Method for labeling training data of deep neural network training model for estimating remaining useful lifetime of battery

Abstract

The present invention relates to a method for labeling training data of a deep neural network training model for estimating remaining useful lifetime (RUL) by collecting measurement data from a battery in a real-world environment, and to a method for training a deep neural network through labeling of learning data, a training system therefor, a method for estimating the RUL of a battery through a training model, and a system therefor.

Description

A method for labeling training data for a deep neural network learning model for estimating remaining battery life.

[0001] The present invention relates to a method for labeling learning data of a deep neural network learning model for estimating the remaining useful lifetime (RUL) by collecting measurement data from a battery in an actual use environment, and relates to a deep neural network learning method through labeling of the learning data and a learning system thereof, and a method for estimating the remaining useful lifetime (RUL) of a battery through the learning model and a system thereof.

[0002] As the popularity of electric vehicles rapidly expands, batteries, their core power source, are facing various challenges in terms of environmental friendliness, convenience, economy, and safety. Therefore, predicting battery life in advance is crucial to reduce or prevent fire risks and accidents, and alleviate user inconvenience.

[0003] In various applications that use batteries as their primary energy source, such as electric vehicles, accurately estimating the remaining battery life is crucial for efficient use and management of batteries and improved stability.

[0004] The Remaining Useful Life (RUL) of a battery is information indicating how long it will last. It is mainly predicted based on physical or chemical changes related to the battery such as capacity reduction, internal resistance increase, voltage change, etc. These changes occur while the battery is being charged and discharged and can cause the battery's performance to deteriorate.

[0005] There are several methods for estimating the RUL. One common method involves collecting battery usage and environmental condition data and using it to build a predictive (learning) model. This predictive (learning) model is then used to estimate battery performance degradation based on the data.

[0006] For example, machine learning and data analysis can be used to monitor the periodic condition of a battery, and based on this, the rate of capacity decline and internal resistance increase can be predicted to calculate the remaining lifespan.

[0007] The remaining battery life plays a critical role in managing the maintenance and replacement schedule for application-side equipment or systems. Predicting remaining battery life in advance can prevent system failures and optimize maintenance schedules.

[0008] The main parameters that affect the remaining life of a battery are capacity degrading, increase in internal resistance, cycle count, usage environment and conditions, battery age, and physical shock or damage.

[0009] Additionally, there are methods for training various deep neural networks (DNNs) such as ANNs (artificial neural networks), TCNNs (temporal convolutional neural networks), RNNs (recursive neural networks), LSTMs (long short-term memory), and CNNs (convolutional neural networks) to estimate the remaining life of a battery.

[0010] The present invention aims to estimate the remaining life of a battery by training a deep neural network. Compared to directly measuring the life of a battery through modeling based on physical characteristics, the more training data there is in various operating environments, the more precise the prediction performance of the learning model becomes.

[0011] For this reason, it's crucial to develop methods for labeling and training data from real-world environments, rather than laboratory data. However, because accurately predicting the lifespan of vehicles like electric vehicles during operation is virtually impossible, there's currently no way to obtain labeled data for deep neural network (DNN) training.

[0012] To solve this problem, the present invention proposes a method of constructing learning data through the cycle count of a battery from measurement data at run time and then automatically labeling the learning data with a target value (RUL).

[0013] In addition, the present invention proposes a method for generating a learning model for estimating the remaining life of a battery by training a deep neural network using the labeled learning data.

[0014] In addition, the present invention proposes a method for accurately estimating the remaining life of a battery by extracting measurement data using an indirect measurement method at run time in an electric vehicle, etc., and then inputting the data into a learning model for estimating the remaining life of the battery.

[0015] 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.

[0016] First, Korean Patent No. 2032229 (October 8, 2019) relates to a battery life state estimation system and method thereof, which adopts a low complexity equivalent circuit model and introduces initial electromotive force as a parameter in the parameter estimation process to enable prediction of the current battery life state based on the initial charge state.

[0017] The above Korean Patent No. 2032229 (October 8, 2019) calculates the battery state of charge (SOC) and open circuit voltage (OCV) from the current (I) based on a mathematical model derived from an equivalent circuit model, and calculates the state of health (SOH) of the battery as the ratio of the initial battery capacity parameter and the current battery capacity parameter. In the present invention, the cycle count is extracted from the measurement data collected from the battery at the run-time to configure learning data, and then the RUL (remaining useful lifetime) is automatically labeled as a target value for the learning data to train a deep neural network to create a learning model, and the remaining life of the battery is estimated through the created learning model, which presents a different technical feature.

[0018] In addition, Korean Patent Publication No. 2016-0067510 (June 14, 2016) relates to a method and device for estimating the state of a battery, which divides sensing data for a battery into predetermined sections, accumulates time information corresponding to the divided sections, extracts time information corresponding to at least one cycle from the accumulated time information, extracts expected time information based on the accumulated time information, the time information corresponding to at least one cycle, and predetermined learning information, and estimates the end-of-life time of the battery based on the expected time information.

[0019] The above Korean Patent Publication No. 2016-0067510 (June 14, 2016) estimates the lifespan by accumulating time information, which is clearly a different technical feature from the method of estimating the lifespan of a battery based on the current direction, charge amount, and effective usage time of the battery presented in the present invention.

[0020] The present invention was created to solve the above problems, and the purpose of the present invention is to present a learning data labeling method of a deep neural network for estimating the remaining useful lifetime (RUL) by collecting cycle count information from a battery as measurement data in an actual use environment.

[0021] In addition, another object of the present invention is to provide a method and system for training a deep neural network through a method of labeling learning data of the deep neural network in a real-world environment.

[0022] In addition, another purpose of the present invention is to provide a method and system for estimating the remaining useful life (RUL) of a battery using a learning model learned through labeling of learning data of a deep neural network in a real-world environment.

[0023] In addition, the present invention has another purpose of collecting cycle count information of a battery in an actual use environment, extracting cycle counts as feature values ​​based on current direction, charge amount (capacity), and time to construct learning data, and labeling the remaining life (RUL) of the battery according to the constructed learning data.

[0024] In addition, the present invention has another purpose of evaluating a learning model for estimating the remaining useful life (RUL) of a battery according to the cycle count of the battery in a real-world environment, and adjusting the internal weights of the learning model through backpropagation or optimizing the learning model through hyperparameter tuning.

