Learning model monitoring device and its operation method

Abstract

A learning model monitoring device according to one embodiment disclosed herein may include: an acquisition unit that acquires time-series data of a battery; an extraction unit that extracts at least one feature value corresponding to at least one feature from each of a plurality of sub-time-series data that constitute the time-series data and correspond to different cycles; a learning unit that trains a state diagnostic model capable of diagnosing the state of the battery based on the extracted at least one feature value; a first importance evaluation unit that evaluates the local importance of the at least one feature to the trained state diagnostic model for each cycle; and a second importance evaluation unit that evaluates the global importance of the at least one feature to the trained state diagnostic model based on the local importance.

Description

Technical Field

[0001] This application claims the benefit of priority based on Korean Patent Application No. 10-2023-0056575 filed on April 28, 2023, and all the contents disclosed in the literature of the Korean patent application are incorporated herein as part of this specification.

[0002] The embodiments disclosed in this document relate to a learning model monitoring apparatus and an operating method thereof.

Background Art

[0003] Recently, research and development on secondary batteries have been actively conducted. Here, a secondary battery is a battery that can be charged and discharged, and includes all conventional Ni / Cd batteries, Ni / MH batteries, etc., and recent lithium ion batteries. Among secondary batteries, lithium ion batteries have the advantage of having a much higher energy density compared to conventional Ni / Cd batteries, Ni / MH batteries, etc. Also, lithium ion batteries can be made small and lightweight, so they are used as power sources for mobile devices. Recently, their range of use as power sources for electric vehicles has expanded, and they are attracting attention as next-generation energy storage media.

[0004] For diagnosing the state of such batteries, a state diagnosis learning model based on machine learning is used. Generally, in order to ensure the explainability of such a model based on machine learning, a method of evaluating the importance of features, which are input factors of the model, and understanding its operating principle is used.

[0005] However, in most cases, methods are used to evaluate the global importance of features for models trained on time-series battery data. Such global importance can become unreliable when the characteristics of the battery data change (e.g., when the cause of battery anomalies changes). Therefore, in order to develop reliable trained models that can understand the results even when the characteristics of the battery data change, there is a need to develop a machine learning-based trained model monitoring system that can provide not only a global explanation for time-series data but also a local explanation for some data. [Overview of the project] [Problems that the invention aims to solve]

[0006] The embodiments disclosed herein provide a learning model monitoring device and a method for operating it that can evaluate the local importance of features in a state diagnostic learning model for each battery cycle and evaluate the global importance of the features based on the evaluated local importance.

[0007] The embodiments disclosed herein can provide a learning model monitoring device and a method for operating the same that can evaluate the causal relationships between factors in a state diagnostic learning model.

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

[0009] A learning model monitoring device according to one embodiment disclosed herein may include: an acquisition unit that acquires time-series data of a battery; an extraction unit that extracts at least one feature value corresponding to at least one feature from each of a plurality of sub-time-series data that constitute the time-series data and correspond to different cycles; a learning unit that trains a state diagnostic model capable of diagnosing the state of the battery based on the extracted at least one feature value; a first importance evaluation unit that evaluates the local importance of the at least one feature to the trained state diagnostic model for each cycle; and a second importance evaluation unit that evaluates the global importance of the at least one feature to the trained state diagnostic model based on the local importance.

[0010] In a learning model monitoring device according to one embodiment disclosed in this document, the first importance evaluation unit can calculate the Shapley value of at least one feature for the learned state diagnostic model and evaluate the local importance based on the calculated Shapley value.

[0011] In a learning model monitoring device according to one embodiment disclosed in this document, the second importance evaluation unit can evaluate the global importance for each of the at least one feature by averaging a plurality of local importances in the cycle.

[0012] A learning model monitoring device according to one embodiment disclosed herein further includes a causality evaluation unit that evaluates causal relationships between factors of the learned state diagnostic model, wherein the factors may include the state diagnostic results of the battery and / or the at least one feature.

[0013] In a learning model monitoring device according to one embodiment disclosed in this document, the causality evaluation unit can evaluate the causal relationship between the state diagnosis result and at least one feature among the factors.

[0014] In a learning model monitoring device according to one embodiment disclosed in this document, the causality evaluation unit may further include an analysis unit that evaluates the causal relationship between at least one feature, determines a second feature that is the cause of the first feature which has the highest global importance among the at least one feature, and analyzes the relationship between the first feature and the second feature.

[0015] In a learning model monitoring device according to one embodiment disclosed herein, the time-series data may include time-series data for at least one of the voltage, current, or temperature of the battery.

[0016] A learning model monitoring device according to one embodiment disclosed in this document may further include a preprocessing unit that classifies the cycles in the time-series data according to whether the battery has been charged or discharged, and extracts the plurality of partial time-series data.

[0017] A learning model monitoring method according to one embodiment disclosed herein may include: acquiring time-series data of a battery; extracting at least one feature value corresponding to at least one feature from each of a plurality of sub-time-series data that constitute the time-series data and correspond to different cycles; training a state diagnostic model capable of diagnosing the state of the battery based on the extracted at least one feature value; evaluating the local importance of the at least one feature to the learned state diagnostic model for each cycle; and evaluating the global importance of the at least one feature to the learned state diagnostic model based on the local importance.

[0018] In a learning model monitoring method according to one embodiment disclosed herein, the operation to evaluate local importance may include the operation to calculate a Shapley value for at least one feature of the learned state diagnostic model, and the operation to evaluate local importance based on the calculated Shapley value.

[0019] In a learning model monitoring method according to one embodiment disclosed herein, the operation to evaluate the global importance may include an operation to evaluate the global importance for each of the at least one feature by averaging a plurality of local importances in the cycle.

