prediction data, weights are assigned to each of the plurality of prediction data based on weights assigned based on characteristics of the battery data, and the plurality of prediction data to which the weights have been assigned are input into an ensemble learning model to generate the plurality of prediction data in the form of a probability distribution.

5. the controller calculates an average value of the plurality of prediction data in the form of the probability distribution; The battery state prediction device according to claim 4 , wherein the average value is compared with previously stored measured values ​​of the amount of gas generated from the battery to determine accuracy of the plurality of machine learning models.

6. generating a plurality of machine learning models that are trained based on the battery data and the characteristics included in the battery data and that predict the amount of gas generation from the battery; applying the battery data to the plurality of machine learning models to obtain a plurality of prediction data predicting the amount of gas generated from the battery; predicting an amount of gas generated from the battery based on the plurality of prediction data; A method of operating a battery state prediction device, comprising:

7. the plurality of machine learning models includes a plurality of deep neural network models; The step of applying the battery data to the plurality of machine learning models to obtain a plurality of prediction data predicting the amount of gas generation from the battery includes: The method of claim 6 , further comprising applying the battery data to the plurality of deep neural network models to obtain the plurality of prediction data.

8. generating a plurality of machine learning models for predicting the amount of gas generation from the battery based on the battery data, generating the plurality of machine learning models based on the battery data including cumulatively measured temperature, SOC, and SOH of the battery; The battery data characteristics are: The method for operating the battery state prediction device according to claim 6 or 7, comprising determining at least one of an electrode type, an assembly process, and a separator type of the battery.

9. The step of predicting the amount of gas generation from the battery based on the plurality of prediction data includes: a step of assigning a weight to each of the plurality of prediction data based on a weight assigned based on characteristics of the battery data during a process of training each of the plurality of deep neural network models that generate each of the plurality of prediction data; and inputting the weighted prediction data into an ensemble learning model to generate the prediction data in the form of a probability distribution.

10. The step of predicting the amount of gas generation from the battery based on the plurality of prediction data includes: calculating an average value of the plurality of prediction data in the form of the probability distribution; and comparing the average value with previously stored gas generation measurement values ​​of the battery to determine accuracy of the plurality of machine learning models.

Images