[0025] In addition, another object of the present invention is to provide a method for collecting information including capacity, temperature, voltage, current or a combination thereof in addition to cycle counts when configuring learning data for a deep neural network, filtering through preprocessing or performing data normalization, and then converting feature values ​​including cycle counts, capacity change rate, temperature change rate, internal resistance change rate or a combination thereof into learning data.

[0026] In addition, another object of the present invention is to provide a method for estimating the remaining life of a battery by converting measurement data measured at run time into input data and inputting the converted input data into a learning model for estimating the remaining life of the battery, thereby estimating the RUL of the battery.

[0027] A method for labeling learning data of a deep neural network learning model for estimating the remaining life of a battery according to one embodiment of the present invention comprises: a learning data configuration step of configuring learning data by collecting measurement data for a battery at run time and converting it into learning data; and a learning data labeling step of labeling the configured learning data with a remaining useful lifetime (RUL) according to a cycle count of the battery; wherein the remaining life is extracted by subtracting the collected cycle count (useful cycle count) from the maximum useful cycle count of the battery, and the labeling is characterized in that the value of the extracted count is automatically assigned as a label of the learning data.

[0028] The above learning data is characterized in that it includes feature values ​​including open circuit voltage (OCV), voltage (V), current (I), charge (Q), temperature (T) of the battery, or a combination thereof, or includes statistical values ​​including the mean, variance, median, standard deviation, difference, or a combination thereof of the feature values, or includes a capacity change rate, a temperature change rate, an internal resistance change rate, or a combination thereof of the battery extracted from the feature values, or further includes a combination thereof.

[0029] The above maximum useful cycle count is characterized in that it is set to have a predetermined margin with respect to the maximum useful cycle count of the battery.

[0030] The above learning data configuration step includes a measurement data collection step of collecting measurement data by filtering or normalizing measurement values ​​measured from a battery at run time; a feature value extraction step of extracting feature values ​​to configure learning data from the collected measurement data; and a learning data conversion step of converting the extracted feature values ​​into learning data in a format for training the deep neural network; wherein the feature value extraction step is characterized in that it calculates or extracts a cycle count of the battery from the collected measurement data, and extracts a remaining useful life (RUL) from the calculated or extracted cycle count.

[0031] Here, the cycle count is a current direction-based cycle count, and the current direction-based cycle count is extracted by accumulating and counting the number of times that there is no change in the current direction flowing through the battery for a predetermined period of time at run time, and the remaining useful life (RUL) of the battery is extracted by subtracting the current direction-based cycle count extracted at the run time from a maximum useful cycle count of the battery, which is predetermined as a maximum useful cycle count of the battery.

[0032] In addition, the cycle count here is a capacity-based cycle count, and the capacity-based cycle count is extracted by adding up a predetermined amount of charge that is charged and discharged in the battery at run time and then accumulating and counting the number of times the predetermined amount of charge is reached, and the remaining useful life (RUL) of the battery is extracted by subtracting the capacity-based cycle count extracted at the run time from a predetermined maximum useful cycle count based on the battery's capacity.

[0033] In addition, the cycle count is a use time-based cycle count, and the use time-based cycle count is extracted by repeatedly counting a predetermined use time interval during which the battery is used in run time, and the remaining useful life (RUL) of the battery is extracted by subtracting the use time-based cycle count extracted in the run time from a maximum useful cycle count predetermined as a maximum useful cycle count of the battery.

[0034] Meanwhile, a learning method for generating a deep neural network learning model for estimating the remaining life of a battery according to another embodiment of the present invention is characterized by including generating a deep neural network learning model for estimating the remaining life of the battery through a learning data labeling method of a deep neural network learning model for estimating the remaining life of the battery according to the present invention.

[0035] Meanwhile, a learning system for generating a deep neural network learning model for estimating the remaining life of a battery according to another embodiment of the present invention comprises: a memory storing a program configured to generate a deep neural network learning model for estimating the life of the battery through a learning data labeling method of the deep neural network learning model for estimating the remaining life of the battery according to the present invention; and a processor configured to execute the program stored in the memory.

[0036] Meanwhile, a method for estimating a remaining life of a battery according to another embodiment of the present invention is characterized by including estimating the remaining life of the battery by inputting measurement data measured from the battery as input data into a deep neural network learning model for estimating the life of the battery generated through a learning data labeling method of a deep neural network learning model for estimating the remaining life of the battery according to the present invention.

[0037] Meanwhile, a battery remaining life estimation system according to another embodiment of the present invention comprises: a memory storing a program configured to estimate the remaining life of a battery by inputting measurement data measured from the battery as input data into a deep neural network learning model for estimating the life of the battery generated through a learning data labeling method of a deep neural network learning model for estimating the remaining life of the battery according to the present invention; and a processor configured to execute the program stored in the memory.

[0038] As described above, the present invention has the effect of automatically labeling learning data of a deep neural network for estimating the remaining useful lifetime (RUL) of a battery by collecting cycle count information from the battery as measurement data in an actual use environment.

[0039] In addition, the present invention has the effect of generating a high-performance battery remaining life estimation learning model by training the deep neural network through the learning data labeling method of the deep neural network.

[0040] In addition, the present invention has the effect of enabling efficient maintenance and management of the battery by estimating the exact remaining life of the battery and knowing the timing of battery replacement in advance.

[0041] In addition, the present invention has the effect of enabling stable and continuous use of an application by estimating the exact remaining life of a battery and knowing the timing of battery replacement in advance, thereby replacing the battery at the right time.

[0042] FIG. 1 is a conceptual diagram illustrating a concept of collecting measurement data for labeling learning data that creates a deep neural network learning model for estimating remaining battery life according to one embodiment of the present invention.

[0043] FIG. 2 is a diagram illustrating a concept of collecting current direction-based cycle counts as measurement data for labeling learning data that generates a deep neural network learning model for estimating the remaining life of a battery according to one embodiment of the present invention.

[0044] FIG. 3 is a diagram illustrating a concept of collecting charge-based cycle counts as measurement data for labeling learning data to create a deep neural network learning model for estimating remaining battery life according to one embodiment of the present invention.