[0020] A learning model monitoring method according to one embodiment disclosed herein further includes an operation to evaluate the causal relationships between factors of the learned state diagnostic model, wherein the factors may include the battery state diagnostic result and / or the at least one feature.

[0021] In a learning model monitoring method according to one embodiment disclosed herein, the operation for evaluating the causal relationship may include an operation for evaluating the causal relationship between the state diagnosis result and at least one feature among the factors.

[0022] In a learning model monitoring method according to one embodiment disclosed herein, the operation to evaluate causal relationships includes an operation to evaluate the causal relationship between at least one feature and to determine a second feature that causes the first feature having the highest global importance among the at least one feature, and the learning model monitoring method may further include an operation to analyze the relationship between the first feature and the second feature.

[0023] A learning model monitoring method according to one embodiment disclosed in this document may further include the operation of classifying the cycles in the time series data according to whether the battery has been charged or discharged, and extracting the plurality of partial time series data. [Effects of the Invention]

[0024] According to the embodiments disclosed in this document, it is possible to provide not only a global description of the time-series data of the learning model for diagnosing the state of the battery, but also a local description of some of the data, thereby improving the reliability of the results of the learning model even when the battery data characteristics change.

[0025] According to the embodiments disclosed in this document, it is possible to confirm causal relationships between factors in a state diagnostic learning model that cannot be confirmed by feature importance evaluation alone.

[0026] Furthermore, various other effects can be understood directly or indirectly from this document. [Brief explanation of the drawing]

[0027] [Figure 1] This is a block diagram of a learning model monitoring device according to one embodiment. [Figure 2] This is a block diagram of the evaluation unit in a learning model monitoring device according to one embodiment. [Figure 3] This figure illustrates an example of how a learning model monitoring device according to one embodiment extracts multiple partial time series data from battery time series data acquired by that device. [Figure 4] This figure illustrates an example of how a learning model monitoring device according to one embodiment extracts feature values ​​from each of multiple partial time series data. [Figure 5] This figure illustrates an example of how a learning model monitoring device according to one embodiment evaluates the local importance of features for a state diagnostic model in each cycle. [Figure 6] This graph shows the local importance and global importance evaluated by a learning model monitoring device according to one embodiment. [Figure 7]This figure illustrates an example of a learning model monitoring device according to one embodiment that evaluates the causal relationships between factors in a state diagnosis model. [Figure 8] This is a flowchart showing the operation of a learning model monitoring device according to one embodiment. [Figure 9] This is a flowchart showing the operation of a learning model monitoring device according to one embodiment. [Figure 10] This is a flowchart showing the operation of a learning model monitoring device according to one embodiment. [Figure 11] This is a flowchart showing the operation of a learning model monitoring device according to one embodiment. [Modes for carrying out the invention]

[0028] Various embodiments of the present invention are described below with reference to the accompanying drawings. However, this should be understood not as limiting the present invention to any particular embodiment, but rather as including various modifications, equivalents, and / or alternatives to the embodiments of the present invention.

[0029] The various embodiments and terminology used in this document should be understood not to limit the technical features described herein to any particular embodiment, but to include various modifications, equivalents, or substitutions of such embodiments. In relation to the description of the drawings, similar reference numerals may be used for similar or related components. The singular form of a noun corresponding to an item may include one or more of such items unless the context clearly indicates otherwise.

[0030] In this document, each of the phrases such as “A or B,” “at least one of A and B,” “at least one of A or B,” “A, B or C,” “at least one of A, B and C,” and “at least one of A, B, or C” may include any one of the items listed together in the phrase, or any possible combination thereof. Terms such as “first,” “second,” “primary,” “second,” “A,” “B,” “(a),” or “(b)” may be used simply to distinguish one component from other components and, in particular, do not limit the component in other aspects (e.g., importance or order) unless otherwise stated.

[0031] In this document, when a component (e.g., component 1) is referred to as being "coupled," "joined," or "connected" to another component (e.g., component 2), with or without such terms, it means that the component can be connected to the other component directly (e.g., by wire), wirelessly, or via a third component.

[0032] According to various embodiments, each of the above-described components (e.g., modules or programs) may include one or more individuals, some of which may be separated and arranged in other components. According to various embodiments, one or more of the above-described components or operations may be omitted, or one or more other components or operations may be added. Generally or further, multiple components (e.g., modules or programs) may be integrated into a single component. In such cases, the integrated component may perform one or more functions of each of the multiple components in the same or similar manner as they were performed by the components of the multiple components before the integration. According to various embodiments, operations performed by modules, programs or other components may be performed sequentially, in parallel, repeatedly, or heuristically, or one or more of the operations may be performed in a different order, omitted, or one or more other operations may be added.

[0033] The configuration of the data processing device will be described below with reference to Figures 1 and 2.

[0034] Figure 1 is a block diagram of a learning model monitoring device according to one embodiment. Figure 2 is a block diagram of the evaluation unit within the learning model monitoring device according to one embodiment.

[0035] Referring to Figure 1, the learning model monitoring device 101 can be connected to the electronic device 103 and the user terminal 105 by wire and / or wirelessly.

[0036] According to one embodiment, the connection 104 between the learning model monitoring device 101 and the electronic device 103 can be a communication connection via a wired and / or wireless network. In one embodiment, the wired network can be based on LAN (local area network) communication or power line communication. In one embodiment, the wireless network can be based on a short-range communication network (e.g., Bluetooth®, WiFi (wireless fidelity), or IrDA (infrared data association)) or a long-range communication network (cellular network, 4G network, 5G network).

[0037] According to other embodiments, the connection 104 between the learning model monitoring device 101 and the electronic device 103 can be a connection via a communication method between devices (e.g., bus, GPIO (general purpose input and output), SPI (serial peripheral interface), or MIPI (mobile industry processor interface)).

