Method for generating learning model for estimating battery state by applying different types of dnns according to battery life cycle

Abstract

The present invention relates to a method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle and, to a method for generating a learning model, which can accurately estimate a battery state specialized according to each life cycle period by determining to which period of a battery life cycle feature data collected from a battery belongs and classifying the feature data, configuring training data to be input into different types of DNNs specialized in respective periods, by using the classified battery feature data, and allowing each of the different types of DNNs to learn the training data.

Description

A method for creating a learning model for battery state estimation by applying different types of DNNs according to the battery life cycle.

[0001] The present invention relates to a method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, and more specifically, to a method for generating a learning model capable of accurately estimating a battery state by specializing it according to a life cycle section, by determining and classifying which section of the battery life cycle the feature data collected from the battery belongs to, and by configuring learning data to be input to different types of DNNs specialized for each section using the classified battery feature data, and having each of the DNNs learn.

[0002] Batteries that can be used for long periods of time by repeatedly charging and discharging (i.e., secondary batteries) are used as an essential power source for various application devices such as electric vehicles and smartphones due to their advantage of being able to produce electricity whenever needed without continuous connection to a commercial power source.

[0003] Accurately estimating the battery's condition is crucial for efficient battery management and improving the stability of various battery-powered application devices. Battery condition can include state of charge (SOC), state of health (SOH), and state of power (SOP).

[0004] Meanwhile, research is currently being conducted on a device that estimates the condition of a battery using artificial intelligence technology.

[0005] Conventional techniques for estimating the state of a battery using artificial intelligence create a learning model by learning learning data generated based on the characteristic values ​​of the battery, and then input the characteristic values ​​of the actual battery into the created learning model to estimate the state of the battery.

[0006] Typically, batteries deteriorate over time. For example, battery capacity deteriorates and decreases over time. However, in the initial cycle of a battery's life cycle (overall life cycle), capacity tends to increase regardless of time (usage time).

[0007] These properties are due to the charge stored in the negative electrode region connected to the positive electrode of the battery.

[0008] This shows that the initial cycle period in the battery life cycle has a low correlation with time, while the period after the initial cycle period has a high correlation with time.

[0009] Therefore, when the battery state is estimated through simple learning, there is a problem that the battery state may be estimated higher than the actual battery state in the initial cycle period.

[0010] Accordingly, the present invention proposes a method of dividing feature data, which is data of feature values ​​measured from a battery, into predetermined cycle sections according to the battery life cycle and classifying them to form learning data for each cycle section, and using the learning data for each cycle section, training different types of DNNs specialized for each cycle section, thereby creating a learning model for estimating a battery state specialized for each cycle section.

[0011] That is, the present invention enables accurate estimation of the battery state by using a learning model for estimating the battery state specialized for each cycle section in the battery life cycle.

[0012] 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.

[0013] First, Korean Patent No. 2215450 (February 5, 2021) relates to a device and method for learning and estimating battery status information. The device and method learn and estimate battery status information by collecting battery information including any one of a voltage signal, a current signal, and a temperature signal, inputting the interval information of a predetermined range and the reference information of the battery into a learner to learn parameters.

[0014] That is, Korean Patent No. 2215450 simply learns section information based on the voltage signal, current signal, and temperature signal of the battery and the reference information of the battery set in advance, and does not classify the characteristic data of the battery proposed in the present invention into which cycle section it belongs in the life cycle of the battery to generate learning data, and does not describe at all a method of generating a learning model for estimating the battery state for each cycle section by training different types of DNNs specialized for each cycle section through each learning data. Therefore, the two inventions have significant differences in their technical configuration, purpose, and effect.

[0015] In addition, Korean Patent No. 2171807 (October 23, 2020) relates to a system for predicting product defects during a process and a method for creating a learning model for predicting defects, which collects process-specific measurement information including voltage, current, and accumulated gas volume, and accumulates the collected process-specific measurement information to create a learning model for predicting defects through a deep neural network (DNN), an artificial neural network (ANN), etc. The system relates to a system for predicting product defects during a process and a method for creating a learning model for predicting defects.

[0016] That is, Korean Patent No. 2171807 creates a learning model using learning data that accumulates measurement information for each process.

[0017] On the other hand, the present invention is for battery state estimation, and it creates specialized learning models for battery estimation for each cycle section of the battery life cycle. Therefore, the two inventions differ significantly in their technical structure, purpose, and effectiveness.

[0018] The present invention was created to solve the above problems, and the purpose of the present invention is to provide a method for generating a learning model for estimating a battery state by applying different types of DNNs according to the battery life cycle, which divides the life cycle of a battery into predetermined cycle sections and generates a learning model for estimating a battery state specialized for each cycle section, and which generates a learning model for estimating a battery state by applying different types of DNNs according to the battery life cycle.

[0019] In addition, the present invention provides a method for classifying battery characteristic data by determining which cycle section of the battery life cycle it belongs to, generating learning data for each section using the classified battery characteristic data, and training a DNN specialized for each cycle section to generate a learning model for battery estimation for each section.

