Method for generating DNN model for battery state estimation adopting multi-variant input vector

Abstract

The present invention relates to a method for generating a DNN model for battery state estimation adopting a multi-variant input vector, and to a method for generating a DNN model, the method enabling: generating a plurality of DNN models adopting input vectors formed by a plurality of mutually different time steps, feature values, or combinations thereof through a diversity design; and estimating a state (SOX) or the remaining useful life (RUL) of a battery by applying the mutually different time steps, feature values, or combinations thereof in a composite manner by using the generated plurality of DNN models according to an application field.

Description

A method for creating a DNN model for battery state estimation that accepts multivariate input vectors.

[0001] The present invention relates to a method for generating a DNN model for estimating a battery state that accepts a multivariate input vector, and more particularly, to a method for generating a DNN model that generates a plurality of DNN models that accept input vectors formed by a plurality of different time steps, feature values, or a combination thereof through a diversity design, and uses the plurality of DNN models generated above to apply different time steps, feature values, or a combination thereof in a complex manner according to an application field, thereby estimating the state (SOX) or remaining useful life (RUL) of a battery.

[0002] Batteries are devices that store and transmit electrical energy. Various types of batteries are designed and operated to meet specific applications and performance requirements. These batteries are primarily used in a variety of fields, including portable electronic devices, electric vehicles, and renewable energy storage.

[0003] It is generally known that the structure and design of the cells that make up a battery have a significant impact on its performance. Furthermore, thermal management is crucial, as overheating can lead to safety issues. If heat is not properly controlled, battery performance can deteriorate and safety issues can arise.

[0004] Additionally, a battery's lifespan can be determined by counting the number of charge and discharge cycles, which indicates the battery's usable lifespan. Furthermore, new materials and nanotechnology are helping to increase battery energy density and improve performance.

[0005] These batteries are used in a variety of applications, from home appliances to electric vehicles. Therefore, it is very important to accurately estimate the battery status depending on the application in order to efficiently manage the battery status and improve stability.

[0006] Meanwhile, there are methods that utilize AI to accurately estimate battery condition. When using AI to estimate battery condition, the performance of the AI ​​model can be optimized by applying different time steps depending on the application environment, such as embedded systems or cloud systems.

[0007] In AI models, increasing the timestep size is beneficial for capturing long-term dependencies, but it also increases computational costs, consumes significant memory, and slows learning. Therefore, the optimal timestep size varies depending on the given problem and data. Choosing the appropriate timestep size should take into account factors such as the complexity of the problem, the characteristics of the time-series data, and available computational resources.

[0008] To improve the performance of AI models, it is necessary to apply multiple timestep sizes to find the optimal timestep size that effectively learns long-term dependencies for specific applications while also being computationally and memory-efficient. Here, the timestep size has the same meaning as the timestep interval, time interval, and sequence length.

[0009] Therefore, in the present invention, a plurality of DNN models are generated with a plurality of different time steps and feature values ​​through diversity design, and then the results of estimating the battery state (SOX) or remaining life (RUL) of each of the generated DNN models according to the application field are combined to finally determine a DNN model according to the application field, and thereby estimate the battery state (SOX) or remaining life (RUL).

[0010] Next, we will briefly explain the prior art existing in the technical field of the present invention, and then describe the technical details that the present invention seeks to achieve differently from the prior art.

[0011] First, Korean Patent Publication No. 2023-0166196 (December 7, 2023) relates to a battery data preprocessing method and system for battery condition diagnosis, and a battery condition prediction system. It describes normalizing and standardizing battery data and dividing the data into specific intervals (e.g., time and cycle, etc.) and inputting them into a neural network. However, it does not intend to generate multiple DNN models with multiple different time step sizes to determine a DNN model according to the application field.

[0012] In addition, Korean Patent No. 2544179 (June 12, 2023) relates to a method and device for determining a kernel size for changing the structure of an artificial intelligence model. In order to reduce the weight of an artificial intelligence model, a search model is used to determine the kernel size for each layer, and a training model is trained based on the determined kernel size. This does not intend to determine a DNN model according to an application field by generating multiple DNN models with multiple different time step sizes.

[0013] That is, the above-mentioned cited inventions do not describe, suggest, or imply any technical features of generating multiple DNN models that accept input vectors formed by multiple different time steps, feature values, or combinations thereof through the diversity design presented in the present invention, and using the multiple DNN models generated above to apply different time steps, feature values, or combinations thereof in a complex manner according to the application field to estimate the state (SOX) or remaining life (RUL) of the battery.

[0014] The present invention was created to solve the above problems, and its purpose is to provide a method for generating a DNN model for battery state estimation that accepts a multivariate input vector and a learning system thereof.