[0045] FIG. 4 is a diagram illustrating a concept of collecting usage time-based cycle counts as measurement data for labeling learning data that creates a deep neural network learning model for estimating remaining battery life according to one embodiment of the present invention.

[0046] FIG. 5 is a diagram illustrating a method for labeling learning data with RUL to create a deep neural network learning model for estimating the remaining life of a battery according to one embodiment of the present invention.

[0047] FIG. 6 is a conceptual diagram illustrating a concept of estimating the remaining life of a battery using a deep neural network learning model for estimating the remaining life of a battery according to one embodiment of the present invention.

[0048] FIG. 7 is a block diagram showing the configuration of a learning system that generates a deep neural network learning model that estimates the remaining life of a battery through a learning data labeling method of a deep neural network according to one embodiment of the present invention.

[0049] Figure 8 is a block diagram showing the configuration of a measurement data collection unit in a learning system according to one embodiment of the present invention.

[0050] FIG. 9 is a block diagram showing the configuration of a battery remaining life estimation system that estimates the remaining life of a battery by measuring the characteristic values ​​of the battery and inputting them into a deep neural network learning model that estimates the remaining life of the battery generated through a learning data labeling method of a deep neural network according to one embodiment of the present invention.

[0051] Fig. 10 is a block diagram showing the configuration of an input data collection unit in a battery remaining life estimation system according to one embodiment of the present invention.

[0052] FIG. 11 is a flowchart illustrating a procedure for labeling and learning learning data according to one embodiment of the present invention and a procedure for estimating the remaining life of an actual battery using a learning model for estimating the remaining life of a battery generated through the learning.

[0053] [Description of symbols] 100: learning system; 100a: learning engine; 300: battery remaining life estimation system; 300a: battery remaining life estimation engine; 200: database; 110: learning data composition unit; 111, 310: measurement value receiving unit; 112, 320: steady state recognition unit; 113: measurement data collection unit; 114: learning data conversion unit; 1131: measurement value filtering unit; 1132: measurement value expansion unit; 1133: cycle count unit; 1134: feature value extraction unit; 120: remaining life calculation unit; 130: learning data labeling unit; 140: learning model generation unit; 330: input data composition unit; 331: estimation-use measurement value filtering unit; 332: estimation-use measurement value expansion unit; 333: estimation-use feature value extraction unit; 340: Remaining life estimation section.

[0054] Hereinafter, with reference to the attached drawings, a preferred embodiment of a method for labeling learning data of a deep neural network learning model for estimating the remaining life of a battery according to the present invention will be described in detail. The same reference numerals 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 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.

[0055] FIG. 1 is a conceptual diagram illustrating a concept of collecting measurement data for labeling learning data of a deep neural network learning model for estimating remaining battery life according to one embodiment of the present invention.

[0056] As illustrated in FIG. 1, the learning system (100) for generating a deep neural network learning model for estimating the remaining life of a battery according to the present invention receives measurement values ​​from the battery in any case, whether the application is an electric vehicle, whether the electric vehicle is driving, parked, or charging, collects measurement data, converts the collected measurement data into learning data, and then labels it with the remaining life (RUL) according to the present invention to train a deep neural network. As a result of learning the deep neural network, a learning model is generated.

[0057] The above-generated learning model is stored and managed in a database (200) for each condition (temperature, battery specifications, etc.). Measurement data, learning data, and labeled learning data can also be stored and managed in the database (200).

[0058] Additionally, measurements according to the present invention are measured at runtime for each battery. However, using measurements measured at runtime as-is may not properly train a deep neural network.

[0059] Therefore, in the present invention, a process of preprocessing the measurement values ​​and then processing them into measurement data suitable for training a deep neural network is required.

[0060] That is, the learning system (100) receives run-time measurement values ​​from an application in a life cycle according to use in various environments for multiple batteries, collects measurement data by removing noise or normalizing the measurement values ​​so that the measurement values ​​can be practically used to predict the remaining life of the batteries, converts the collected measurement data into learning data, labels the converted learning data with the remaining life, and then trains a deep neural network (DNN). After testing whether sufficient accuracy is guaranteed through the learning, a learning model is created and stored in a database (200).

[0061] The above database (200) stores the collected measurement data and the training data converted into training data, and stores not only the deep neural network to be trained but also the deep neural network for which training has been completed. The above measurement data and training data include battery cycle counts, and information regarding the basis of the cycle counts can also be stored together.

[0062] For example, the measurement data detects a case where the direction of the current does not change from the measurement value and remains in one direction, determines the cycle time, and extracts measurement data including further characteristic values ​​in the cycle, such as OCV, voltage (V), current (I), charge (Q), and temperature (T). The process of extracting the measurement data includes a process of filtering out or normalizing unnecessary information.

[0063] In general, when the application side is running, operating, or in use at runtime, the battery measurement values ​​are unstable, so it is not appropriate to use those measurements as learning data.

[0064] However, when performing a cycle count based on the current direction at run time according to the present invention, since the direction of the current does not change in the cycle, the state of the battery is stabilized even at run time, so it is possible to collect measurement data by extracting cycle counts and feature values ​​in the stabilized state.

[0065] In addition, when performing a cycle count based on the charge amount according to the present invention, since it is impossible to calculate a point where a predetermined charge amount is achieved by accumulating the charge amount in the run time, the accurate charge amount is extracted when the battery is in a stable state including a rest period in which the charge amount can be accurately measured, and interpolation is performed using an average value of the charge amounts measured in two stabilized states between the intervals in which the stabilized state is repeated, and interpolation is performed between the stabilized sections and the stabilized sections in this way, and measurement data including the cycle count and the feature value are collected from this.

[0066] In addition, when performing cycle counting based on usage time according to the present invention, if the usage time of the battery is divided into predetermined time intervals and counted in run time, the battery's idle period has relatively less influence on the remaining lifespan than the time for performing charging and discharging, so it would be desirable to reflect this.

[0067] Therefore, when the stabilized period including the battery's idle period is reached, the measured value is received, and between the stabilized periods, the measured value is determined by interpolating the average value of the measured values ​​of the two stabilized periods, and the duration of the stabilized period is relatively less reflected in the usage time compared to the period of repeating the charging and discharging, and the cycle of the usage time is counted, and measurement data including the characteristic value corresponding to the cycle count is collected.