[0020] In addition, the present invention provides a method for generating a learning model for estimating a battery state by using an RNN-based DNN for a cycle section in which the state of a battery changes in relation to time and a CNN-based DNN for a cycle section that is slightly related to time, by using the properties of feature data for a relationship that transforms according to time, space, or a combination thereof.

[0021] In addition, the present invention provides a method for estimating a battery state by inputting characteristic data for characteristic values ​​of an actual battery corresponding to a specific cycle section into a battery state estimation learning model specialized for each cycle section, thereby estimating the battery state.

[0022] A method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle according to an embodiment of the present invention is characterized by including a feature data configuration step of collecting feature values ​​measured from a battery and configuring feature data, a feature data classification step of dividing the configured feature data into predetermined cycle sections according to the battery life cycle and classifying them, a DNN configuration step of configuring a DNN (deep neural network) selected according to an attribute of the feature data classified into the predetermined cycle sections, a learning data configuration step of configuring learning data according to the predetermined cycle sections using the feature data classified into the predetermined cycle sections, and a learning step of generating a learning model for estimating a battery state according to each of the cycle sections by training the DNNs selected as corresponding to each of the cycle sections with the learning data configured according to the cycle sections.

[0023] In addition, the property of the above feature data is characterized in that the feature data is a property of a relationship that changes according to time, space, or a combination thereof.

[0024] In addition, the feature data classification step is characterized in that if the number of cycles included in the feature data is less than a predetermined first critical cycle number and the rate of change in the amount of charge included in the feature data is a positive value, it is classified as a first cycle section, if the number of cycles is within a predetermined critical cycle number range and the rate of change in the amount of charge is within a predetermined range, it is classified as a second cycle section, and if the number of cycles exceeds a predetermined second critical cycle number and the rate of change in the amount of charge is a negative value, it is classified as a third cycle section.

[0025] In addition, the DNN selected to correspond to the first cycle section is a DNN based on a CNN (convolutional neural network), and the DNN selected to correspond to the second and third cycle sections is a DNN based on a RNN (recurrent neural network).

[0026] In addition, according to another embodiment of the present invention, a method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle is characterized by including a feature data configuring step of collecting feature values ​​measured from a battery in an agent and configuring feature data, a feature data classification step of dividing the configured feature data into predetermined cycle sections according to a battery life cycle in the agent and classifying them, a DNN configuring step of configuring a DNN (deep neural network) selected according to an attribute of the feature data classified into predetermined cycle sections in the agent, and a learning data configuring step of configuring learning data according to the predetermined cycle sections using the feature data classified into the predetermined cycle sections in the agent, and a reinforcement learning step of generating a learning model for estimating a battery state according to the cycle sections by performing reinforcement learning on the DNNs selected as corresponding to each of the cycle sections using the learning data configured according to the cycle sections in the agent.

[0027] In addition, the reinforcement learning step includes a reinforcement learning result output step for inputting learning data configured for each cycle section into each DNN selected as a corresponding one for each cycle section in the reinforcement learning process and outputting a reinforcement learning result which is the output of each DNN, and a reward calculation step for calculating a reward for each reinforcement learning result output for each DNN, and when the reward calculated for each DNN exceeds a predetermined threshold value, the reinforcement learning is performed so as to be applied to each DNN to output a reinforcement learning result in which the reward is maximized, and when the reward does not exceed a predetermined threshold value, the reinforcement learning is performed.

[0028] In addition, a device for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle according to an embodiment of the present invention is characterized in that it generates a learning model for estimating a battery state according to the method for generating a learning model for estimating a battery state.

[0029] As described above, the present invention has the effect of estimating the battery state more accurately by dividing the battery life cycle into predetermined cycle sections and generating a learning model for estimating the battery state through a DNN specialized for each cycle section.

[0030] FIG. 1 is a diagram illustrating the life cycle of a battery according to one embodiment of the present invention.

[0031] FIG. 2 is a diagram illustrating an initial cycle section in the life cycle of a battery according to one embodiment of the present invention.

[0032] FIG. 3 is a drawing illustrating a cycle section according to one embodiment of the present invention.

[0033] FIG. 4 is a diagram illustrating a method for generating a learning model for estimating a battery state according to one embodiment of the present invention.

[0034] FIG. 5 is a diagram illustrating in detail a method for generating a learning model for estimating a battery state according to one embodiment of the present invention.

[0035] FIG. 6 is a diagram illustrating a method for generating a learning model for estimating a battery state according to another embodiment of the present invention.

[0036] FIG. 7 is a diagram illustrating in detail a method for generating a learning model for estimating a battery state according to another embodiment of the present invention.

[0037] FIG. 8 is a block diagram showing the configuration of a device that generates a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle according to one embodiment of the present invention.

[0038] FIG. 9 is a block diagram showing the configuration of a device that generates a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle according to another embodiment of the present invention.

[0039] Figure 10 is a flowchart illustrating a procedure for generating a learning model for estimating a battery state according to one embodiment of the present invention.

[0040] Figure 11 is a flowchart illustrating a procedure for generating a learning model for estimating a battery state according to another embodiment of the present invention.