[0015] In addition, the present invention aims to provide a method for generating a DNN model for battery state estimation, in which a plurality of different time steps and feature values ​​are applied in a complex manner by generating a DNN model with a plurality of different time steps and feature values ​​through a diversity design and combining the results of the plurality of generated DNN models.

[0016] In addition, the present invention aims to create a DNN model capable of estimating the state (SOX) or remaining life (RUL) of a battery by comprehensively reflecting DNN models of various time steps and feature values ​​by using DNN models created with multiple different time steps and feature values ​​depending on the application field.

[0017] In addition, the present invention aims to provide an estimation system that estimates the state (SOX) or remaining life (RUL) of a battery through a DNN model for estimating the state of a battery in which various time steps and feature value DNN models are comprehensively reflected.

[0018] A method for generating a DNN model for estimating a battery state that accepts a multivariate input vector according to one embodiment of the present invention includes a DNN model generating step of generating a plurality of different DNN models with a plurality of different time steps, feature values, or a combination thereof; and a step of delay-integrating and outputting learning results of the generated plurality of different DNN models according to an application field; and is characterized in that a DNN model for estimating a battery state is generated to which various time steps, feature values, or a combination thereof are applied in a complex manner according to the application field.

[0019] The above-described plurality of different DNN models are characterized in that they are configured to estimate the battery state (SOX) or remaining useful life (RUL) including the SOC, SOH, SOP, SOE, SOF or a combination thereof of the battery.

[0020] The above DNN model generation step includes a hyperparameter optimization step of optimizing hyperparameters for each of the plurality of different DNN models trained with the plurality of different time steps, feature values, or a combination thereof; wherein the hyperparameters include a learning rate, a number of epochs, a batch size, a number of hidden layers, a number of neurons in each hidden layer, weight initialization, an activation function, a dropout ratio, regularization, or a combination thereof.

[0021] The above-described delayed integration processing is characterized by including selecting results of the plurality of different DNN models, delay-aligning other results by applying a smaller time step to the result of the largest time step among the selected results, comparing the aligned results with each other, and selecting and outputting one of the maximum value, minimum value, average value, or median value through the compared results.

[0022] The method for generating a DNN model for estimating a battery state that accepts the above multivariate input vector is characterized by generating a DNN model with multiple different time steps and feature values ​​through diversity design, combining the results of the multiple generated DNN models to generate a DNN model for estimating a battery state in which multiple time steps and feature values ​​are comprehensively reflected, thereby generating a safe DNN model through redundancy design.

[0023] The DNN model with the above time step of 1 estimates SOX or RUL immediately, and as the above time step becomes larger than 1, the DNN model requires interpolation and deeper layers, and when a gradient vanishing problem occurs in the backpropagation process due to the deep layers, the gradient vanishing problem is solved by selecting one of an activation function, weight initialization, use of memory cells, or an optimization algorithm (AdaGrad, RMSprop, Adam).

[0024] In addition, the application field is either an embedded system or a cloud server, and the embedded system applies the time step smaller than the cloud server, and as a result, the embedded system estimates the SOX or RUL faster than the cloud server.

[0025] Meanwhile, a device for generating a DNN model for estimating a battery state that accepts a polygonal input vector according to another embodiment of the present invention is characterized by including: a memory storing a program configured to generate the DNN model for estimating a battery state according to a method for generating a DNN model for estimating a battery state that accepts a polygonal input vector according to an embodiment of the present invention; and a processor configured to execute the program stored in the memory.

[0026] Meanwhile, a method for estimating a battery state according to a DNN model for estimating a battery state that accepts a polyvariate input vector according to another embodiment of the present invention is characterized by including estimating the state of the battery by inputting a polyvariate input vector extracted from measurement data measured from the battery into the DNN model for estimating a battery state that is generated according to a method for generating a DNN model for estimating a battery state that accepts a polyvariate input vector according to an embodiment of the present invention.

[0027] Meanwhile, a battery state estimation device according to a DNN model for estimating a battery state that accepts a polyvariate input vector according to another embodiment of the present invention is characterized by including: a memory storing a program configured to estimate the state of the battery by inputting a polyvariate input vector extracted from measurement data measured from the battery into the DNN model for estimating a battery state generated according to a method for generating a DNN model for estimating a battery state that accepts a polyvariate input vector according to an embodiment of the present invention; and a processor configured to execute the program stored in the memory.

[0028] As described above, the present invention has the effect of generating a DNN model with multiple different time steps and feature values ​​through diversity design, and combining the results of the multiple DNN models generated above to generate a DNN model for estimating a battery state in which various time steps and feature values ​​are comprehensively reflected.

