Safe-DNN model generation method for safely estimating battery state and device therefor

Abstract

The present invention relates to a safe-DNN model generation method for safely estimating a battery state and a device therefor, wherein a safe-DNN model is generated to estimate the state of a battery (SOX) more conservatively biased than an actual battery state (SOX), and the state of the battery is safely estimated in advance even in the event of sudden degradation or thermal runaway of the battery, thereby enabling the battery to be replaced or checked in advance before damage occurs.

Description

Method for generating a SAFE-DNN model for safely estimating battery status and device thereof

[0001] The present invention relates to a method and a device for generating a safe-DNN model for safely estimating a battery state, and more particularly, to a method and a device for generating a safe-DNN model for safely estimating a battery state (SOX) in a biased manner more conservatively than the actual battery state (SOX), thereby enabling the battery to be replaced or inspected in advance even in the event of sudden battery degradation or thermal runway.

[0002] A secondary battery (hereinafter referred to as a "battery") is a device that converts external electrical energy into chemical energy, stores it, and then generates electricity when needed. In other words, batteries have the advantage of being able to be used for extended periods of time, through repeated charging and discharging.

[0003] These batteries are used in a wide range of applications, from home appliances to electric vehicles. Therefore, accurately estimating the battery's condition is crucial for efficiently managing battery health and improving stability in diverse applications.

[0004] Here, the battery status may include SOC (state of charge), SOH (state of health), SOP (state of power), etc.

[0005] Meanwhile, due to the recent rapid development of artificial intelligence technology, research is being conducted on devices that use artificial intelligence to estimate the condition of batteries.

[0006] When estimating battery status using artificial intelligence, the estimated battery status must be smaller than the actual battery status.

[0007] For example, in the case of SOC, if the SOC estimated through AI is higher than the actual SOC, it means the estimated SOC is higher than the actual battery charge level. This can lead to applications ceasing to operate due to the battery being judged to have a high charge level even when it actually needs to be charged.

[0008] In other words, when estimating the battery status through artificial intelligence, it is possible to replace or inspect the battery in advance before damage occurs due to battery deterioration or thermal runaway only if the estimated status is the same as or lower than the actual battery status.

[0009] Accordingly, the present invention proposes a method to safely estimate the state of the battery in advance by generating a safe-DNN model to estimate SOX (state of battery) with a conservative bias compared to the actual state of the battery using a pre-learning DNN model, thereby enabling the battery to be replaced or inspected in advance before damage occurs due to battery deterioration or thermal runaway.

[0010] That is, the present invention proposes a method for enabling a safe-DNN model to safely estimate SOX by setting a target value for SOX that is biased from the SOX estimation results using multiple pre-learning DNN models in the process of generating a safe-DNN model and proceeding with learning according to the set target value for SOX.

[0011] 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.

[0012] First, Korean Patent Publication No. 2017-0060516 (June 1, 2017) relates to a battery management device and method, which acquires location information of an electric vehicle equipped with a battery, acquires one or more environmental factors among temperature, humidity, climate, terrain, road conditions, and city characteristics of the region according to the location information, and inputs them into an estimation model to estimate the internal state of health (SOH) of the battery.

[0013] In other words, Korean Patent Publication No. 2017-0060516 estimates the internal state of a battery using an estimation model, but does not describe at all how the learning process was performed to create the estimation model. In other words, the estimation model in Korean Patent Publication No. 2017-0060516 appears to have been created using a typical learning method, which could potentially lead to an overestimation of the battery's internal state.

[0014] On the other hand, the present invention generates a safe-DNN that conservatively biases the actual SOX and safely estimates SOX by updating the weights when the learning result of the safe-DNN model is equal to or less than the target value set using the estimation results of multiple pre-learning DNN models in the process of generating a safe-DNN model, and not performing an update on the weights when the learning result of the safe-DNN model exceeds the target value. Korean Patent Publication No. 2017-0060516 does not describe, suggest, or imply any technical features of the present invention.

[0015] In addition, Korean Patent Publication No. 2021-0024962 (March 8, 2021) relates to a device and method for diagnosing the status and predicting the life of an ESS battery, which collects a battery status signal from a battery included in an ESS (energy storage system), generates battery status data from the collected battery status signal using signal analysis and statistical analysis, and then infers result data for diagnosing the status and predicting the life of the battery from the battery status data using a pre-trained deep learning-based diagnosis prediction model.

[0016] Korean Patent Publication No. 2021-0024962 discloses a method for diagnosing battery status and predicting lifespan using a pre-generated diagnostic prediction model. However, the diagnostic prediction model only describes the extent to which it is generated through training based on battery status data and result data. In other words, Korean Patent Publication No. 2021-0024962 has a problem in that the results of battery status diagnosis and lifespan prediction may exceed actual values.