[0041] [Description of symbols] 100, 200: Device for generating a learning model for estimating battery status; 110, 210: Feature value collection unit; 220: Agent; 120, 221: Feature data configuration unit; 222: Classification condition setting unit; 130, 223: Feature data classification unit; 140, 224: DNN configuration unit; 150, 225: Learning data configuration unit; 160: Learning unit; 226: Reward calculation unit; 227: Reinforcement learning unit.

[0042] Hereinafter, with reference to the attached drawings, a preferred embodiment of a method for generating a learning model for estimating a battery state by applying different types of DNNs according to the battery life cycle of the present invention will be described in detail. The same reference numerals in each drawing represent the same components. 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.

[0043] FIG. 1 is a diagram illustrating the life cycle of a battery according to one embodiment of the present invention.

[0044] As illustrated in Figure 1, in the life cycle of a battery, which involves repeated charge / discharge cycles according to one embodiment of the present invention, the battery's capacity generally inevitably declines over time. This change in battery capacity may vary over time, space, or a combination thereof.

[0045] As shown in Figure 1, in the life cycle of a battery undergoing repeated charging and discharging, the capacity increases during the initial cycle, then maintains a certain level before gradually decreasing, and the capacity decrease accelerates near the end of the battery's lifespan. In other words, there is a weak correlation between the initial capacity and the battery's lifespan.

[0046] At this time, the increase in capacity is due to the charge stored in the cathode region. That is, the correlation with time is weak in the initial cycle period, and thereafter, the correlation with time increases.

[0047] FIG. 2 is a diagram illustrating an initial cycle section in the life cycle of a battery according to one embodiment of the present invention.

[0048] As illustrated in FIG. 2, in the life cycle of a battery according to one embodiment of the present invention, when the initial cycle section (e.g., 100 cycles) is expanded, the charge capacity (Qc) (or charge amount) rather maintains an increased state (1.0 or more). Thereafter, the capacity decreases as illustrated in FIG. 1.

[0049] Similarly, during the initial cycle of a battery, the discharge capacity (Qd) (or discharge charge) starts out with an increased (over 1.0). After this, the capacity decreases in the same manner as the capacity change during the charge cycle.

[0050] As shown in Figure 1, in the battery life cycle including charge cycles and discharge cycles, it can be seen that the capacity increases in the initial cycle section, then maintains a certain level and then decreases over time.

[0051] That is, during the charge / discharge life cycle of a battery, a phenomenon can be observed in the initial cycle period that shows a different aspect from the change in the state of charge according to the life of the battery.

[0052] When looking at this phenomenon, when trying to create a learning model for battery state estimation using artificial intelligence, it is closer to a classification problem in the early cycle section with low correlation with time, and closer to a regression or time-series problem in the cycle section with high correlation with time.

[0053] Therefore, in the present invention, the life cycle of a battery is divided into multiple cycle sections, and a learning model for estimating the battery state is created that is specialized for each cycle section that is closer to a classification problem and each cycle section that is closer to a regression problem, thereby enabling the battery state to be estimated very accurately.

[0054] FIG. 3 is a drawing illustrating a cycle section according to one embodiment of the present invention.

[0055] As illustrated in FIG. 3, a device (100) (hereinafter referred to as a learning model generation device) that generates a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle according to one embodiment of the present invention divides the battery life cycle into first to third cycle sections.

[0056] The first cycle section refers to a cycle section in the battery life cycle in which the rate of change of the charge (e.g., Qc, charge amount) among the battery characteristic values ​​is greater than 0 and the number of cycles is less than a predetermined number of cycles (less than a first critical cycle number), the second cycle section refers to a cycle section in which the rate of change of the charge is within a predetermined range and within a predetermined critical cycle number range, and further, the third cycle section refers to a cycle section in which a capacity decrease occurs based on a capacity decrease point (capacity fade point) after the second cycle, and refers to a cycle section in which the rate of change of the charge is less than 0 and exceeds a predetermined third cycle number.

[0057] The rate of change of the above charge amount can be calculated by differentiating the charge amount (charge amount (Qc) or discharge amount (Qd)) with respect to the unit time. Here, the unit time must be determined to be a predetermined time or longer so as not to incorrectly reflect the rate of change due to an instantaneous rise or fall in the charge amount. In other words, an excessively short time should not be set as the unit time.

[0058] In the second cycle section, the predetermined range for the rate of change in the amount of charge can be set as the rate of change (max) for the amount of charge of the battery cell with the largest capacity and the rate of change (min) for the amount of charge of the battery cell with the smallest capacity in the cycle section where there is little change in capacity among the plurality of battery cells constituting the battery.

[0059] In addition, the learning model generation device (100) of the present invention is configured to solve the first cycle section, in which the capacity rather increases, as a classification problem because the correlation with time is weak, and to solve the second and third cycle sections as regression problems because the correlation with time is higher than that of the first cycle section.

[0060] That is, the learning model generation device (100) is configured to generate a learning model for predicting the battery status corresponding to the first cycle section using a CNN-based DNN (DNN 1), and to generate a learning model for predicting the battery status corresponding to the second and third cycle sections using an RNN-based DNN (DNN 2, DNN 3).

[0061] Although only the charge amount (Qc) is described in Fig. 3, the same method is applied to the discharge charge amount (Qd). That is, the charge amount (Q) is differentiated with respect to unit time and the cycle section of the composition is classified according to the rate of change of the charge amount.