[0038] According to one embodiment, the connection 106 between the learning model monitoring device 101 and the user terminal 105 can be a communication connection via a wired and / or wireless network.

[0039] According to one embodiment, the electronic device 103 can be a mobile device (e.g., a mobile phone, a laptop computer, a smartphone, a smartpad), an electric vehicle (e.g., an EV (electric vehicle), an HEV (hybrid EV), a PHEV (plug-in HEV), an FCEV (fuel cell EV)), an energy storage system (ESS), or a battery swapping system (BSS).

[0040] According to one embodiment, the electronic device 103 may include one or more batteries 111, 113, 115. Each of the one or more batteries 111, 113, 115 may be a battery cell, a battery module, a battery pack, or a battery rack.

[0041] According to one embodiment, the user terminal 105 can be a mobile device (e.g., a mobile phone, a laptop computer, a smartphone, a smartpad) or a PC (personal computer).

[0042] According to one embodiment, the learning model monitoring device 101 may include a communication circuit 120, a sensor 130, a memory 140, and a processor 150. Depending on the embodiment, the learning model monitoring device 101 illustrated in Figure 1 may further include at least one component other than the components illustrated in Figure 1 (e.g., a display, an input device, or an output device).

[0043] According to one embodiment, the communication circuit 120 creates a wired communication channel and / or wireless communication channel between the learning model monitoring device 101 and the electronic device 103 and / or the user terminal 105, and can send and receive data with the electronic device 103 and / or the user terminal 105 via the created communication channel.

[0044] According to one embodiment, the sensor 130 can acquire values ​​related to the state of the batteries 111, 113, and 115 of the electronic device 103. In one embodiment, the state-related values ​​may represent one or more values ​​for the voltage, current, resistance, state of charge (SOC), state of health (SOH), or temperature of the batteries 111, 113, and 115, or a combination thereof. Hereinafter, the state-related values ​​may be referred to as "state values".

[0045] According to one embodiment, the memory 140 may include volatile memory and / or non-volatile memory.

[0046] According to one embodiment, the memory 140 can store data used by at least one component of the learning model monitoring device 101 (e.g., the processor 150). For example, the data may include software (or related instructions), input data, or output data. In one embodiment, the instructions may cause the learning model monitoring device 101 to perform an operation defined by the instructions when executed by the processor 150.

[0047] According to one embodiment, the memory 140 may include one or more software (e.g., acquisition unit 141, preprocessing unit 142, extraction unit 143, learning unit 144, evaluation unit 145, and analysis unit 146).

[0048] Referring to Figure 2, the evaluation unit 145 may include a first importance evaluation unit 145-1, a second importance evaluation unit 145-2, and / or a causality evaluation unit 145-3.

[0049] Referring also to Figure 1, the processor 150 may include a central processing unit, an application processor, a graphics processing unit, a neural network processing unit (NPU), an image signal processor, a sensor hub processor, or a communication processor.

[0050] According to one embodiment, the processor 150 can execute software (e.g., acquisition unit 141, preprocessing unit 142, extraction unit 143, learning unit 144, evaluation unit 145, and analysis unit 146) to control at least one other component (e.g., hardware or software component) of the learning model monitoring device 101 connected to the processor 150, and perform various data processing or calculations.

[0051] The following describes how the learning model monitoring device 101 monitors a state diagnostic model that can diagnose the state of batteries 111, 113, or 115, via an acquisition unit 141, a preprocessing unit 142, an extraction unit 143, a learning unit 144, an evaluation unit 145, and an analysis unit 146, with reference to Figures 3, 4, 5, 6, and 7.

[0052] Figure 3 illustrates an example of how a learning model monitoring device according to one embodiment extracts multiple partial time series data from battery time series data acquired by that device. Figure 4 illustrates an example of how a learning model monitoring device according to one embodiment extracts feature values ​​for each of the multiple partial time series data. Figure 5 illustrates an example of how a learning model monitoring device according to one embodiment evaluates the local importance of features for the state diagnostic model for each cycle. Figure 6 is a graph showing the local and global importance evaluated by a learning model monitoring device according to one embodiment. Figure 7 illustrates an example of how a learning model monitoring device according to one embodiment evaluates the causal relationships between factors in the state diagnostic model.

[0053] Referring to Figure 3, you can see graphs 310 and 320 showing time-series data for batteries 111, 113, or 115.

[0054] According to one embodiment, the acquisition unit 141 can acquire time-series data of batteries 111, 113, or 115, as shown in graph 310. Here, the time-series data may include time-series data for at least one of the following: voltage, current, or temperature of batteries 111, 113, or 115. Therefore, the y-axis of graphs 310 and 320 can show the state values ​​(voltage, current, or temperature) of batteries 111, 113, or 115.

[0055] According to one embodiment, the preprocessing unit 142 can extract a plurality of partial time-series data from the time-series data. Here, the time-series data can consist of the plurality of partial time-series data. In one embodiment, as shown in graph 320, the preprocessing unit 142 extracts cycles C1, C2, C3, C4, ..., and C in the time-series data according to whether or not batteries 111, 113, or 115 are charged or discharged. n (where n is a natural number) can be classified. Here, cycles C1, C2, C3, C4, ..., and C n These may include different charging, discharging, and resting periods. The pre-processing unit 142 processes cycles C1, C2, C3, C4, ..., and C n Multiple partial time series data corresponding to each can be extracted. For example, the preprocessor 142 extracts the first partial time series data showing the battery status value in the first cycle C1, the second partial time series data showing the battery status value in the second cycle C2, the third partial time series data showing the battery status value in the third cycle C3, the fourth partial time series data showing the battery status value in the fourth cycle C4, and the nth cycle C n It is possible to extract the nth part of the time series data showing the battery status value.