[0068] The above-mentioned idle period is defined as a period in which the measured data (OCV, etc.) changes only within a specified range and stabilizes over a specified period of time. In other words, the above-mentioned stabilized period refers to a period of time during which the battery is stabilized without charging or discharging, including the above-mentioned idle period.

[0069] Accordingly, the present invention collects the cycle count and feature values ​​by interpolating the feature values ​​between the two periods with the average value between the two periods based on the feature values ​​measured in the previous stabilization period and the feature values ​​measured in the next stabilization period in two adjacent stabilization periods in the life cycle of the battery, converts the collected measurement data into learning data, and labels the learning data with the remaining life extracted through the collected measurement data.

[0070] Therefore, by estimating the remaining useful life (RUL) using the label estimated from the output of the learning model for estimating the remaining battery life, it becomes possible to estimate the life of the battery. In other words, the remaining battery life is calculated based on the RUL label, which is the output of the learning model for estimating the remaining battery life.

[0071] Accordingly, the remaining life estimation system (300) configures the measurement data of the actual battery as input data to be input into a learning model for estimating the remaining life of the battery, and inputs the configured input data into the learning model for estimating the remaining life of the battery to estimate the remaining life of the battery according to the RUL label, which is the output of the learning model for estimating the remaining life of the battery.

[0072] In addition, the deep neural network in the present invention can be composed of various machine learning networks such as an artificial neural network (ANN), a deep convolutional neural network (DCNN), a transformer, and a temporal convolution neural network (TCNN).

[0073] FIG. 2 is a diagram illustrating a concept of collecting current direction-based cycle counts as measurement data for labeling learning data that creates a deep neural network learning model for estimating the remaining life of a battery according to one embodiment of the present invention.

[0074] As illustrated in Fig. 2, the change in current can be measured by measuring the amount of charge from the battery at runtime and calculating the change in the amount of charge per unit time (dQ / dt). Current direction-based cycle counting can be performed by increasing the count of the corresponding cycle when the change in the current only changes within a predetermined threshold value.

[0075] If the change in charge per unit time is within a certain range, this indicates that the application-side battery remains in a stable state after operation has ceased. This stable state is maintained even during the battery's idle period.

[0076] Typically, when a battery undergoes repeated charging and discharging during runtime, the battery's state becomes unstable, making it impossible to extract feature values. However, in a current-direction-based cycle, the current direction remains constant, so the battery's state remains stable even during runtime. Therefore, in this stabilized state, it is possible to collect measurement data by extracting cycle counts and feature values ​​from measured values ​​from the battery.

[0077] The feature data including the cycle count and feature value extracted from the measurement value measured for each current direction-based cycle is converted into learning data, and the cycle count extracted from the measurement data is called a current direction-based cycle count, and the current direction-based cycle count is extracted by accumulating and counting the number of times there is no change in the current direction flowing through the battery for a predetermined time or longer at run time, and the remaining useful life (RUL) of the battery is extracted by subtracting the current direction-based cycle count extracted at the run time from a maximum current direction-based cycle count predetermined as the maximum useful cycle count of the battery.

[0078] The remaining life is extracted by subtracting the collected cycle count (useful cycle count) from the maximum useful cycle count of the battery, and the labeling is performed so that the value of the extracted count is automatically assigned as a label of the learning data.

[0079] Here, the maximum useful cycle count is set to have a predetermined margin with respect to the maximum lifetime cycle count of the battery.

[0080] Meanwhile, in the present invention, the learning data includes feature values ​​including open circuit voltage (OCV), voltage (V), current (I), charge (Q), temperature (T) of the battery, or a combination thereof, or includes statistical values ​​including the mean, variance, median, standard deviation, difference, or a combination thereof of the feature values, or includes a capacity change rate, a temperature change rate, or a combination thereof of the battery extracted from the feature values, or further includes a combination thereof.

[0081] FIG. 3 is a diagram illustrating a concept of collecting charge-based cycle counts as measurement data for labeling learning data to create a deep neural network learning model for estimating remaining battery life according to one embodiment of the present invention.

[0082] As shown in Fig. 3, when performing a cycle count based on charge amount, the charge amount at run time will be affected by the load and will vary depending on the environment, so it will be impossible to accumulate this and calculate the point at which the charge amount becomes a certain amount.

[0083] Therefore, the present invention extracts the exact charge amount when the battery reaches a stable state, including a rest period during which the charge amount can be accurately measured. Between the intervals where the stabilized state is repeated, interpolation is performed using the average of the charge amounts measured in two stabilized states to complete the overall measurement data. From the completed measurement data, measurement data including cycle counts and feature values ​​are collected.

[0084] Here, the cycle count is a capacity-based cycle count, and the capacity-based cycle count is extracted by adding up a predetermined amount of charge that is charged and discharged in the battery at run time and then accumulating and counting the number of times the predetermined amount of charge is reached, and the remaining useful life (RUL) of the battery is extracted by subtracting the capacity-based cycle count extracted at the run time from a predetermined maximum useful cycle count based on the battery's capacity.

[0085] FIG. 4 is a diagram illustrating a concept of collecting usage time-based cycle counts as measurement data for labeling learning data that creates a deep neural network learning model for estimating remaining battery life according to one embodiment of the present invention.

[0086] As illustrated in FIG. 4, when performing cycle counting based on usage time according to the present invention, if the usage time of the battery is divided into predetermined time intervals and counted in run time, the idle period of the battery has relatively less influence on the remaining lifespan compared to the time for performing charging and discharging, so it would be desirable to reflect this.

[0087] Therefore, when the stabilization period including the battery's idle period is reached, the measurement value is received, and between the stabilization periods, the measurement values ​​of the two stabilization periods are averaged and interpolated to complete the measurement data.

[0088] Here, it is desirable to count the cycle of the usage time by reflecting the duration of the stabilized period relatively less than the period of repeating the charging and discharging.

[0089] Considering these circumstances, measurement data including cycle counts and feature values ​​are collected from the completed measurement data. The collected measurement data is converted into training data, and the training data is labeled using the cycle counts extracted from the measurement data.