[0062] FIG. 4 is a diagram illustrating a method for generating a learning model for estimating a battery state according to one embodiment of the present invention.

[0063] As illustrated in FIG. 4, a learning model generation device (100) according to one embodiment of the present invention divides feature data composed of feature values ​​measured from a battery into predetermined cycle sections in the battery life cycle and classifies them.

[0064] The above characteristic values ​​are composed of voltage (V), current (I), temperature (T), charge (Q), number of cycles, or a combination thereof.

[0065] Configuring the above feature data means converting it into data that the DNN can understand (or read), i.e., a specific code, integer, or binary number, in order to create a learning model for estimating the battery state.

[0066] The above cycle section is composed of the first cycle section to the third cycle section, and the first to third cycle sections have been described with reference to FIG. 2, so a detailed description thereof will be omitted.

[0067] The above learning model generation device (100) configures learning data for each cycle section using the classified feature data, and configures a DNN selected according to the properties of the feature data for each cycle section.

[0068] The above learning data is configured for each cycle section using feature data classified for each cycle section. The learning data includes voltage, current, temperature, charge, or a combination thereof as feature values, and may further include voltage statistics, current statistics, temperature statistics, charge statistics, or a combination thereof. Here, it is preferable that the learning data exclude the cycle count of the feature data. This is because the cycle count is used when classifying the feature data into cycle sections. However, this is not limited to this, and the cycle count may be included when configuring the learning data.

[0069] The above statistical values ​​may include the mean, variance, median, standard deviation, or a combination thereof for each of voltage, current, temperature, and charge.

[0070] The above DNN is configured by selecting a specialized one for each cycle section according to the properties of the classified feature data, and the DNN specialized for the first cycle section (DNN 1) may be configured as a convolutional neural network (CNN), and the DNNs specialized for the second and third cycle sections may be configured as RNN-based DNNs. At this time, the DNN specialized for the second cycle section (DNN 2) may be configured as a bi-directional LSTM DNN, and the DNN specialized for the third cycle section (DNN 3) may be configured as a recurrent neural network (RNN).

[0071] The CNN is configured to process learning data (first learning data) consisting of a sequence of feature data including voltage, current, temperature, charge, or a combination thereof, which constitute the learning data. The sequence of feature data constituting the learning data is input to the input channel of the CNN.

[0072] The above CNN generates a feature map by convolving the learning data input to each channel with a kernel having a predetermined weight corresponding to each channel, subsamples the feature map by average pooling or max pooling, and inputs the subsampled result to a fully connected layer (FCL) to finally output a single value, which is the learning result (i.e., battery status).

[0073] That is, the CNN that has completed learning becomes a learning model for battery estimation specialized for the first cycle section, and the input is a sequence (sequence according to a predetermined time step) of voltage, current, temperature, charge, or a combination thereof, and the output is the battery status.

[0074] Bidirectional LSTM is a RNN-based DNN that can capture dependencies across both past and future time steps. This is achieved by maintaining two hidden states for each time step.

[0075] One of the two hidden states is performed by processing the sequence of feature data constituting the input training data (second training data) in a forward direction (right direction in Fig. 4), and the other is performed by processing the input sequence of feature data constituting the input training data in a backward direction (left direction in Fig. 4). The final output of each time step is calculated based on the hidden state.

[0076] The above bidirectional LSTM uses a sequence of feature data including voltage, current, temperature, charge, or a combination thereof according to the learning data as input, and the output calculated based on the hidden state becomes a series of context vectors.

[0077] The above context vector is a weighted sum of hidden states according to forward and backward processing at a time step, and the weights can be determined by an attention mechanism.

[0078] The context vector is then fed to a fully connected layer (FCL), which is configured to output a single value representing the learning outcome (e.g., battery status) during the learning process.

[0079] That is, the forward LSTM that has completed learning becomes a learning model for battery estimation specialized for the second cycle section, and the input is a sequence of voltage, current, temperature, charge, or a combination of these according to the time step, and the output is the battery status.

[0080] The above RNN is also a type of DNN that is very suitable for prediction related to time series data, and uses a sequence of feature data including voltage, current, temperature, charge, or a combination thereof according to learning data (third learning data) as input.

[0081] The above RNN performs learning by updating the weights of the learning parameters (Whh, Wxh, Why) while updating the hidden states (h0 to ht) at each time step. The final output (learning result) during the learning process consists of a single output representing the battery status.

[0082] That is, the RNN that has completed learning becomes a learning model for battery estimation specialized for the third cycle section, and the input is a sequence of voltage, current, temperature, charge, or a combination thereof according to the time step, and the output is the battery status.

[0083] That is, the DNN specialized for each cycle section is configured to learn by receiving learning data for each cycle section composed of a sequence of voltage, current, temperature, charge, or a combination thereof for each time step.

[0084] The above battery state includes state of charge (SOC), state of health (SOH), state of energy (SOE), state of power (SOP), remaining useful life (RUL), or a combination thereof.

[0085] The above SOC or SOE is information that tells you how long the battery can be used, SOH is information that shows how good the battery's performance is, SOP is information that shows how powerful the battery is, and RUL is information that shows how much life the battery has left.