[0029] In addition, the present invention has the effect of enabling learning and estimation more suitable for the application field by estimating the state (SOX) or remaining life (RUL) of the battery through a DNN model for estimating the state of the battery in which various time steps and feature values ​​of the DNN model are comprehensively reflected depending on the application field.

[0030] FIG. 1 is a diagram for explaining the meaning of a time step in a method for generating a DNN model for battery state estimation that accepts a multivariate input vector according to one embodiment of the present invention.

[0031] FIG. 2 is a diagram illustrating a learning network that inputs a DNN model by configuring an input vector with a specific time step and feature value and inputting it to the DNN model, and outputs SOX or RUL from the DNN model, in a method for generating a DNN model for estimating a battery state that accepts a multivariate input vector according to one embodiment of the present invention.

[0032] FIG. 3 is a drawing illustrating a method for generating a DNN model for battery state estimation that accepts a polygonal input vector according to one embodiment of the present invention, in which the results of a plurality of DNN models having polygonal input vectors are finally delayed and integrated to obtain results according to an application field.

[0033] FIG. 4 is a diagram illustrating a method for generating a DNN model for estimating a battery state that accepts a multivariate input vector according to another embodiment of the present invention, in which delayed integration processing is performed according to a specific application field to obtain a result according to the application field.

[0034] FIG. 5 is a block diagram showing the configuration of a DNN model generation device for battery state estimation that accepts a multivariate input vector according to one embodiment of the present invention.

[0035] FIG. 6 is a flowchart illustrating a method for generating a DNN model for battery state estimation that accepts a multivariate input vector according to one embodiment of the present invention.

[0036] FIG. 7 is a diagram illustrating a configuration of a DNN model generation device for battery state estimation and a battery state estimation device that accepts a multivariate input vector according to one embodiment of the present invention.

[0037] [Description of symbols] 100: DNN model generation device; 110: Measurement data receiving unit; 120: Feature value extraction unit; 130: Time step determination unit; 140: Learning data composition unit; 150: DNN learning unit; 151: Learning data application unit; 152: Learning result evaluation unit; 153: Weight update unit; 160: Learning result processing unit; 200: Battery status estimation device.

[0038] Hereinafter, with reference to the attached drawings, a preferred embodiment of a method for generating a DNN model for battery state estimation that accepts a multivariate input vector of the present invention will be described in detail. The same reference numerals presented in each drawing represent the same elements. In addition, specific structural and functional descriptions of the embodiments of the present invention are merely illustrative for the purpose of explaining the embodiments according to the present invention, and unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by a person of ordinary skill in the art to which the present invention pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning they have in the context of the related technology, and it is preferable not to interpret them in an ideal or excessively formal sense unless explicitly defined in this specification.

[0039] FIG. 1 is a diagram for explaining the meaning of a time step in a method for generating a DNN model for battery state estimation that accepts a multivariate input vector according to one embodiment of the present invention.

[0040] As illustrated in FIG. 1, a method for generating a DNN model for estimating a battery state that accepts a multivariate input vector according to an embodiment of the present invention collects current (I), voltage (V), temperature (T), and changes in charge amount (Qc, Qd) according to charging and discharging as feature values, calculates a target value through the feature values, and configures some or all of the feature values ​​as learning data according to the size of a time step, and then trains the DNN.

[0041] The above training data is in the form of an input vector composed of feature values ​​with a predetermined time step. The input vector is labeled with the target value, and during the DNN training process, learning is performed through backpropagation to minimize the error by comparing the learning result with the target value.

[0042] Here, the input vector format of the DNN is determined by the number of time steps and feature values. If the time step is 1 and the feature values ​​are voltage (V), current (I), and temperature (T), the input vector format is 1x3. If the time step is 3, the structure is such that three 1x3 input vectors are input consecutively. In other words, a 3x3 input vector is input as the training data of the DNN.

[0043] Furthermore, the accuracy of battery state estimation will vary depending on the timestep size. For example, if the timestep is 1, the DNN model estimates SOX or RUL at a single instant (timestep). Therefore, if the timestep size is small, it is difficult to detect patterns. On the other hand, larger timesteps generally require deeper layers, but these deeper layers cause the vanishing gradient problem during backpropagation. This vanishing gradient problem must be addressed by selecting an activation function, weight initialization, memory cell use, or an optimization algorithm.

[0044] Optimization algorithms used here include AdaGrad (Adaptive Gradient Algorithm), RMSprop (Root Mean Square Propagation), and Adam (Adaptive Moment Estimation). These are all based on gradient descent and are well-known techniques that dynamically adjust the learning rate to facilitate efficient learning. Therefore, a detailed description thereof will be omitted in this invention.