[0017] However, since the present invention relates to generating a safe-DNN that safely estimates SOX by being biased more conservatively than the actual SOX, there are significant differences between the present invention and Korean Patent Publication No. 2021-0024962 in terms of their technical configuration, purpose, and effect.

[0018] The present invention was created to solve the above problems, and the purpose of the present invention is to provide a method for generating a safe-DNN model for safely estimating the state of a battery and a device therefor, by generating a safe-DNN model to estimate SOX with a conservative bias compared to the SOX (state of battery) of an actual battery, thereby safely estimating the state of a battery, thereby enabling the battery to be replaced or inspected in advance before damage due to battery deterioration or thermal runaway occurs.

[0019] In addition, the present invention provides a method and a device for generating a safe-DNN by calculating an error for a SOX estimation result of a battery using a plurality of pre-learning DNN models, determining an upper limit and a lower limit from a distribution of the calculated errors, setting a target value for a biased SOX based on the upper limit, and not updating the weight when the SOX learning result exceeds the target value in the process of generating a safe-DNN model, and updating the weight when it is equal to or less than the target value.

[0020] In addition, the present invention aims to provide a method and device for generating a safe-DNN model including a lower model composed of a plurality of pre-trained DNN models and an upper model that has learned the estimation results of the plurality of pre-trained DNN models.

[0021] In addition, the present invention aims to provide a method and device for safely estimating and providing SOX of a battery when a characteristic value of an actual battery is input.

[0022] A method for generating a safe-DNN model according to one embodiment of the present invention includes a step of generating a plurality of pre-trained DNN models using feature values ​​of a battery as first learning data, and a safe-DNN model generating step of training a deep neural network (DNN) using the first learning data to generate a safe-DNN model, wherein the safe-DNN model generating step includes a target value setting step of setting a target value for SOX biased according to the input first learning data from estimation results of the plurality of pre-trained DNN models generated when the first learning data is input to the deep neural network, and a weight update step of updating a weight of the deep neural network according to the learning result of the process of generating the safe-DNN model and the set target value, wherein the weight update step is characterized in that the safe-DNN model that safely estimates the SOX is generated by updating the weight only when the learning result of the process of generating the safe-DNN model is equal to or less than the set target value.

[0023] In addition, the SOX is characterized by including a state of charge (SOC), a state of energy (SOE), a state of health (SOH), a state of power (SOP), or a combination thereof.

[0024] In addition, the safe-DNN model generation method further includes an error calculation step of calculating an error for an estimation result obtained by estimating the SOX using the plurality of pre-learning DNN models generated above, and an error limit determination step of determining an error limit by determining an upper limit and a lower limit of the error from a distribution of the calculated errors, and the target value setting step is characterized in that the target value is set by adding the determined upper limit to the actual SOX of the battery for each of the input first learning data, or by subtracting the determined upper limit from the actual SOX of the battery.

[0025] In addition, the target value setting step is characterized by setting the target value as the actual SOX of the battery for each of the input first learning data.

[0026] In addition, a method for generating a safe-DNN model according to another embodiment of the present invention includes a pre-training DNN model generation step of generating a plurality of pre-training DNN models using feature values ​​of a battery as first learning data and configuring them as lower-level models, and an upper-level model generation step of training a deep neural network using SOX estimation results using the plurality of pre-training DNN models generated above as second learning data to generate an upper-level model, wherein the upper-level model generation step includes a target value setting step of setting a target value for SOX biased according to the input second learning data from the estimation results of the plurality of pre-training DNN models generated above when second learning data is input to the deep neural network, and a weight update step of updating weights of the deep neural network according to the learning result of the process of generating the upper-level model and the target value, and the weight update step is characterized in that the safe-DNN model for safely estimating the SOX including the lower-level model and the upper-level model is generated by updating the weight only when the learning result of the process of generating the upper-level model is equal to or less than the set target value.

[0027] In addition, the safe-DNN model creation method further includes an error calculation step of calculating an error for an estimation result obtained by estimating the SOX using a plurality of pre-learning DNN models configured as the sub-models, and an error limit determination step of determining an error limit by determining an upper limit and a lower limit of the error from a distribution of the calculated errors, and the target value setting step is characterized in that the target value is set by adding the determined upper limit to the actual SOX of the battery for each of the input second learning data, or by subtracting the determined upper limit from the actual SOX of the battery.

[0028] In addition, the target value setting step is characterized by setting the target value as the actual SOX of the battery for each of the second learning data input.

[0029] In addition, according to another embodiment of the present invention, a method for generating a safe-DNN model according to an embodiment of the present invention includes a pre-learning DNN model generation step of generating a plurality of pre-learning DNN models using feature values ​​of a battery as first learning data, and a SOX safety module generation step of generating a SOX safety module that finally estimates the SOX by selecting a minimum value among SOX estimation results for the battery of the plurality of pre-learning DNN models, and is characterized in that the number of the pre-learning DNN models is increased so that one of the estimation results for SOX of the plurality of pre-learning DNN models is always equal to or less than the actual SOX of the battery, thereby configuring the plurality of pre-learning DNN models.