[0086] Meanwhile, in Fig. 4, DNN 1 is depicted as a CNN, but it is not limited thereto, and various CNN-based DNNs can be used, and DNN 2 and DNN 3 are depicted as a bidirectional LSTM and RNN, but it is not limited thereto, and various RNN-based DNNs such as a transformer, RES-net (residual net), LSTM (ling short term memory), GRU (gated recurrent unit), etc. can be used.

[0087] Below, the process of creating a learning model for estimating battery status through learning is described in detail with reference to Fig. 5.

[0088] FIG. 5 is a diagram illustrating in detail a method for generating a learning model for estimating a battery state according to one embodiment of the present invention.

[0089] As illustrated in FIG. 5, a learning model generation device (100) according to one embodiment of the present invention collects feature values ​​measured from a battery. These feature values ​​may be collected in advance and stored in a database (not shown) or collected from a user application device (electric vehicle). Here, these feature values ​​are collected for learning.

[0090] The above learning model generation device (100) configures the collected feature values ​​into feature data. The feature data is configured by converting the feature values ​​into a specific code or binarizing (or integerizing) them so that the DNN can understand them, in order to generate a learning model for battery state estimation.

[0091] The above learning model generation device (100) divides and classifies feature data into predetermined cycle sections according to the battery life cycle.

[0092] In addition, the above classification is performed using the charge amount and number of cycles included in the feature data, and as described with reference to FIGS. 3 and 4, a detailed description thereof will be omitted.

[0093] That is, the learning model generation device (100) classifies the feature data into cycle sections by determining which cycle section the feature data belongs to among the battery life cycle.

[0094] In addition, the learning model generation device (100) configures a DNN selected according to the properties of feature data corresponding to a predetermined cycle section, and configures learning data for each predetermined section using the classified feature data.

[0095] That is, the learning model generation device (100) selects and configures a DNN according to the properties of the relationship that changes according to time, space, or a combination thereof of the classified feature data.

[0096] At this time, the feature data classified as the first cycle section has the property that the rate of change in the amount of charge has a positive value (i.e., the capacity increases) below a predetermined first critical cycle number according to time, space, or a combination thereof, the feature data classified as the second cycle section has the property that the rate of change in the amount of charge changes within a predetermined range within a predetermined critical cycle range, and the feature data classified as the third cycle section has the property that the rate of change in the amount of charge has a negative value when the predetermined second critical cycle number is exceeded.

[0097] Additionally, the learning model generation device (100) configures some or all of the feature values ​​constituting the classified feature data as learning data, depending on the size of the time step. The time step refers to a predetermined time unit. At this time, the learning data is labeled with the actual battery status.

[0098] That is, the learning model generation device (100) organizes feature data into a time series according to the size of the time step and labels the battery status, thereby organizing learning data for each cycle section. Here, the feature data is organized for learning and consists of feature values ​​measured according to the battery's life cycle.

[0099] For example, if the feature data consists of voltage, current, and temperature, and the time step is 3, the three feature data are continuous (i.e., time series) and constitute learning data in the form of 3 x 3. The time step can be set to an appropriate scale for each cycle section. If the time step increases, interpolation can be performed by interpolating between adjacent values ​​with the average or median value, or by padding with a specific value (e.g., 0).

[0100] Configuring the above DNN involves optimizing hyperparameters for the DNN. The hyperparameters include the depth (number of layers) of the DNN, dropout, learning rate, epoch, batch size of training data according to time steps, loss function, regularization parameter, or a combination thereof.

[0101] The labeled battery status is added based on the learning data, and can be calculated using a predetermined method. Since the learning data can be organized by each cycle section in the present invention, there are no restrictions on the method for labeling the battery status.

[0102] In addition, the learning model generation device (100) trains a DNN specialized for each cycle section using learning data configured for each cycle section to generate a learning model for estimating the battery state for each cycle section.

[0103] The above learning is performed by updating the weights of the DNN through the backpropagation method, which minimizes the error between the output (learning result) of each DNN and the label (i.e., correct answer) of the learning data when learning data is input during the learning process.

[0104] The input of the learning model for estimating the battery status for each cycle section thus generated is a sequence of feature data for each time step configured through feature values ​​received from an application device equipped with an actual battery, and the output is the battery status according to the input.

[0105] FIG. 6 is a diagram illustrating a method for generating a learning model for estimating a battery state according to another embodiment of the present invention.

[0106] As illustrated in FIG. 6, a learning model generation device (100) according to another embodiment of the present invention is configured to include an agent (220).

[0107] The above agent (220) configures feature data from the collected feature values, classifies the configured feature data into predetermined cycle sections, and configures a DNN specialized for each cycle section according to the properties of the feature data corresponding to the classified cycle section.

[0108] The configuration of the above feature data, classification of the feature data, and configuration of the DNN are performed through the same method as described with reference to Fig. 4.

[0109] The above agent (220) uses learning data for each cycle section to perform reinforcement learning on the DNN corresponding to each cycle section, thereby generating a learning model (i.e., a reinforcement learning model) for estimating the battery status for each cycle section.