[0045] Also, the charge and discharge amount (Qc, Qd) is expressed as capacity (Ah), and its range is exemplified as changing from approximately 0 to 1. Also, the voltage continuously increases when charging occurs and decreases as discharge occurs. The current is (+) when charging and changes to (-) when discharging begins. Also, the temperature (T) increases when charging, decreases when charging is complete, and increases again when discharging.

[0046] The time step size can vary from 1 to hundreds, and as the time step increases, data can be added through interpolation in the middle. The interpolation can be performed by averaging feature values ​​from adjacent time steps spaced apart by the time step size, interpolating the median value, or padding with a specific value (0, 1, -1, etc., as needed).

[0047] FIG. 2 is a diagram illustrating a learning network that inputs a DNN model by configuring an input vector with a specific time step and feature value and inputting it to the DNN model, and outputs SOX or RUL from the DNN model, in a method for generating a DNN model for estimating a battery state that accepts a multivariate input vector according to one embodiment of the present invention.

[0048] As illustrated in FIG. 2, each DNN according to the present invention is a deep neural network (DNN) composed of various input sides, hidden layers, and output layers depending on the shape of the input vector.

[0049] The above input vector refers to a vector input through the input layer of a deep neural network (DNN) according to the number of time steps and feature values. That is, when the input format is (1, 3), a 1x3 input vector is input through the input layer of the DNN and trained.

[0050] The input vector entered in this manner has a target value, and if the learning result of training the DNN with the input vector has an error with the target value, the learning parameters are modified through backpropagation to minimize the error. The learning parameters are also called weights.

[0051] The estimation result of the above DNN becomes SOX (state of x) or RUL (remaining useful lifetime). SOX and RUL represent the state and remaining useful life of the battery, respectively. In the present invention, SOX and RUL are collectively called battery states, but it is more preferable to specifically represent the state of the actual battery by distinguishing between SOX and RUL. That is, SOC (state of charge), SOH (state of health), SOP (state of power), SOE (state of energy), and SOF (state of function) belonging to SOX each have different target values, and RUL also has a different target value from SOX.

[0052] The above SOC indicates the state of charge of the battery (e.g., 0: fully discharged, 1: completely discharged), SOH indicates the degree to which the capacity or performance of the battery is maintained, SOP indicates the output capability or output status (energy supply rate) of the battery, SOE indicates the energy storage status of the battery (total amount of stored energy), and SOF indicates the functional status of the battery (normal, abnormal, etc.).

[0053] Using the structure of these individual DNNs, a DNN model for estimating a battery state that accepts a multivariate input vector according to the present invention is described below, which comprises multiple DNNs.

[0054] FIG. 3 is a drawing illustrating a method for generating a DNN model for battery state estimation that accepts a polygonal input vector according to one embodiment of the present invention, in which the results of a plurality of DNN models having polygonal input vectors are finally delayed and integrated to obtain results according to an application field.

[0055] As illustrated in FIG. 3, the DNN model for estimating a battery state that accepts a multivariate input vector according to the present invention is configured to select all or part of the SOX and RUL output from each DNN trained using input vectors composed of different time steps, feature values, or a combination thereof for a plurality of DNN models having a plurality of different input vector forms, sort them according to the time step, and then compare them, and select and output one of the maximum value, minimum value, average value, and median value from the result of the comparison.

[0056] Here, we train the DNN and perform hyperparameter optimization during the DNN design process to extract the optimal time step size.

[0057] That is, in the process of generating each of the above DNN models, a hyperparameter optimization process is performed to optimize hyperparameters for each of the multiple different DNN models trained with the multiple different time steps, feature values, or a combination thereof. Here, the hyperparameters include a learning rate, number of epochs, batch size, number of hidden layers, number of neurons in each hidden layer, weight initialization, activation function, dropout ratio, regularization, or a combination thereof.

[0058] By adjusting the learning rate, the DNN model determines how many weight updates it will make at each training step. The number of epochs indicates the number of times the entire input data set passes through the DNN model at once. The batch size indicates the number of data samples fed into the DNN model at once. Furthermore, the number of hidden layers and the number of neurons determine the structure of the deep neural network.

[0059] Additionally, the dropout rate is the rate at which some neurons are randomly disabled during training to prevent overfitting. Meanwhile, weight initialization describes how to set the initial weight values ​​of a DNN model. The activation function selects the activation function used in the hidden layer.