[0056] According to one embodiment, the extraction unit 143 can extract at least one feature value for each of the plurality of partial time series data that constitute the time series data and correspond to different cycles. According to one embodiment, the extraction unit 143 can set at least one feature related to the battery status value (e.g., mean, maximum, minimum, gradient, standard deviation, standard score, etc. of the battery status value). The extraction unit 143 can extract at least one feature value corresponding to the at least one feature for each of the plurality of partial time series data.

[0057] Referring to Figure 4, the extraction unit 143 extracts multiple features X1, X2, X3, X4, ..., X pcan be set. According to one embodiment, the extraction unit 143 processes a plurality of cycles C1, C2, C3, C4, ···, and C n For each of the corresponding plurality of partial time series data, a plurality of features X1, X2, X3, X4, …, X p The corresponding plurality of feature values can be extracted. As an example, the extraction unit 143, like the dataset 400, based on the first partial time series data corresponding to the first cycle C1, a plurality of features X1, X2, X3, X4, …, X p The corresponding plurality of feature values F 1-1 , F 2-1 , F 3-1 , F 4-1 , ···, F p-1 can be extracted.

[0058] For example, the first feature X1 is the average value of the battery state value, the second feature X2 is the maximum value of the battery state value, the third feature X3 is the minimum value of the battery state value, the fourth feature X4 is the gradient of the battery state value, and the p-th feature X p can be the standard deviation of the battery state value. In this case, the extraction unit 143, based on the first partial time series data, takes the average value of the battery state value in the first cycle C1 as the first - 1 feature value F 1-1 , takes the maximum value of the battery state value in the first cycle C1 as the second - 1 feature value F 2-1 , takes the minimum value of the battery state value in the first cycle C1 as the third - 1 feature value F 3-1 , takes the gradient of the battery state value in the first cycle C1 as the fourth - 1 feature value F 4-1 , and takes the standard deviation of the battery state value in the first cycle C1 as the p - 1 feature value F p-1 and extract them. The extraction unit 143, in the manner as described above, from the plurality of partial time series data corresponding to the plurality of cycles C1, C2, C3, C4, ···, and C n extracts a plurality of feature values F 1-1 , F 2-1 [[ID=3​​​​​​​​​According to one embodiment, the learning unit 144 can be trained to learn a state diagnostic model that can diagnose the state of battery 111, 113, or 115 based on at least one feature value extracted by the extraction unit 143. Here, the state diagnostic model can be a machine learning-based learning model that outputs a state diagnostic result for battery 111, 113, or 115 based on input factors related to the state values ​​of battery 111, 113, or 115.

[0060] According to one embodiment, the learning unit 144 can learn a state diagnostic model based on the state diagnostic results of batteries 111, 113, or 115 for each cycle, in addition to the at least one feature value. For example, the learning unit 144 can learn a state diagnostic model based on the dataset 400 shown in Figure 4. The state diagnostic results Y1, Y2, Y3, Y4, ..., Y for batteries 111, 113, or 115 for each cycle shown in the dataset 400 n This may be a result of a diagnosis performed in advance by the electronic device 103, or a result of a diagnosis performed in advance by the learning model monitoring device 101 based on multiple partial time series data. According to one embodiment, the state diagnosis model can learn the relationship between multiple feature values ​​and the diagnosis result for each cycle. For example, the state diagnosis model learns the relationship between multiple feature values ​​F in the first cycle C1. 1-1 F 2-1 F 3-1 F 4-1 , , , F p-1 The relationship between this and the first diagnostic result Y1 can be learned.

[0061] According to one embodiment, the first importance evaluation unit 145-1 can evaluate the local importance of at least one feature of the state diagnostic model learned by the learning unit 144 for each cycle.

[0062] Referring to Figure 5, the first importance evaluation unit 145-1 evaluates multiple cycles C1, C2, C3, C4, ..., and C, as in the dataset 500. nFor each of these, multiple features X1, X2, X3, X4, ..., and X for the state diagnostic model. p Local importance I 1-1 , I 2-1 , I 3-1 , I 4-1 , , , and I p-n This can be evaluated. Here, each of the local importance levels corresponds to a specific cycle C1, C2, C3, C4, ..., or C n Within, specific features X1, X2, X3, X4, ..., or X p This could mean an impact on the performance of the diagnostic model.

[0063] According to one embodiment, the first importance evaluation unit 145-1 determines that the state diagnosis model has multiple feature values ​​F in the first cycle C1. 1-1 F 2-1 F 3-1 F 4-1 , , , F p-1 The process of learning the relationship between and the first diagnostic result Y1, or the state diagnostic model that has learned the relationship, can be analyzed. Based on the analysis, the first importance evaluation unit 145-1 evaluates multiple features X1, X2, X3, X4, ..., X in the first cycle C1. p Each local importance level I 1-1 , I 2-1 , I 3-1 , I 4-1 , , , I p-1 This can be evaluated. For example, the first importance evaluation unit 145-1 evaluates multiple features X1, X2, X3, X4, ..., X for the state diagnosis model based on the analysis. p Each Shapley value can be calculated. The first importance evaluation unit 145-1 evaluates multiple features X1, X2, X3, X4, ..., X based on the calculated Shapley values. p Each local importance level I 1-1 , I 2-1 , I 3-1 , I 4-1 , , , I p-1 This can be evaluated. Here, the first importance evaluation unit 145-1 evaluates the calculated Shapley value to the local importance I1-1 , I 2-1 , I 3-1 , I 4-1 , , , I p-1 It can also be determined as follows. The first importance evaluation unit 145-1 determines multiple cycles C1, C2, C3, C4, ..., and C n Each of these has multiple features X1, X2, X3, X4, ..., and X for the state diagnostic model. p Local importance I 1-1 , I 2-1 , I 3-1 , I 4-1 , , , and I p-n This can be evaluated.