[0110] The input of the learning model for estimating the battery status for each cycle section generated through the above reinforcement learning is a sequence of feature data according to a time step of a predetermined size, and the output is the battery status.

[0111] FIG. 7 is a diagram illustrating in detail a method for generating a learning model for estimating a battery state according to another embodiment of the present invention.

[0112] As illustrated in FIG. 7, an agent (220) of a learning model generation device (200) according to another embodiment of the present invention sets classification conditions and classifies feature data configured based on feature values ​​into cycle sections according to the set classification conditions.

[0113] In addition, the agent (220) configures learning data for each cycle section using feature data classified by cycle section, and selects and configures a DNN specialized for each cycle section according to the properties of each feature data configured for each cycle section.

[0114] Additionally, the agent (220) organizes learning data by cycle section using the feature values ​​classified by classification conditions.

[0115] The above agent (220) uses learning data configured for each cycle section to reinforce learning the DNN configured for each cycle section, thereby generating a learning model for battery state estimation specialized for each cycle section.

[0116] The above agent (220) calculates a reward according to the reinforcement learning result (output result) of each DNN for the learning data input during the reinforcement learning process, and performs reinforcement learning for each DNN to output a reinforcement learning result that maximizes the reward by applying the reward to each DNN.

[0117] The above reward is calculated based on the error between the reinforcement learning result and the actual battery state. The agent (220) is configured to produce a higher reward as the error approaches zero, and to produce a lower reward as the error increases.

[0118] The above agent (220) can be configured to perform reinforcement learning only when the reward exceeds a predetermined value, and not perform reinforcement learning when the reward is below the predetermined value. This reinforcement learning has the advantage of being able to estimate the optimal battery condition by adapting to feature values ​​that change in real time during the cycle.

[0119] FIG. 8 is a block diagram showing the configuration of a device that generates a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle according to one embodiment of the present invention.

[0120] As illustrated in FIG. 8, a learning model generation device (100) according to one embodiment of the present invention is configured to include a feature value collection unit (110), a feature data configuration unit (120), a feature data classification unit (130), a DNN configuration unit (140), a learning data configuration unit (150), and a learning unit (160).

[0121] The above-mentioned feature value collection unit (110) collects the feature values ​​of the battery measured from the battery. The feature values ​​can be collected from the user's application device.

[0122] The above-mentioned feature data configuration unit (120) configures the collected feature values ​​into feature data by binarizing or encoding the feature values. In other words, the feature values ​​are preprocessed to create a learning model for estimating the battery status.

[0123] The above-mentioned feature data classification unit (130) classifies the feature data into a predetermined cycle section by determining which cycle section the feature data belongs to in the battery life cycle based on pre-set classification conditions. The above-mentioned classification has been described with reference to FIG. 3, and thus is omitted here.

[0124] The above DNN configuration unit (140) selects and configures a DNN specialized for each cycle section according to the properties of the feature data classified for each cycle section.

[0125] Configuring the above DNN includes optimizing hyperparameters, and the DNN selected to be specialized for the first cycle section may be a CNN-based DNN, and the DNN selected to be specialized for the second and third cycle sections may be an RNN-based DNN.

[0126] The above learning data configuration unit (150) configures learning data according to a DNN selected as specialized for each cycle section from the classified data for each cycle section. The configuration of the learning data has been described with reference to FIGS. 4 and 5, and thus is omitted here.

[0127] The above learning unit (160) uses learning data configured for each cycle section to train each DNN selected for each cycle section, thereby generating a learning model for battery state estimation specialized for each cycle section.

[0128] FIG. 9 is a block diagram showing the configuration of a device that generates a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle according to another embodiment of the present invention.

[0129] As illustrated in FIG. 9, a learning model generation device (200) according to another embodiment of the present invention is configured to include a feature value receiving unit (210) and an agent (220).

[0130] The above-described feature value receiving unit (210) receives the feature value of the battery. The above-described feature value receiving unit (210) performs the same function as the feature value receiving unit (110) described with reference to FIG. 8.

[0131] The above agent (220) is configured to divide the battery life cycle into predetermined cycle sections (first to third cycle sections) and to generate a learning model for estimating the battery state for each cycle section by performing reinforcement learning on a DNN specialized for each cycle section, and includes a feature data configuration section (221), a classification condition setting section (222), a feature data classification section (223), a DNN configuration section (224), a learning data configuration section (225), a reward calculation section (226), and a reinforcement learning section (227).

[0132] The above feature data configuration unit (221) configures feature data for feature values.

[0133] The above classification condition setting unit (222) sets classification conditions for classifying the above-configured feature data into predetermined cycle sections. The above classification conditions are for classifying the feature data into predetermined cycle sections in which the capacity changes differently depending on the characteristics of the battery, and are described with reference to FIG. 3, so they are omitted here.

[0134] The above-described feature data classification unit (223) classifies the feature data into predetermined cycle sections according to the above-described classification conditions, the DNN configuration unit (224) selects and configures a DNN specialized for each cycle section according to the properties of the feature data classified into the cycle sections, and the learning data configuration unit (225) configures learning data for each cycle section using the feature data classified for each cycle section.

[0135] The above-described feature data configuration unit (221) to learning data configuration unit (225) perform the same function as the feature data configuration unit (120) to learning data configuration unit (150) illustrated in FIG. 8.