[0060] The above optimization algorithm refers to the optimization algorithm used for weight updates, and the above normalization includes data normalization and model regularization. Data normalization is a process that adjusts the range of input data to assist in the learning of the DNN model. For example, it stabilizes learning by adjusting or standardizing the scale of input features to a certain range. In addition, model regularization is a technique that controls the complexity of the DNN model to prevent overfitting and improve generalization performance. It can simplify the model by imposing penalties on the model's weights, making some of them zero or keeping them small.

[0061] Meanwhile, the present invention selects the output values ​​of each DNN, sorts them according to time steps, and finally compares the results to select and output one of the minimum, maximum, average, or median values. This process is called delayed integration processing.

[0062] That is, the delayed integration processing is characterized by including selecting results of the plurality of different DNN models, delay-sorting other results by applying a smaller time step to the result of the largest time step among the selected results, comparing the sorted results with each other, and selecting and outputting one of the maximum value, minimum value, average value, or median value through the compared results.

[0063] For example, if the input vector has the form of time steps 1, 10, and 100, the result of the DNN with time step 1 will be output instantly, and the result of the DNN with time step 10 will be output as a stable result after 10 time steps have passed, so the results of the DNN with time steps 1 and 10 are delayed and sorted to the result of 100, which has the largest time step size, and then the results are integrated and processed.

[0064] In particular, a safe design is required in an automated system, and the method according to the present invention generates a DNN model with multiple different time steps and feature values ​​through diversity design, and combines the results of the multiple DNN models generated to generate a DNN model for estimating a battery state in which multiple time steps and feature values ​​are comprehensively reflected, thereby exhibiting the effect of generating a safe DNN model through redundancy design.

[0065] That is, the present invention estimates the state (SOX) or remaining life (RUL) of a battery through a DNN model for battery state estimation in which various time steps and feature value DNN models are comprehensively reflected depending on the application field, thereby enabling learning and estimation that is more suitable and safe for the application field.

[0066] In addition, the DNN model for estimating battery status according to the present invention can be used to estimate not only SOX but also RUL.

[0067] FIG. 4 is a diagram illustrating a method for generating a DNN model for estimating a battery state that accepts a multivariate input vector according to another embodiment of the present invention, in which delayed integration processing is performed according to a specific application field to obtain a result according to the application field.

[0068] As illustrated in FIG. 4, a DNN model for estimating a battery state that accepts a multivariate input vector according to another embodiment of the present invention can selectively and comprehensively utilize the results of estimating the state of the battery depending on the application field.

[0069] For example, in the case of embedded systems and cloud systems, the application areas are different, so there's no need to use the same DNN results. For example, in embedded systems, it's more appropriate to apply DNN results with smaller time steps. This is because DNNs used in embedded systems are less complex than those with larger time steps. Furthermore, DNNs used in embedded systems need to respond quickly and output results.

[0070] Additionally, there is a risk that data may not be transmitted properly during transmission from the embedded system to the cloud. In this case, a DNN trained with a large time step size in the cloud may perform worse than a DNN trained with a small time step size in the embedded system. Therefore, it is necessary to compare the results of multiple DNNs to produce more stable and robust results.

[0071] Additionally, DNNs trained with small time steps in the cloud may produce better results than DNNs trained with large time steps due to data corruption issues.

[0072] Meanwhile, the DNN model according to the present invention can be trained with a combination of various time sequences. That is, multiple DNNs can be trained with time step sizes of (1, 2, 3, 4, 5, ...), or (1, 3, 5, 7, ...), or (2, 4, 6, 8, ...), or (1, 5, 10, ...).

[0073] If the timestep size is greater than 1, it is possible to expand the data using interpolation. Furthermore, raw data with a timestep size of 1000 can be converted to data with a timestep size of 100 and used.

[0074] FIG. 5 is a block diagram showing the configuration of a DNN model generation device for battery state estimation that accepts a multivariate input vector according to one embodiment of the present invention.

[0075] As illustrated in FIG. 5, a DNN model generation device (100) for battery state estimation that accepts a multivariate input vector according to one embodiment of the present invention is configured to include a measurement data receiving unit (110), a feature value extraction unit (120), a time step determination unit (130), a learning data configuration unit (140), a DNN learning unit (150), and a delay integration processing unit (160).

[0076] The above measurement data receiving unit (110) converts the battery measurement values ​​received in real time into a specific format (bit count, integer, floating point, double, character, etc.) so that they can be used in the DNN, thereby creating measurement data. If the measurement data is previously stored in a database, the measurement data can be received from the database and used immediately.