[0090] Here, the cycle count is a use time-based cycle count, and the use time-based cycle count is extracted by repeatedly counting a predetermined use time interval during which the battery is used in run time, and the remaining useful life (RUL) of the battery is extracted by subtracting the use time-based cycle count extracted in the run time from a maximum useful cycle count predetermined as the maximum useful cycle count of the battery.

[0091] FIG. 5 is a diagram illustrating a method for labeling learning data with RUL to create a deep neural network learning model for estimating the remaining life of a battery according to one embodiment of the present invention.

[0092] As shown in Fig. 5, RUL is extracted from measurement data measured at run time, the measurement data is converted into learning data, and then the learning data is automatically labeled with the RUL.

[0093] The above run-time measurement data includes cycle counts and characteristic values ​​of the battery, such as OCV (open circuit voltage), V (measured voltage), I (current), Q (capacity), T (temperature), Vs (voltage statistics), CC (cycle count), QC (capacity change rate), TC (temperature change rate), RC (internal resistance change rate), or a combination thereof.

[0094] Here, the feature values ​​are converted into learning data, and the learning data is labeled with the RUL extracted from each learning data. Here, the learning data may further include statistical values, including the mean, variance, median, standard deviation, difference, or a combination thereof, of each feature value extracted from the collected feature values.

[0095] The above RUL is assigned as a label to each training data, and the training data can be arranged in various sizes and forms, such as a 3x3 or 1x9 matrix, depending on the input form of the deep neural network to be trained. Figure 5 is only one example, and various modifications are of course possible. In addition, new training data items can be added as needed, and the deep neural network can also be trained by modifying them to various sizes and depths.

[0096] That is, the format of the above learning data has a very diverse structure depending on the arrangement of each feature value and the time series range of the feature values ​​measured continuously.

[0097] Here, Vs provides statistical information (mean, variance, median, standard deviation, difference, etc.) about the measured voltage V, allowing for controlling the learning of the learning network through various combinations with the measured voltage. For example, when interpolating the measured voltage, the interpolated voltage value can be set to 0 or the average value to eliminate the reflection of the learning results in the learning network or control the degree of reflection.

[0098] FIG. 6 is a conceptual diagram illustrating a concept of estimating the remaining life of a battery using a deep neural network learning model for estimating the remaining life of a battery according to one embodiment of the present invention.

[0099] As illustrated in FIG. 6, the deep neural network learning model generated by the learning system (100) according to the present invention is stored and managed in a database (200). In this case, the deep neural network learning model may be utilized to independently estimate the remaining life of a battery through input data provided from each device in an application area such as an electric vehicle, or may be provided to a battery remaining life estimation system (300) on a network to be used to remotely estimate the remaining life of a battery.

[0100] The above learning system (100) can be implemented on a server on a network (or in the cloud) to also perform the role of a system (300) for estimating the remaining life of a battery.

[0101] In other words, the deep neural network learning model can be uploaded to each device or cloud server (which functions as a remaining life estimation system) in the application area and used to estimate the remaining life of each battery. In application areas such as electric vehicles, self-collected measurement data can be converted into input data for the deep neural network learning model to locally estimate the remaining life of the battery.

[0102] Meanwhile, when computing power is insufficient in each device of the application area, the input data of the self-generated deep neural network learning model is transmitted to a remote cloud or network-based remaining life estimation system (300), and the remaining life of the battery can be estimated remotely by inputting the input data into the deep neural network learning model uploaded to the remaining life estimation system (300).

[0103] In each case, the deep neural network learning model can be uploaded and used from a database (200) on a network (cloud).

[0104] In addition to the learning model for estimating the remaining battery life, the above database (200) can also store and manage run-time measurement data or labeled learning data.

[0105] Finally, by collecting measurement data from each device in each application area and configuring input data, and then inputting it into a deep neural network learning model, which is a learning model for estimating the remaining battery life, the RUL label is estimated as a result, and the remaining battery life can be estimated based on the RUL label.

[0106] FIG. 7 is a block diagram showing the configuration of a learning system that generates a deep neural network learning model that estimates the remaining life of a battery through a learning data labeling method of a deep neural network according to one embodiment of the present invention.

[0107] As illustrated in FIG. 7, a learning system (100) according to another embodiment of the present invention comprises, as its core components, a learning engine (100a) including a memory storing a program configured to generate a deep neural network learning model for estimating the remaining life of the battery through the learning data labeling method of the deep neural network, and a processor configured to execute the program stored in the memory.

[0108] The above learning engine (100a) is configured to include a learning data configuration unit (110), a remaining life calculation unit (120), a learning data labeling unit (130), and a learning model generation unit (140).

[0109] The above learning data configuration unit (110) includes a measurement value receiving unit (111), a stable state recognition unit (112), a measurement data collection unit (113), and a learning data conversion unit (114).

[0110] The above learning data configuration unit (110) collects measurement data from measurement values ​​received from a battery, and converts the collected measurement data into learning data in a predetermined format according to a deep neural network to configure learning data.

[0111] The remaining life calculation unit (120) calculates the remaining life (RUL) from the measurement data collected from the battery to label the configured learning data. What is actually extracted from the measurement data is the cycle count of the battery, and the remaining life (RUL) is calculated by subtracting the cycle count from the maximum life of the battery.

[0112] The method of calculating the remaining life based on the above cycle count is omitted here as it is explained based on current direction, charge amount, and usage time.

[0113] However, the remaining life of the battery is calculated by subtracting the extracted cycle count (useful cycle count) from the maximum useful cycle count of the battery.

[0114] In addition, the learning data labeling unit (130) is configured to automatically label the learning data configured in the learning data configuration unit (110) with the calculated remaining battery life (RUL). The labeling is configured to automatically assign the remaining battery life calculated by subtraction as a label for the learning data.

[0115] Here, the learning data includes basic measurement values ​​including open circuit voltage (OCV), voltage (V), current (I), charge (Q), temperature (T) of the battery, or a combination thereof as feature values, or includes statistical values ​​including the mean, variance, median, standard deviation, difference, or a combination thereof of the feature values, or includes a capacity change rate, temperature change rate, internal resistance change rate, or a combination thereof of the battery extracted from the feature values.

[0116] In the above learning data configuration unit (110), the measurement value receiving unit (111) receives a measurement value measured from the battery. The measurement value basically includes the open circuit voltage (OCV), voltage (V), current (I), charge (Q), temperature (T) of the battery, or a combination thereof. The measurement value is referred to as a basic measurement value.