[0064] According to one embodiment, the second importance evaluation unit 145-2 can evaluate the global importance of at least one feature for the state diagnosis model learned by the learning unit 144, based on the local importance evaluated by the first importance evaluation unit 145-1.

[0065] According to one embodiment, the second importance evaluation unit 145-2 evaluates a plurality of features X1, X2, X3, X4, ..., and X p Each of the multiple cycles C1, C2, C3, C4, ..., and C n Multiple local importance levels I 1-1 , I 2-1 , I 3-1 , I 4-1 , , , I p-n Averaging these, multiple features X1, X2, X3, X4, ..., and X for the state diagnostic model are obtained. p The global importance of can be assessed. Here, global importance is the overall cycle C1, C2, C3, C4, ..., and C n In a specific feature X1, X2, X3, X4, ..., or X p This could mean an impact on the performance of the diagnostic model.

[0066] For example, the second importance evaluation unit 145-2 evaluates multiple cycles C1, C2, C3, C4, ..., and Cn The plurality of local importance degrees I of the first feature X1 in 1-1 , I 1-2 , I 1-3 , I 1-4 , ···, I 1-n can be evaluated as the global importance degree of the first feature X1 by averaging the values. As another example, the second importance degree evaluation unit 145-2 includes a plurality of cycles C1, C2, C3, C4, ···, and C n The plurality of local importance degrees I of the second feature X2 in 2-1 , I 2-2 , I 2-3 , I 2-4 , ···, I 2-n can be evaluated as the global importance degree of the second feature X2 by averaging the values. The second importance degree evaluation unit 145-2 can evaluate the global importance degrees of a plurality of features X1, X2, X3, X4, ···, and X p in the manner as described above.

[0067] Referring to FIG. 6, the learning model monitoring device 101 can show the transition of the local importance degree for each feature evaluated by the first importance degree evaluation unit 145-1 as shown in graph 610. Also, the learning model monitoring device 101 can show the global importance degree for each feature evaluated by the second importance degree evaluation unit 145-2 as shown in graph 620. For example, the learning model monitoring device 101 can display the graph (610 and / or 620) via at least one component (e.g., a display or an output device) included in the learning model monitoring device 101 or via an external device.

[0068] According to one embodiment, the causality evaluation unit 145-3 can evaluate the causal relationships between factors of the state diagnosis model learned by the learning unit 144. Here, the factors may include the state diagnosis results and / or at least one feature of the battery 111, 113, or 115. For example, the state diagnosis result is the state diagnosis result Y included in the dataset 400 (or 500) of Figure 4 (or Figure 5), and at least one feature is a plurality of features X1, X1, X3, X4, ..., and X included in the dataset 400 (or 500) of Figure 4 (or Figure 5). p It can be.

[0069] According to one embodiment, the causality evaluation unit 145-3 can evaluate causal relationships between factors using a causal discovery algorithm. Here, the causal discovery algorithm can include various algorithms for deriving causal relationships between learning model factors (e.g., a constraint-based causal discovery algorithm, a scored-based causal discovery algorithm, a continuous optimization method, etc.).

[0070] Referring to Figure 7, the causality evaluation unit 145-3 can evaluate the causal relationships between factors X1, X2, X3, X4, X5, and Y of the state diagnosis model learned by the learning unit 144.

[0071] According to one embodiment, the causality evaluation unit 145-3 can evaluate the causal relationship between the state diagnosis result and at least one feature among the factors. For example, as shown in Figure 7, the causality evaluation unit 145-3 can evaluate the first feature X1, the second feature X2, and the third feature X3 as factors that cause the state diagnosis result Y.

[0072] According to one embodiment, the causality evaluation unit 145-3 can evaluate the causal relationship between at least one feature. For example, as shown in Figure 7, the causality evaluation unit 145-3 can evaluate the second feature X2 and the third feature X3 as factors causing the first feature X1, the fourth feature X4 as a factor causing the second feature X2, the fifth feature X5 as a factor causing the third feature X3, and the fifth feature X5 and the sixth feature X6 as factors causing the first feature X4.

[0073] According to one embodiment, the learning model monitoring device 101 can display the causal relationships evaluated by the causality evaluation unit 145-3 via at least one component included in the learning model monitoring device 101 (e.g., a display or output device) or via an external device.

[0074] According to one embodiment, the analysis unit 146 can analyze the relationship between features (e.g., second feature X2 and / or third feature X3) that cause the feature with the highest global importance among the at least one feature (e.g., first feature X1), based on the causal relationship between at least one feature evaluated by the causality evaluation unit 145-3. For example, the analysis unit 146 can quantitatively analyze the changes in the first feature X1 due to numerical changes in the second feature X2 and / or third feature X3.

[0075] According to one embodiment, the learning model monitoring device 101 can display the analysis results of the analysis unit 146 via at least one component included in the learning model monitoring device 101 (e.g., a display or output device) or via an external device.

[0076] Figure 8 is a flowchart of the operation of a learning model monitoring device according to one embodiment. Figure 8 can be used to explain the operation of the learning model monitoring device 101 in Figure 1, and can be explained using the configuration of Figure 1 (e.g., learning model monitoring device 101).

[0077] The embodiment shown in Figure 8 is just one embodiment, and the order of steps in various embodiments of the present invention may differ from that shown in Figure 8. Some of the steps shown in Figure 8 may be omitted, the order of the steps may be changed, or steps may be merged.