[0077] The above-described feature value extraction unit (120) selects a feature value to be used in the DNN from the received measurement data, or, if a new feature value must be generated from the received specific data, extracts a feature value including the generated feature value. The new feature value includes, for example, statistical values ​​(average value, maximum value, minimum value, median value, etc.) for the measurement value in the measurement data or feature values ​​extracted (cycle count, etc.) from the measurement value.

[0078] The above time step determination unit (130) is configured to determine the time step of the feature value together with the extracted feature value. The DNN model according to the present invention can be trained with a combination of various time sequences. That is, a plurality of DNNs can be trained with a time step size of (1, 2, 3, 4, 5, ...), or a size of (1, 3, 5, 7, ...), or a size of (2, 4, 6, 8, ...), or a size such as (1, 5, 10, ...). In addition, the size of the time step can be determined to an arbitrary size depending on the application field. That is, a predetermined time step size can be determined in advance according to each application field, stored in memory, and then the size of the time step to be currently applied can be determined by referring to it as needed.

[0079] Learning data is constructed based on the extracted feature values ​​and time step size, and the learning data is labeled to construct the learning data.

[0080] The above learning data configuration unit (140) configures an input vector of a specific form according to the feature values ​​extracted by the feature value extraction unit (120) and the time step determined by the time step determination unit (130). That is, the form of the input vector is determined by the size of the time step and the number of feature values. For example, if the size of the time step is 3 and the number of feature values ​​is 3, a 3x3 input vector is configured as the learning data of the DNN. The learning data has a target value, and the target value is a label of the learning data.

[0081] The above DNN learning unit (150) is configured to include a learning data application unit (151), a learning result evaluation unit (152), and a weight update unit (153).

[0082] The above learning data application unit (151) is configured to perform learning by inputting the learning data as an input vector into the DNN. The learning is a process of determining learning parameters determined by passing through a learning network that starts from the input layer, passes through multiple hidden layers, and reaches the output layer.

[0083] The above learning result evaluation unit (152) is configured to evaluate the learning result extracted through the learning so that the error between the learning result and the target value is minimized.

[0084] If the above-mentioned evaluation result does not converge to the desired minimum error, backpropagation is used to repeatedly learn to reduce the error. In other words, the learning data application unit (151) and the learning result evaluation unit (152) are repeatedly performed until the evaluation result converges to the desired level. During this iteration, the learning parameters will continuously change.

[0085] The above weight update unit (153) is configured to update the learning parameters for a given learning data when the learning result for that learning data is satisfied and the learning process is terminated. Once the learning parameters are updated, the next learning parameter is input and learning continues through the learning data application unit (151), learning result evaluation unit (152), and weight update unit (153).

[0086] A plurality of DNNs are trained using a plurality of training data generated according to the results of the above-described feature value extraction unit (120) and the above-described time step determination unit (130). Therefore, in the DNN model generation device according to one embodiment of the present invention, there are as many DNN learning units (150) as the number of DNN models to be generated. In other words, since each DNN model outputs a battery state estimation result, it is a structure in which a plurality of battery state estimation results are derived.

[0087] The above delay integration processing unit (160) is configured to provide a battery state estimation result in which the results of the plurality of DNN models are applied in a composite manner according to the application field.

[0088] The above delay integration processing unit (160) selects the results of the plurality of different DNN models, delay-aligns other results by applying a smaller time step to the result of the largest time step among the selected results, compares the aligned results with each other, and selects and outputs one of the maximum value, minimum value, average value, or median value through the compared results.

[0089] The purpose is to create multiple DNN models that accept multiple different input vectors, and to apply the learning results of each of the multiple DNN models in combination according to the selection of the application field to provide an estimation result on the state of the battery suitable for the application field.

[0090] By comparing multiple DNN learning results, it is preset to output one of the minimum, maximum, average, or median values ​​as the estimated result for the battery status, depending on the application field selected.

[0091] FIG. 6 is a flowchart illustrating a method for generating a DNN model for battery state estimation that accepts a multivariate input vector according to one embodiment of the present invention.

[0092] As illustrated in FIG. 6, a method for generating a DNN model for battery state estimation according to one embodiment of the present invention first receives battery measurement data, extracts feature values, and determines each feature value to be applied to each of the multiple DNNs (S110). Furthermore, the size of each time step to be applied to each of the multiple DNNs is determined (S120).

[0093] Next, various types of learning data are configured with the size of the time step and the feature values ​​to be applied to each of the multiple DNNs, and the target value of the learning data is determined (S130).

[0094] Next, the DNN is trained using the generated training data as input vectors for each of the plurality of DNNs (S140). This includes a hyperparameter optimization step for optimizing hyperparameters for each of the plurality of different DNN models trained using the plurality of different time steps, feature values, or combinations thereof.