[0117] The above-mentioned stable state recognition unit (112) determines that the battery is in a stable state when the change in OCV, current, voltage, and charge amount from the received measurement values ​​is within a predetermined range, and in this case, recognizes the corresponding measurement value as a valid measurement value. The valid measurement value is used as a basic feature value.

[0118] The above-mentioned stabilized state includes the battery's relaxation period, which refers to the period of time after a significant period of time has elapsed since operation or use ceased in an application such as an electric vehicle, during which the battery remains in a stabilized state. Beyond simply measuring time, such as 3 or 6 hours, this is also determined by whether the battery's OCV changes within a specified range over time.

[0119] Figure 8 is a block diagram showing the configuration of a measurement data collection unit in a learning system according to one embodiment of the present invention.

[0120] As illustrated in FIG. 8, the measurement data collection unit (113) according to the present invention is configured to include a measurement value filtering unit (1131), a measurement value expansion unit (1132), a cycle count unit (1133), and a feature value extraction unit (1134).

[0121] The above measurement value filtering unit (1131) performs filtering to remove unnecessary measurement values ​​from the valid measurement values ​​in the stabilized state. The unnecessary measurement values ​​refer to measurement values ​​that cannot be used as valid measurement data among the multiple measurement values ​​in the stabilized state.

[0122] The above measurement value expansion unit (1132) completes the measurement value by interpolating the measurement values ​​for sections other than the stabilized state using the filtered measurement values. The interpolation is a process of inserting data extracted through padding with an average, linearity, arbitrary value, or a specific functional relationship using at least one filtered measurement value of the stabilized state.

[0123] The above cycle counting unit (1133) is a process for counting the total cycles for a specific battery using the completed measurement values. The cycle counting is performed based on the current direction, charge amount, and usage period according to the present invention.

[0124] The above-mentioned feature value extraction unit (1134) extracts feature values ​​for training a deep neural network by calculating the cycle count extracted from the completed measurement value and other capacity change rates, temperature change rates, and internal resistance change rates. The present invention is not limited to the feature values ​​listed above, and various feature values ​​can be further extracted from the completed measurement value and used to train a learning model.

[0125] Ultimately, the above measurement data collection unit (113) receives basic measurement values ​​including OCV-SOC, voltage (V), current (I), charge (Q), temperature (T), or a combination thereof measured by an indirect measurement method, and processes them into measurement data including various feature values ​​suitable for use as learning data. For example, this may include unifying the format (resolution, integer, floating point, etc.) of each measurement data.

[0126] The above learning data conversion unit (114) is configured to convert the feature values ​​collected by the measurement data collection unit (113) into learning data in a format usable by a deep neural network. The learning data may further include statistical values ​​including the mean, variance, median, standard deviation, difference, or a combination thereof extracted from the feature values ​​of the collected measurement data. The format of the learning data is described with reference to FIG. 5, and thus is omitted here.

[0127] The remaining useful lifetime calculation unit (120) extracts the cycle count extracted from the cycle count unit (1133) by subtracting it from the maximum useful cycle count according to each method, based on the current direction, charge amount, and usage period according to the present invention. This is called the remaining useful lifetime (RUL) and becomes the label of the learning data.

[0128] The above-described learning data labeling unit (130) is configured to automatically label the converted learning data with the calculated RUL. The label may be recognized as the target value of the RUL. In practice, estimating the RUL using the learning model according to the present invention is synonymous with estimating the target value.

[0129] In the above deep neural network, all data points of the learning data used for learning are data columns that list measurement data through indirect measurement regardless of time interval.

[0130] For example, in the learning data labeling method according to the present invention, the learning data of the deep neural network is set to a first window having a fixed sampling interval, and the first window is set to be larger than a second window having a predetermined sampling interval for the received measurement value, and for sampling points that are insufficient in the measurement value, the measurement data is configured by interpolating with predetermined data.

[0131] The above learning model generation unit (140) is configured to generate a learning model for estimating the remaining battery life by training a deep neural network using the labeled learning data.

[0132] The above learning system (100) labels the learning data with a target value according to the above learning data, and is configured to perform deep learning on the labeled learning data by minimizing the error by backpropagating the error for the output of the deep neural network.

[0133] The learning model for estimating the remaining battery life thus generated is configured to output a value that is probabilistically closest to the target value of the remaining battery life (RUL) when input data composed of measurement data measured at run time is input.

[0134] FIG. 9 is a block diagram showing the configuration of a battery remaining life estimation system that estimates the remaining life of a battery by measuring the characteristic values ​​of the battery and inputting them into a deep neural network learning model that estimates the remaining life of the battery generated through a learning data labeling method of a deep neural network according to one embodiment of the present invention.

[0135] As illustrated in FIG. 9, a battery remaining life estimation system (300) according to another embodiment of the present invention comprises a battery remaining life estimation engine (300a) including a memory (2000) storing a program configured to estimate the remaining life of the battery by extracting a feature value of the battery and inputting it into a deep neural network learning model that estimates the remaining life of the battery generated through the learning data labeling method of the deep neural network, and a processor (1000) configured to execute the program stored in the memory.

[0136] Here, the user interface (3000), data interface (4000), and network interface (5000) are means for providing user, database, and network interfaces, which are typically provided on a user terminal or a server on a cloud or network. Furthermore, the web server (6000) is provided to facilitate easy access of user terminals from a server on a cloud or network.

[0137] The above battery remaining life estimation engine (300a) is configured to estimate the remaining life of an actual battery using actual run-time measurement data and a learning model for estimating the remaining life of a battery, and includes a measurement value receiving unit (310), a stable state recognition unit (320), an input data composition unit (330), and a remaining life estimation unit (340).

[0138] The above measurement value receiving unit (310) is configured to receive measurement values ​​of the battery from an application provided by the user at runtime. The measurement values ​​include basic characteristic values ​​including OCV-SOC, voltage (V), current (I), charge (Q), temperature (T), or a combination thereof.

[0139] The above stable state recognition unit (320) determines that the battery is in a stable state when the change in OCV, current, voltage, and charge amount from the received measurement values ​​is within a predetermined range, and in this case, recognizes the measurement value as a valid measurement value to be used as input data for a deep neural network learning model that estimates the remaining life of the battery.