[0136] The above reinforcement learning unit (227) is for performing reinforcement learning on DNN for each cycle section using learning data configured for each cycle section, and inputs the learning data configured for each cycle section into a DNN selected as specialized for each cycle section, and provides the reinforcement learning result, which is the output of each DNN, to the reward calculation unit (226).

[0137] The above reward calculation unit (226) calculates a reward based on the reinforcement learning result for each DNN and provides it to the reinforcement learning unit (227).

[0138] Afterwards, the reinforcement learning unit (227) performs reinforcement learning for each DNN to output a reinforcement learning result that maximizes the reward by applying a reward according to the input learning data to each DNN.

[0139] Since the above-mentioned compensation calculation and reinforcement learning are described with reference to FIGS. 6 and 7, further detailed description will be omitted.

[0140] Additionally, estimating the battery state through a learning model for estimating the battery state generated for each cycle section can be performed through a separate battery state estimation device (not shown).

[0141] The above battery state estimation device is configured to receive a characteristic value of a battery from a user's application device, configure it as characteristic data, classify the characteristic data into cycle sections according to predetermined classification conditions, and configure input data of a learning model for battery state estimation specialized for the corresponding cycle section using the characteristic data, thereby estimating the battery state.

[0142] Meanwhile, when classifying feature data, if it does not meet the classification conditions and is not classified into a specific cycle section, it is determined that there is a problem with the battery and an alarm is immediately sent to the user's application device so that the battery can be inspected or replaced.

[0143] The above battery condition estimation device may be provided on the cloud or on an application device.

[0144] Figure 10 is a flowchart illustrating a procedure for generating a learning model for estimating a battery state according to one embodiment of the present invention.

[0145] FIG. 10 illustrates a procedure for generating a learning model for battery state estimation for each cycle section based on the learning model generation device (100) described with reference to FIG. 7.

[0146] As illustrated in Fig. 10, the learning model creation device (100) performs a feature value collection step of collecting feature values ​​of a battery measured from the battery (S110).

[0147] The above feature values ​​are collected for learning and can be collected from an application device.

[0148] Next, the learning model generation device (100) performs a feature data configuration step of configuring feature data for the collected feature values. The feature data refers to the encoded or binarized feature values.

[0149] Next, the learning model generation device (100) performs a feature data classification step of classifying feature data by cycle section according to pre-set classification conditions (S130).

[0150] As described above, the above classification is performed based on the rate of change and number of cycles for the amount of charge included in the feature data.

[0151] Since the above feature data is generated based on feature values ​​measured according to the life cycle of the battery, the set of feature data classified by cycle section is composed of time series data.

[0152] Next, the learning model generation device (100) performs a DNN configuration step of selecting and configuring a DNN specialized for a cycle section according to the properties of the classified feature data (S140).

[0153] That is, the DNN configuration step further includes configuring DNNs for each cycle section and optimizing hyperparameters for each DNN.

[0154] Next, the learning model generation device (100) performs a learning data configuration step of configuring learning data for each cycle section using the classified feature data (S150).

[0155] The above training data is structured according to the time step size, and the actual battery status is labeled as described above. Here, the training data can be structured using some or all of the feature data (i.e., voltage, current, temperature, charge).

[0156] Next, the learning model generation device (100) performs a learning step in which the learning data configured for each cycle section is trained on each DNN selected as corresponding to the cycle section (S160).

[0157] In addition, the learning model generation device (100) repeats steps S160 to S170 until learning for the DNN for each cycle section is completed (S170) using all learning data.

[0158] At this point, the DNN for each cycle section where learning has completed becomes a specialized learning model for battery state estimation for each cycle section. As described above, this learning is performed via backpropagation.

[0159] Figure 11 is a flowchart illustrating a procedure for generating a learning model for estimating a battery state according to another embodiment of the present invention.

[0160] Figure 11 illustrates a procedure for generating a learning model for battery state estimation for each cycle section based on the learning model generation device (200) described with reference to Figure 9.

[0161] As illustrated in Fig. 11, the learning model creation device (200) performs a feature value collection step of collecting feature values ​​in the feature value collection unit (210) (S210).

[0162] Next, the learning model generation device (200) performs a feature data configuration step (S220) for configuring feature data for feature values ​​in the agent (220), a feature data classification step (S230) for classifying feature data by cycle section according to classification conditions set in advance, a DNN configuration step (S240) for selecting and configuring a DNN specialized for a cycle section according to the properties of the classified feature data, and a learning data configuration step (S250) for configuring learning data for each cycle section using the classified feature data.

[0163] The above steps S210 to S250 are each performed in the same manner as steps S110 to S150 of FIG. 10, and a detailed description thereof is omitted.

[0164] Next, the learning model generation device (100) performs a reinforcement learning result output step (S260) in which learning data configured for each cycle section is input into the DNN selected for each cycle section in the agent (220) to output a reinforcement learning result, and performs a reward calculation step in which a reward is calculated according to the output reinforcement learning result (S270).

[0165] The above reinforcement learning results refer to the output results of each DNN according to the learning data input during the reinforcement learning process.