[0095] The above hyperparameters include learning rate, number of epochs, batch size, number of hidden layers, number of neurons in each hidden layer, weight initialization, activation function, dropout ratio, regularization, or a combination thereof.

[0096] The learning results of each DNN are evaluated by comparing them with the target values ​​of the corresponding learning data (S150).

[0097] Based on the above evaluation results, steps S140 and S150 are repeated until the difference (error) between the learning result and the target value satisfies a predetermined minimum value (S160). This process is a backpropagation process, and the learning parameters of the DNN are determined so that the error converges to a minimum value. Once the error converges to a minimum value (S160), the learning parameters (weights) of the DNN are updated, indicating that learning has been completed for the corresponding input vector (S170).

[0098] Through this process, the process of training a single DNN on a single training data set is completed. Since the present invention requires that this training process be performed for each DNN, steps S140 to S170 are repeated until all DNNs complete training (S180).

[0099] In this way, after all DNNs are trained in the form of different input vectors, a step is performed to select, compare, or combine the training results of the multiple DNNs to output one of the minimum value, maximum value, average value, and median value (S190).

[0100] That is, a delayed integration process is performed by selecting the results of the above-described multiple different DNN models, delay-sorting other results by applying a smaller time step to the result of the largest time step among the selected results, comparing the sorted results with each other, and selecting and outputting one of the maximum value, minimum value, average value, or median value through the compared results.

[0101] Ultimately, a method for generating a DNN model for estimating a battery state that accepts a multivariate input vector according to an embodiment of the present invention includes a DNN model generation step of generating a plurality of different DNN models using a plurality of different time steps, feature values, or a combination thereof, and a step of processing and outputting learning results of the generated plurality of different DNN models according to an application field. That is, a DNN model for estimating a battery state to which various time steps, feature values, or a combination thereof are applied in a complex manner according to the application field is generated.

[0102] The above multiple different DNN models are configured to estimate the battery state of charge (SOX) or remaining useful life (RUL) including SOC, SOH, SOP, SOE, SOF or a combination thereof of the battery.

[0103] FIG. 7 is a diagram illustrating a configuration of a DNN model generation device for battery state estimation and a battery state estimation device that accepts a multivariate input vector according to one embodiment of the present invention.

[0104] As illustrated in FIG. 7, a device (100) for generating a DNN model for estimating a battery state that accepts a polygonal input vector according to an embodiment of the present invention includes a memory (2000) storing a program configured to generate a DNN model for estimating a battery state that accepts a polygonal input vector according to the present invention, and a processor (1000) configured to execute the program stored in the memory.

[0105] In addition, it may further include a user interface (3000), a database interface (4000), a network interface (5000), and a web server (6000). The above components are connected to each other by a bus line.

[0106] The user interface (3000) above is a passage through which a user accesses the device via a user interface (UI) or a graphical user interface (GUI), and the database interface (4000) is responsible for receiving measurement values ​​or measurement data from a local or network database, or for storing the generated DNN model in the database. The network interface (5000) serves as an interface through which the DNN model generation device (100) for battery state estimation or the battery state estimation device (200) according to the present invention accesses a network to receive measurement values ​​from a battery, or transmits or receives a DNN model to or from the network. In addition, the web server (6000) is a function that allows easy web-based access when an external terminal accesses the DNN model generation device (100) for battery state estimation or the battery state estimation device (200) according to the present invention.

[0107] Meanwhile, a method for estimating a state of a battery using a DNN model for estimating a state of a battery that accepts a polygonal input vector according to another embodiment of the present invention is configured to estimate the state of the battery by inputting a feature value from measurement data measured from the battery as an input vector into the DNN model for estimating a state of the battery that accepts the polygonal input vector.

[0108] Meanwhile, a battery state estimation device for estimating a state of a battery using a DNN model for estimating a state of a battery that accepts a polygonal input vector according to another embodiment of the present invention is configured to include a memory (2000) storing a program configured to estimate a state or remaining life of a battery using a DNN model for estimating a state of a battery that accepts a polygonal input vector according to the present invention, and a processor (1000) configured to execute the program stored in the memory.

[0109] As described above, the present invention has the effect of generating a DNN model with multiple different time steps and feature values ​​through diversity design, and combining the results of the multiple DNN models generated above to generate a DNN model for estimating a battery state in which various time steps and feature values ​​are comprehensively reflected.

[0110] In addition, the present invention has the effect of enabling learning and estimation more suitable for the application field by estimating the state (SOX) or remaining life (RUL) of the battery through a DNN model for estimating the state of the battery in which various time steps and feature values ​​of the DNN model are comprehensively reflected depending on the application field.