[0140] The above input data configuration unit (330) is configured to convert the valid measurement values ​​received by the measurement value receiving unit (310) for the period recognized as a stable state by the stable state recognition unit (320) into input data to be input into a learning model for estimating the remaining battery life.

[0141] The above input data can be configured for at least one stable period, and depending on the learning model, it is possible to configure the input data for not only one stable period but also multiple stable periods.

[0142] Fig. 10 is a block diagram showing the configuration of an input data configuration unit in a battery remaining life estimation system according to one embodiment of the present invention.

[0143] As illustrated in FIG. 10, the input data configuration unit (330) according to the present invention comprises an estimation measurement value filtering unit (331), an estimation measurement value expansion unit (332), and an estimation feature value extraction unit (333). Of course, if sufficient input data is configured to estimate the remaining life of a battery in the battery remaining life estimation system (300), any of the estimation measurement value filtering unit (331), the estimation measurement value expansion unit (332), or the estimation feature value extraction unit (333) may be omitted.

[0144] The above-mentioned estimation measurement value filtering unit (331) performs filtering on the measurement values ​​received from the measurement value receiving unit (310) to remove unnecessary measurement values ​​based on the valid measurement values ​​in the stabilized state of the stable state recognition unit (320). The unnecessary measurement values ​​refer to measurement values ​​that cannot be used as valid measurement data among the plurality of valid measurement values ​​in the stabilized state.

[0145] The above-mentioned estimation measurement value expansion unit (332) interpolates the measurement values ​​for a section other than the stabilized state using the filtered measurement values ​​to complete the measurement values ​​to be input as input data. When a request to estimate the remaining life of a battery is input, the remaining life is estimated using the measurement values ​​input in the immediately preceding stabilized period from the time the request is input, or the data extracted through an average, linear, or specific functional relationship using the measurement values ​​measured in at least one previous stabilized period from the time the request is input is inserted up to the time the request is input, thereby expanding the measurement values ​​up to the time when the estimation of the remaining life of the battery is requested.

[0146] The above-mentioned estimation feature value extraction unit (333) calculates the expanded measurement value and the capacity change rate, temperature change rate, and internal resistance change rate of the battery that can be extracted from the expanded measurement value, and extracts them as feature values ​​of the learning model for estimating the remaining life of the battery. The present invention is not limited to the feature values ​​listed above, and various feature values ​​can be further extracted from the basic measurement value and used as inputs of the learning model for estimating the remaining life of the battery.

[0147] Accordingly, the input data configuration unit (330) may include filtering, expanding, and extracting feature values ​​from the measurement values, and is configured to extract the input data using the extracted feature values. This makes it possible to estimate the remaining life of the battery at all data points requesting an estimation of the remaining life of the battery.

[0148] For example, if a measurement value measured at a specific point in time is not suitable for estimating the remaining life of a battery, i.e., if the state of the battery is not stabilized, the remaining life of the battery can be estimated using a measurement value measured during a previously stabilized period, or by configuring input data with feature values ​​extended from the measurement value measured during the previously stabilized period to a point in time when estimation of the remaining life is requested, so as to estimate the remaining life of the battery.

[0149] The above remaining life estimation unit (340) is configured to input the input data configured in the above input data configuration unit (330) into a learning model for estimating the remaining life of the battery to estimate the remaining life.

[0150] That is, the remaining life estimation unit (340) uploads a learning model from the database (200) and then estimates a remaining life (RUL) label using the input data configured in the input data configuration unit (330). The remaining life of the battery is estimated based on the label. The remaining life estimation can be performed using current direction, charge amount, usage period, or a combination thereof. For example, the remaining life of the battery can be estimated using at least one of a current direction-based learning model for estimating remaining battery life, a charge amount-based learning model for estimating remaining battery life, a usage period-based learning model for estimating remaining battery life, or a combination thereof.

[0151] Meanwhile, the learning system (100) according to the present invention requires high-performance computing power because it learns labeled learning data to create a learning model for estimating the remaining battery life.

[0152] However, since the battery remaining life estimation system (300) estimates the remaining life of a battery using a learning model for battery remaining life estimation generated by the learning system (100), it can be executed even with relatively low computing power. Accordingly, the battery remaining life estimation engine (300a) according to the present invention can be directly provided to a user's application in the form of an application program.

[0153] Therefore, the battery remaining life estimation engine (300a) can be installed in a user application and used to locally check the remaining battery life in real time. Of course, a more accurate remaining life estimation is also possible by continuously estimating the remaining battery life over multiple stabilization periods.

[0154] FIG. 11 is a flowchart illustrating a procedure for labeling and learning learning data according to one embodiment of the present invention and a procedure for estimating the remaining life of an actual battery using a learning model for estimating the remaining life of a battery generated through the learning.

[0155] As illustrated in FIG. 11, the procedure for creating a learning model for estimating the remaining life of a battery by labeling learning data and training a deep neural network according to an embodiment of the present invention first includes: a learning system (100) selects measurement values ​​collected during a stable period of a battery from measurement values ​​stored in a database (200) and calculates a cycle count therefrom (S110). Feature values ​​to be used for learning the learning model are extracted from measurement data collected from the measurement values, including the calculated cycle count (S110).

[0156] Next, learning data is constructed using the extracted feature values ​​(S120). The learning data is obtained by converting the extracted feature values ​​into a format according to a deep neural network.

[0157] Next, the remaining life (RUL) is extracted based on the above-described cycle count, and the learning data is labeled with the extracted remaining life (RUL) (S130).

[0158] Next, the learning system (100) trains a deep neural network using the labeled learning data to create a learning model for estimating the remaining battery life (S140).

[0159] The learning model for estimating the remaining battery life thus generated is stored and managed in a database (200).

[0160] In addition, as a procedure for estimating the remaining life of a battery installed in an actual user's application using a learning model for estimating the remaining life of a battery, first, the battery remaining life estimation system (300) receives a measurement value for the corresponding battery from the user's application (S210).

[0161] Next, the battery remaining life estimation system (300) converts the received actual measurement value into input data to configure input data for inputting the learning model for estimating the battery remaining life (S220).