[0166] Next, a reinforcement learning step is performed in which the agent (220) applies the reward calculated based on the reinforcement learning result of each DNN to each DNN, thereby performing reinforcement learning so that the DNN for each cycle section outputs a reinforcement learning result that maximizes the reward based on the input (S280).

[0167] Thereafter, the learning model generation device (200) performs steps S260 to S290 using all learning data until reinforcement learning is completed. At this time, the DNN for each cycle section where reinforcement learning is completed becomes a learning model for battery state estimation specialized for each cycle section.

[0168] As described above, the present invention has the effect of accurately estimating the battery state by creating a learning model for estimating the battery state specialized for each cycle section in the battery life cycle.

[0169] 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.

[0170] As described above, the present invention has industrial applicability because it allows for more accurate estimation of the battery state by dividing the battery life cycle into predetermined cycle sections and generating a learning model for estimating the battery state through a DNN specialized for each cycle section.

Claims

1. Feature data composition step of collecting feature values measured from a battery and composing feature data; A feature data classification step of dividing the above-described feature data into predetermined cycle sections according to the battery life cycle; A DNN configuration step for configuring a DNN (deep neural network) selected according to the properties of the feature data classified into the above-mentioned predetermined cycle section; A learning data configuration step for configuring learning data according to the predetermined cycle section using the feature data classified into the predetermined cycle section; and A method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, characterized in that it includes a learning step for generating a learning model for estimating a battery state according to each cycle section by training the DNNs selected as corresponding to each cycle section with learning data configured according to the cycle section.

2. In claim 1, The properties of the above feature data are: A method for generating a learning model for estimating a battery state by applying different types of DNNs according to the battery life cycle, wherein the above-mentioned feature data is characterized by a property of a relationship that changes according to time, space, or a combination thereof.

3. In claim 1, The above feature data classification step is, If the number of cycles included in the above characteristic data is less than a predetermined first threshold cycle number and the rate of change in the amount of charge included in the above characteristic data is a positive value, it is classified as a first cycle section. If the above cycle number is within a predetermined critical cycle number range and the rate of change of the charge amount is within a predetermined range, it is classified as a second cycle section, A method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, characterized in that if the number of cycles exceeds a predetermined second threshold number of cycles and the rate of change in the amount of charge is a negative value, the learning model is classified as a third cycle section.

4. In claim 3, The DNN selected to correspond to the first cycle section above is a DNN based on a CNN (convolutional neural network). A method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, wherein the DNN selected to correspond to the second and third cycle sections is a DNN based on a recurrent neural network (RNN).

5. A feature data configuration step in which the agent collects feature values measured from the battery and configures feature data; In the above agent, a feature data classification step of dividing the configured feature data into predetermined cycle sections according to the battery life cycle and classifying them; In the above agent, a DNN configuration step of configuring a DNN (deep neural network) selected according to the properties of feature data classified into a predetermined cycle section; In the above agent, a learning data configuration step of configuring learning data according to each predetermined cycle section using the feature data classified into the predetermined cycle section; and A method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, characterized in that it includes a reinforcement learning step for generating a learning model for estimating a battery state according to the cycle section by performing reinforcement learning on the DNNs selected as corresponding to each cycle section using learning data configured according to the cycle section in the agent.

6. In claim 5, The properties of the above feature data are: A method for generating a learning model for estimating a battery state by applying different types of DNNs according to the battery life cycle, wherein the above-mentioned feature data is characterized by a property of a relationship that changes according to time, space, or a combination thereof.

7. In claim 5, The above feature data classification step is, If the number of cycles included in the above characteristic data is less than a predetermined first threshold cycle number and the rate of change in the amount of charge included in the above characteristic data is a positive value, it is classified as a first cycle section. If the above cycle number is within a predetermined critical cycle number range and the rate of change of the charge amount is within a predetermined range, it is classified as a second cycle section, A method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, characterized in that if the number of cycles exceeds a predetermined second threshold number of cycles and the rate of change in the amount of charge is a negative value, the learning model is classified as a third cycle section.

8. In claim 6, The DNN selected to correspond to the first cycle section above is a DNN based on a CNN (convolutional neural network). A method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, wherein the DNN selected to correspond to the second and third cycle sections is a DNN based on a recurrent neural network (RNN).

9. In claim 6, The above reinforcement learning step is, In the process of performing the above reinforcement learning, a reinforcement learning result output step of inputting the learning data configured for each cycle section into each DNN selected as corresponding to each cycle section and outputting the reinforcement learning result which is the output of each DNN; and It includes a reward calculation step for calculating the reward for each reinforcement learning result output by each DNN; A method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, characterized in that when the reward calculated for each DNN exceeds a predetermined threshold, the reinforcement learning is performed for each DNN to output a reinforcement learning result that maximizes the reward, and when the reward does not exceed a predetermined threshold, the reinforcement learning is performed.

10. A device for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, characterized in that a learning model for estimating a battery state is generated according to a method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle as described in any one of claims 1 to 4.

11. A device for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle, characterized in that a learning model for estimating a battery state is generated according to a method for generating a learning model for estimating a battery state by applying different types of DNNs according to a battery life cycle as set forth in any one of claims 5 to 9.

Images