[0111] In addition, although the preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above, and various modifications can be implemented by a person having ordinary skill in the art to which the invention pertains without departing from the gist of the present invention claimed in the claims, and such modifications should not be understood individually from the technical idea or prospect of the present invention.

[0112] As described above, the present invention can generate a DNN model for estimating a battery state by generating a DNN model with multiple different time steps and feature values ​​through a diversity design, and by combining the results of the multiple DNN models generated above, generating a DNN model for estimating a battery state in which various time steps and feature values ​​are comprehensively reflected, and further, by estimating the state (SOX) or remaining useful life (RUL) of the battery through a DNN model for estimating a battery state in which various time steps and feature values ​​are comprehensively reflected according to the application field, thereby enabling learning and estimation more suitable for the application field, and thus has industrial applicability.

Claims

1. A DNN model generation step for generating multiple different DNN models with multiple different time steps, feature values, or a combination thereof; and A step of delay-integrating and outputting the learning results of the multiple different DNN models generated above according to the application field; A method for generating a DNN model for battery state estimation that accepts a multivariate input vector, characterized in that it generates a DNN model for battery state estimation to which various time steps, feature values, or a combination thereof are applied in a complex manner according to the above application field.

2. In claim 1, A method for generating a DNN model for estimating a battery state that accepts a multivariate input vector, wherein the plurality of different DNN models are configured to estimate a battery state (SOX) or remaining useful life (RUL) including SOC, SOH, SOP, SOE, SOF or a combination thereof of the battery.

3. In claim 1, The above DNN model generation step includes a hyperparameter optimization step of optimizing hyperparameters for each of the plurality of different DNN models trained with the plurality of different time steps, feature values, or combinations thereof; A method for generating a DNN model for battery state estimation that accepts a multivariate input vector, wherein the hyperparameters include a learning rate, number of epochs, batch size, number of hidden layers, number of neurons in each hidden layer, weight initialization, activation function, dropout ratio, regularization, or a combination thereof.

4. In claim 1, The above delayed integration processing is, A method for generating a DNN model for estimating a battery state that accepts a multivariate input vector, characterized in that it comprises selecting the results of the plurality of different DNN models, delay-sorting other results by applying a smaller time step to the result of the largest time step among the selected results, comparing the sorted results with each other, and selecting and outputting one of the maximum value, minimum value, average value, or median value through the compared results.

5. In claim 1, The above method, A method for generating a DNN model for estimating a battery state that accepts a multivariate input vector, characterized in that a DNN model is generated with multiple different time steps and feature values through diversity design, and the results of the multiple DNN models generated are combined to generate a DNN model for estimating a battery state in which multiple time steps and feature values are comprehensively reflected, thereby generating a safe DNN model through redundancy design.

6. In claim 1, The above DNN model with a time step of 1 immediately estimates the battery state (SOX) or remaining life (RUL). As the above time step becomes larger than 1, the DNN model requires interpolation and deeper layers. A method for generating a DNN model for battery state estimation that accepts a multivariate input vector, characterized in that when a gradient vanishing problem occurs in the backpropagation process due to the deep layer, the gradient vanishing problem is solved by selecting one of an activation function, weight initialization, use of memory cells, or an optimization algorithm (AdaGrad, RMSprop, Adam).

7. In claim 1, The above application fields are, Either an embedded system or a cloud server, A method for generating a DNN model for estimating a battery state that accepts a multivariate input vector, wherein the embedded system applies a smaller time step than the cloud server, and as a result, the embedded system estimates the battery state (SOX) or remaining life (RUL) faster than the cloud server.

8. A memory storing a program configured to generate a DNN model for estimating a battery state according to a method for generating a DNN model for estimating a battery state that accepts a multivariate input vector according to any one of claims 1 to 7; and A device for generating a DNN model for estimating a battery state that accepts a multivariate input vector, characterized in that it comprises a processor configured to execute a program stored in the above memory.

9. A method for estimating a battery state, characterized in that it comprises estimating the state of the battery by inputting a polyvariate input vector extracted from measurement data measured from the battery into the DNN model for estimating the battery state generated according to the method for generating a DNN model for estimating the battery state that accepts a polyvariate input vector according to any one of claims 1 to 7.

10. A memory storing a program configured to estimate the state of the battery by inputting a polyvariate input vector extracted from measurement data measured from the battery into the DNN model for estimating the state of the battery generated according to the method for generating a DNN model for estimating the state of the battery that accepts a polyvariate input vector according to any one of claims 1 to 6; and A battery state estimation device, characterized in that it comprises a processor configured to execute a program stored in the above memory.

Images