[0078] Referring to Figure 8, in operation 805, the learning model monitoring device 101 can acquire time-series data of batteries 111, 113, or 115. Here, the time-series data may include time-series data for at least one of the following: voltage, current, or temperature of batteries 111, 113, or 115.

[0079] In operation 810, the learning model monitoring device 101 can extract at least one feature value for each of the multiple partial time series data. Here, the multiple partial time series data constitute time series data and can correspond to different cycles. According to one embodiment, the learning model monitoring device 101 can set at least one feature related to the battery status value (e.g., mean, maximum, minimum, gradient, standard deviation, standard score, etc. of the battery status value). The learning model monitoring device 101 can extract at least one feature value corresponding to the at least one feature for each of the multiple partial time series data.

[0080] In operation 815, the learning model monitoring device 101 can train a state diagnostic model capable of diagnosing the state of battery 111, 113, or 115 based on at least one feature value extracted in operation 810. Here, the state diagnostic model can be a machine learning-based learning model that outputs a state diagnostic result for battery 111, 113, or 115 based on input factors related to the state values ​​of battery 111, 113, or 115.

[0081] According to one embodiment, the learning model monitoring device 101 can learn a state diagnostic model based on the state diagnostic results of batteries 111, 113, or 115 for each cycle, in addition to the at least one feature value. The state diagnostic results of batteries 111, 113, or 115 for each cycle may be results previously diagnosed by the electronic device 103, or results previously diagnosed by the learning model monitoring device 101 based on a plurality of partial time series data. According to one embodiment, the state diagnostic model can learn the relationship between at least one feature value and the diagnostic result for each cycle.

[0082] In operation 820, the learning model monitoring device 101 can evaluate the local importance of at least one feature for the state diagnostic model learned in operation 815, for each cycle. Here, each local importance may represent the influence that a particular feature has on the performance of the state diagnostic model's diagnosis within a particular cycle.

[0083] According to one embodiment, the learning model monitoring device 101 can analyze the process by which the state diagnostic model learns the relationship between at least one feature value in a particular cycle and the battery diagnostic result in the particular cycle, or the state diagnostic model that has learned the relationship. Based on the analysis, the learning model monitoring device 101 can evaluate the local importance of each of the at least one feature in the particular cycle. For example, based on the analysis, the learning model monitoring device 101 can calculate a Shapley value for each of the at least one feature in the state diagnostic model. Based on the calculated Shapley value, the learning model monitoring device 101 can evaluate the local importance of each of the at least one feature. Here, the learning model monitoring device 101 can also determine the calculated Shapley value as the local importance.

[0084] In operation 825, the learning model monitoring device 101 can evaluate the global importance of at least one feature for the state diagnostic model learned in operation 815, based on the local importance evaluated in operation 820. Here, global importance may represent the influence that a particular feature has on the performance of the state diagnostic model's diagnosis throughout the entire cycle.

[0085] According to one embodiment, the learning model monitoring device 101 can evaluate the global importance of at least one feature for the state diagnostic model by averaging at least one local importance over multiple cycles for each feature.

[0086] Figure 9 is a flowchart of the operation of a learning model monitoring device according to one embodiment. Figure 9 can be used to explain the operation of the learning model monitoring device 101 in Figure 1, and can be explained using the configuration of Figure 1 (e.g., learning model monitoring device 101).

[0087] The embodiment shown in Figure 9 is just one embodiment, and the order of steps in various embodiments of the present invention may differ from that shown in Figure 9. Some of the steps shown in Figure 9 may be omitted, the order of the steps may be changed, or steps may be merged.

[0088] Referring to Figure 9, in operation 905, the learning model monitoring device 101 can acquire time-series data of batteries 111, 113, or 115. Here, the time-series data may include time-series data for at least one of the following: voltage, current, or temperature of batteries 111, 113, or 115.

[0089] In operation 910, the learning model monitoring device 101 can extract multiple partial time series data from the time series data. Here, the time series data can consist of the multiple partial time series data. In one embodiment, the learning model monitoring device 101 can classify cycles in the time series data according to whether or not batteries 111, 113, or 115 are being charged or discharged. Here, the cycles can include different charging intervals, discharging intervals, and rest intervals. The learning model monitoring device 101 can extract multiple partial time series data corresponding to each cycle.

[0090] In operation 915, the learning model monitoring device 101 can extract at least one feature value for each of the multiple partial time series data extracted in operation 910. According to one embodiment, the learning model monitoring device 101 can set at least one feature related to the battery status value (e.g., mean, maximum, minimum, gradient, standard deviation, standard score, etc., of the battery status value). The learning model monitoring device 101 can extract at least one feature value corresponding to the at least one feature for each of the multiple partial time series data.

[0091] In operation 920, the learning model monitoring device 101 can train a state diagnostic model that can diagnose the state of battery 111, 113, or 115 based on at least one feature value extracted in operation 810. Here, the state diagnostic model can be a machine learning-based learning model that outputs a state diagnostic result for battery 111, 113, or 115 based on input factors related to the state values ​​of battery 111, 113, or 115.

[0092] According to one embodiment, the learning model monitoring device 101 can learn a state diagnostic model based on the state diagnostic results of batteries 111, 113, or 115 for each cycle, in addition to the at least one feature value. The state diagnostic results of batteries 111, 113, or 115 for each cycle may be results previously diagnosed by the electronic device 103, or results previously diagnosed by the learning model monitoring device 101 based on a plurality of partial time series data. According to one embodiment, the state diagnostic model can learn the relationship between at least one feature value and the diagnostic result for each cycle.