[0162] Next, the battery remaining life estimation system (300) inputs the above input data into a learning model for estimating the battery remaining life to estimate the RUL label (S230).

[0163] After this, the battery status estimation system (300) estimates the remaining life of the battery using the estimated RUL label (S240).

[0164] As described above, the present invention has the effect of automatically labeling learning data of a deep neural network for estimating the remaining useful lifetime (RUL) of a battery by collecting cycle count information from the battery as measurement data in an actual use environment.

[0165] In addition, the present invention has the effect of generating a high-performance battery remaining life estimation learning model by training the deep neural network through the learning data labeling method of the deep neural network.

[0166] In addition, the present invention has the effect of enabling efficient maintenance and management of the battery by estimating the exact remaining life of the battery and knowing the timing of battery replacement in advance.

[0167] In addition, the present invention has the effect of enabling stable and continuous use of an application by estimating the exact remaining life of a battery and knowing the timing of battery replacement in advance, thereby replacing the battery at the right time.

[0168] 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.

[0169] As described above, the present invention can automatically label training data of a deep neural network for estimating the remaining useful lifetime (RUL) of a battery by collecting cycle count information from a battery as measurement data in an actual use environment, and can train the deep neural network through the training data labeling method of the deep neural network to generate a high-performance battery remaining useful lifetime estimation training model, and can estimate the accurate remaining useful lifetime of the battery to know the timing of battery replacement in advance, thereby enabling efficient maintenance and management of the battery, and can estimate the accurate remaining useful lifetime of the battery to know the timing of battery replacement in advance, thereby enabling stable and continuous use of an application by replacing the battery at the right time, and thus has industrial applicability.

Claims

1. A learning data configuration step that constructs learning data by collecting measurement data about the battery at runtime and converting it into learning data; and It includes a learning data labeling step of labeling the above-configured learning data with the remaining useful lifetime (RUL) according to the cycle count of the battery; A method for labeling learning data of a deep neural network learning model for estimating the remaining life of a battery, characterized in that the remaining life is extracted by subtracting the collected cycle count (useful cycle count) from the maximum useful cycle count of the battery, and the labeling is configured to automatically assign the value of the extracted count as a label of the learning data.

2. In claim 1, The above learning data includes feature values including the open circuit voltage (OCV), voltage (V), current (I), charge (Q), temperature (T) of the battery, or a combination thereof, or Including statistical values including the mean, variance, median, standard deviation, difference, or a combination thereof of the above feature values, or Including the capacity change rate, temperature change rate, internal resistance change rate or a combination thereof of the battery extracted from the above characteristic values, or A method for labeling training data of a deep neural network learning model for estimating the remaining life of a battery, characterized by further including a combination of these.

3. In claim 1, A method for labeling learning data of a deep neural network learning model for estimating the remaining life of a battery, characterized in that the maximum useful cycle count is set to have a predetermined margin with respect to the maximum lifetime cycle count of the battery.

4. In claim 1, The above learning data configuration step is: A measurement data collection step that collects measurement data by filtering or normalizing the measurement values measured from the battery at run time; A feature value extraction step for extracting feature values to constitute learning data from the collected measurement data; and It includes a learning data conversion step for converting the extracted feature values into learning data in a format for training the deep neural network; A method for labeling learning data of a deep neural network learning model for estimating the remaining life of a battery, characterized in that the above-mentioned feature value extraction step calculates or extracts the cycle count of the battery from the collected measurement data and extracts the remaining life (RUL) from the calculated or extracted cycle count.

5. In claim 1, The above cycle count is a current direction-based cycle count, The above current direction-based cycle count is extracted by accumulating and counting the number of times there is no change in the current direction flowing through the battery for a predetermined period of time at run time. A method for labeling learning data of a deep neural network learning model for estimating the remaining life of a battery, characterized in that the remaining useful life (RUL) of the battery is extracted by subtracting the current direction-based cycle count extracted from the run time from the maximum current direction-based cycle count, which is predetermined as the maximum useful cycle count based on the current direction of the battery.

6. In claim 1, The above cycle count is a capacity-based cycle count, The above charge-based cycle count is extracted by adding up a predetermined charge amount that is charged and discharged from the battery at run time and then accumulating and counting the number of times the predetermined charge amount is reached. A method for labeling learning data of a deep neural network learning model for estimating the remaining life of a battery, characterized in that the remaining useful life (RUL) of the battery is extracted by subtracting the charge-based cycle count extracted from the run-time from a predetermined maximum charge-based cycle count (maximum useful cycle count) of the battery.

7. In claim 1, The above cycle count is a use time-based cycle count. The above usage time-based cycle count is extracted by repeatedly counting a predetermined usage time interval during which the battery is used at run-time. A method for labeling learning data of a deep neural network learning model for estimating the remaining life of a battery, characterized in that the remaining life (RUL) of the battery is extracted by subtracting the use time-based cycle count extracted from the run time from the maximum useful cycle count based on the battery's use time, which is predetermined as the maximum useful cycle count based on the battery's use time.

8. A learning method for generating a deep neural network learning model for estimating the remaining life of a battery, characterized in that it comprises generating a deep neural network learning model for estimating the remaining life of the battery through a learning data labeling method of a deep neural network learning model for estimating the remaining life of the battery according to at least one of claims 1 to 7.

9. A learning system for generating a deep neural network learning model for estimating the remaining life of a battery, characterized by comprising: a memory storing a program configured to generate a deep neural network learning model for estimating the life of a battery through a learning data labeling method of a deep neural network learning model for estimating the remaining life of a battery according to at least one of claims 1 to 7; and a processor configured to execute the program stored in the memory.

10. A method for estimating the remaining life of a battery, characterized in that it comprises estimating the remaining life of the battery by inputting measurement data measured from the battery as input data into a deep neural network learning model for estimating the life of the battery generated through a learning data labeling method of a deep neural network learning model for estimating the remaining life of the battery according to at least one of claims 1 to 7.

11. A battery remaining life estimation system, comprising: a memory storing a program configured to estimate the remaining life of a battery by inputting measurement data measured from the battery as input data into a deep neural network learning model for estimating the life of the battery generated through a learning data labeling method of a deep neural network learning model for estimating the remaining life of the battery according to at least one of claims 1 to 7; and a processor configured to execute the program stored in the memory.

Images