[0093] In operation 925, the learning model monitoring device 101 can evaluate the local importance of at least one feature for the state diagnostic model learned in operation 920, for each cycle. Here, each local importance may represent the influence that a particular feature has on the performance of the state diagnostic model's diagnosis within a particular cycle.

[0094] According to one embodiment, the learning model monitoring device 101 can analyze the process by which the state diagnostic model learns the relationship between at least one feature value in a particular cycle and the battery diagnostic result in the particular cycle, or the state diagnostic model that has learned the relationship. Based on the analysis, the learning model monitoring device 101 can evaluate the local importance of each of the at least one feature in the particular cycle. For example, based on the analysis, the learning model monitoring device 101 can calculate a Shapley value for each of the at least one feature in the state diagnostic model. Based on the calculated Shapley value, the learning model monitoring device 101 can evaluate the local importance of each of the at least one feature. Here, the learning model monitoring device 101 can also determine the local importance from the calculated Shapley value.

[0095] In operation 930, the learning model monitoring device 101 can evaluate the global importance of at least one feature for the state diagnostic model learned in operation 920, based on the local importance evaluated in operation 925. Here, global importance may represent the influence that a particular feature has on the performance of the state diagnostic model's diagnosis throughout the entire cycle.

[0096] According to one embodiment, the learning model monitoring device 101 can evaluate the global importance of at least one feature for the state diagnostic model by averaging at least one local importance over multiple cycles for each feature.

[0097] Figure 10 is a flowchart of the operation of a learning model monitoring device according to one embodiment. Figure 10 can be used to explain the operation of the learning model monitoring device 101 in Figure 1, and can be explained using the configuration of Figure 1 (e.g., learning model monitoring device 101).

[0098] The embodiment shown in Figure 10 is just one embodiment, and the order of steps in various embodiments of the present invention may differ from that shown in Figure 10. Some of the steps shown in Figure 10 may be omitted, the order of the steps may be changed, or steps may be merged.

[0099] Referring to Figure 10, in operation 1005, the learning model monitoring device 101 can acquire time-series data of batteries 111, 113, or 115. Here, the time-series data may include time-series data for at least one of the following: voltage, current, or temperature of batteries 111, 113, or 115.

[0100] In operation 1010, the learning model monitoring device 101 can extract at least one feature value for each of the multiple partial time series data. Here, the multiple partial time series data constitute time series data and can correspond to different cycles. According to one embodiment, the learning model monitoring device 101 can set at least one feature related to the battery status value (e.g., mean, maximum, minimum, gradient, standard deviation, standard score, etc. of the battery status value). The learning model monitoring device 101 can extract at least one feature value corresponding to the at least one feature for each of the multiple partial time series data.

[0101] In operation 1015, the learning model monitoring device 101 can train a state diagnostic model that can diagnose the state of battery 111, 113, or 115 based on at least one feature value extracted in operation 1010. Here, the state diagnostic model can be a machine learning-based learning model that outputs a state diagnostic result for battery 111, 113, or 115 based on input factors related to the state values ​​of battery 111, 113, or 115.

[0102] According to one embodiment, the learning model monitoring device 101 can learn a state diagnostic model based on the state diagnostic results of batteries 111, 113, or 115 for each cycle, in addition to the at least one feature value. The state diagnostic results of batteries 111, 113, or 115 for each cycle may be results previously diagnosed by the electronic device 103, or results previously diagnosed by the learning model monitoring device 101 based on a plurality of partial time-series data. According to one embodiment, the state diagnostic model can learn the relationship between at least one feature value and the diagnostic result for each cycle.

[0103] In operation 1020, the learning model monitoring device 101 can evaluate the local importance of at least one feature for the state diagnostic model learned in operation 1015, for each cycle. Here, each local importance may represent the influence that a particular feature has on the performance of the state diagnostic model's diagnosis within a particular cycle.

[0104] According to one embodiment, the learning model monitoring device 101 can analyze the process by which the state diagnostic model learns the relationship between at least one feature value in a particular cycle and the battery diagnostic result in the particular cycle, or the state diagnostic model that has learned the relationship. Based on the analysis, the learning model monitoring device 101 can evaluate the local importance of each of the at least one feature in the particular cycle. For example, based on the analysis, the learning model monitoring device 101 can calculate a Shapley value for each of the at least one feature in the state diagnostic model. Based on the calculated Shapley value, the learning model monitoring device 101 can evaluate the local importance of each of the at least one feature. Here, the learning model monitoring device 101 can also determine the calculated Shapley value as the local importance.

[0105] In operation 1025, the learning model monitoring device 101 can evaluate the global importance of at least one feature for the state diagnostic model learned in operation 1015, based on the local importance evaluated in operation 1020. Here, global importance may represent the influence that a particular feature has on the performance of the state diagnostic model's diagnosis throughout the entire cycle.

[0106] According to one embodiment, the learning model monitoring device 101 can evaluate the global importance of at least one feature for the state diagnostic model by averaging at least one local importance over multiple cycles for each feature.

[0107] In operation 1030, the learning model monitoring device 101 can evaluate the causal relationships between factors of the state diagnostic model learned in operation 1015. Here, the factors may include the state diagnostic results and / or at least one feature of the battery 111, 113, or 115.

[0108] According to one embodiment, the learning model monitoring device 101 can evaluate causal relationships between factors using a causal discovery algorithm. Here, the causal discovery algorithm can include various algorithms for deriving causal relationships between learning model factors (e.g., a constraint-based causal discovery algorithm, a scored-based causal discovery algorithm, a continuous optimization method, etc.).

[0109] According to one embodiment, the learning model monitoring device 101 can evaluate the causal relationship between the state diagnosis result and at least one feature among the factors. For example, as shown in Figure 7, the learning model monitoring device 101 can evaluate the first feature X1, the second feature X2, and the third feature X3 as factors that cause the state diagnosis result Y.

[0110] According to one embodiment, the learning model monitoring device 101 can evaluate the causal relationship between at least one feature. For example, as shown in Figure 7, the learning model monitoring device 101 can evaluate the second feature X2 and the third feature X3 as factors causing the first feature X1, the fourth feature X4 as a factor causing the second feature X2, the fifth feature X5 as a factor causing the third feature X3, and the fifth feature X5 and the sixth feature X6 as factors causing the first feature X4.

[0111] Figure 11 is a flowchart of the operation of a learning model monitoring device according to one embodiment. Figure 11 can be used to explain the operation of the learning model monitoring device 101 in Figure 1, and can be explained using the configuration of Figure 1 (e.g., learning model monitoring device 101).

[0112] The embodiment shown in Figure 11 is just one embodiment, and the order of steps in various embodiments of the present invention may differ from that shown in Figure 11. Some of the steps shown in Figure 11 may be omitted, the order of the steps may be changed, or steps may be merged.

[0113] Referring to Figure 11, in operation 1105, the learning model monitoring device 101 can determine a second feature that causes the first feature, which has the highest global importance among at least one feature. The learning model monitoring device 101 can determine the first and second features based on the global importance evaluation in operation 1025 and the causal relationship evaluation in operation 1030 in Figure 10.

[0114] In operation 1110, the learning model monitoring device 101 can analyze the relationship between the first and second features determined in operation 1105. For example, the learning model monitoring device 101 can quantitatively analyze the changes in the first feature due to changes in the numerical value of the second feature.

[0115] The terms "contains," "constitutes," or "possesses," as used above, should be interpreted as meaning that the component may be inherent, and not as excluding other components, but as potentially including other components, unless otherwise specified. All terms, including technical or scientific terms, have the same meaning as that generally understood by a person of ordinary skill in the art to which the embodiments disclosed herein belong, unless otherwise defined. Commonly used terms, such as those defined in dictionaries, should be interpreted in accordance with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined herein.

Claims

1. An acquisition unit that acquires time-series data of the battery, An extraction unit that constitutes the aforementioned time series data and extracts at least one feature value corresponding to at least one feature from each of a plurality of partial time series data corresponding to different cycles, A learning unit that trains a state diagnostic model capable of diagnosing the state of the battery based on the extracted at least one feature value, For each cycle, a first importance evaluation unit evaluates the local importance of at least one feature to the learned state diagnostic model, A learning model monitoring device comprising: a second importance evaluation unit that evaluates the global importance of the at least one feature to the learned state diagnostic model based on the local importance.

2. The first importance evaluation unit is, The Shapley value of at least one feature for the learned state diagnostic model is calculated. The learning model monitoring device according to claim 1, which evaluates the local importance based on the calculated Shapley value.

3. The second importance evaluation unit is: The learning model monitoring device according to claim 1, wherein for each of the at least one of the features, the global importance is evaluated by averaging a plurality of local importances in the cycle.

4. The system further includes a causality evaluation unit that evaluates the causal relationships between factors in the learned state diagnosis model, The learning model monitoring device according to claim 1, wherein the factor includes the battery state diagnosis result and / or the at least one feature.

5. The learning model monitoring device according to claim 4, wherein the causality evaluation unit evaluates the causal relationship between the state diagnosis result and at least one feature among the factors.

6. The causality evaluation unit evaluates the causal relationship between the at least one feature and determines the second feature that is the cause of the first feature which has the highest global importance among the at least one feature. The learning model monitoring device according to claim 4, further comprising an analysis unit that analyzes the relationship between the first feature and the second feature.

7. The learning model monitoring device according to claim 1, wherein the time-series data includes time-series data for at least one of the voltage, current, or temperature of the battery.

8. The learning model monitoring device according to claim 1, further comprising a preprocessing unit that classifies the cycles in the time series data according to whether the battery has been charged or discharged, and extracts the plurality of partial time series data.

9. The operation of acquiring time-series data of the battery, The operation of constructing the aforementioned time series data and extracting at least one feature value corresponding to at least one feature for each of the multiple partial time series data corresponding to different cycles, An operation to train a state diagnostic model capable of diagnosing the state of the battery based on the extracted at least one feature value, Each cycle includes an operation to evaluate the local importance of at least one feature for the learned state diagnostic model, A learning model monitoring method, comprising the operation of evaluating the global importance of the at least one feature to the learned state diagnostic model based on the local importance.

10. The operation to evaluate the local importance is, An operation to calculate the Shapley value of at least one feature for the learned state diagnostic model, The learning model monitoring method according to claim 9, further comprising the operation of evaluating the local importance based on the calculated Shapley value.

11. The learning model monitoring method according to claim 9, wherein the operation to evaluate the global importance includes an operation to evaluate the global importance by averaging a plurality of local importances in the cycle for each of the at least one feature.

12. The process further includes evaluating the causal relationships between factors in the learned state diagnosis model, The learning model monitoring method according to claim 9, wherein the factor includes the battery state diagnosis result and / or the at least one feature.

13. The learning model monitoring method according to claim 12, wherein the operation to evaluate the causal relationship includes an operation to evaluate the causal relationship between the state diagnosis result and at least one feature among the factors.

14. The operation to evaluate the causal relationship includes an operation to evaluate the causal relationship between at least one feature and to determine the second feature that causes the first feature which has the highest global importance among the at least one feature. The learning model monitoring method according to claim 12, further comprising the operation of analyzing the relationship between the first feature and the second feature.

15. The learning model monitoring method according to claim 9, further comprising the operation of classifying the cycles in the time series data according to whether the battery has been charged or discharged, and extracting the plurality of partial time series data.

Images