[0030] In addition, a safe-DNN model generation device according to one embodiment of the present invention is characterized by generating a safe-DNN model according to the safe-DNN model generation method.

[0031] As described above, the present invention has the effect of generating a safe-DNN model to safely estimate the SOX by conservatively biasing the actual SOX of the battery, so that when the feature value of the actual battery is input, the SOX for the battery can be safely estimated and provided.

[0032] In addition, the present invention has the effect of preventing damage caused by sudden battery deterioration or thermal runaway in advance by providing a safely estimated SOX to enable battery replacement or inspection.

[0033] FIG. 1 is a diagram illustrating a method for generating a safe-DNN model for safely estimating a battery state according to one embodiment of the present invention.

[0034] FIG. 2 is a diagram illustrating a method for setting a target value in a learning process according to one embodiment of the present invention.

[0035] FIG. 3 is a drawing illustrating a method for setting a target value according to another embodiment of the present invention.

[0036] FIG. 4 is a diagram illustrating the learning results of a safe-DNN model generated by setting a target value according to another embodiment of the present invention.

[0037] FIG. 5 is a diagram illustrating a safe-DNN model according to another embodiment of the present invention.

[0038] FIG. 6 is a diagram illustrating a safe-DNN model according to another embodiment of the present invention.

[0039] Fig. 7 is a block diagram showing the configuration of a safe-DNN model generation device according to one embodiment of the present invention.

[0040] Fig. 8 is a block diagram showing the configuration of a safe-DNN model generation device according to another embodiment of the present invention.

[0041] FIG. 9 is a block diagram showing the configuration of a safe-DNN model generation device according to another embodiment of the present invention.

[0042] FIG. 10 is a flowchart illustrating a procedure for generating a safe-DNN model according to one embodiment of the present invention.

[0043] FIG. 11 is a flowchart illustrating a procedure for generating a safe-DNN model according to another embodiment of the present invention.

[0044] FIG. 12 is a flowchart illustrating a procedure for generating a safe-DNN model according to another embodiment of the present invention.

[0045] FIG. 13 is a flowchart illustrating a procedure for estimating and providing SOX of a battery mounted on a user application according to one embodiment of the present invention.

[0046] [Description of symbols] 100, 100a, 100b: safe-DNN model generation device; 110, 110a, 110b: battery feature value receiving unit; 120, 120a, 120b: SOX estimation unit; 130, 130a, 130b: SOX estimation result providing unit; 140, 140a, 140b: safe alarm output unit; 180, 150a, 150b: safe-DNN model generation unit; 150, 151a, 151b: pre-training DNN model generation unit; 160, 152a: error calculation unit; 152b: SOX safety module generation unit; 170, 153a: error limit determination unit; 181, 154a: learning data input unit; 182, 155a: Target value setting unit; 183, 156a: Learning outcome evaluation unit; 184, 157a: Weight update unit.

[0047] Hereinafter, with reference to the attached drawings, a preferred embodiment of a safe-DNN model generation method and a device thereof for safely estimating a battery state of the present invention will be described in detail. The same reference numerals in each drawing represent the same components. In addition, specific structural and functional descriptions of embodiments of the present invention are merely illustrative for the purpose of explaining embodiments according to the present invention, and unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by a person of ordinary skill in the art to which the present invention pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning they have in the context of the related technology, and it is preferable not to interpret them in an ideal or excessively formal sense unless explicitly defined in this specification.

[0048] FIG. 1 is a diagram illustrating a method for generating a safe-DNN model for safely estimating a battery state according to one embodiment of the present invention.

[0049] As illustrated in FIG. 1, a device (100) for generating a safe-DNN model for safely estimating a battery state according to an embodiment of the present invention (hereinafter, a safe-DNN model generating device) learns the characteristic values ​​of a battery as first learning data to generate a safe-DNN model. Here, the characteristic values ​​of the battery include open circuit voltage (OCV), voltage (V), current (I), charge (Q), temperature (T), or a combination thereof.

[0050] The above battery state refers to SOX (state of X), and SOX is an indicator that can estimate the battery state and includes SOC (state of charge), SOH (state of health), SOE (state of energy), SOP (state of power), or a combination thereof.

[0051] The above SOC or SOE is information that tells you how long the battery can be used, SOH is information that tells you how good the battery's performance is, and SOP is information that tells you how powerful the battery is.

[0052] The above learning is performed through a deep neural network (DNN), and the deep neural network is configured to include an input layer including at least one input node, at least one hidden layer including at least one hidden node, and an output layer including an output node.

[0053] The above input layer refers to a layer into which feature values ​​constituting each learning data labeled with the SOX (label) of the battery are input.

[0054] The feature values ​​of each battery are arranged so that they are input to the appropriate input node. For example, feature value 1 is assigned to the first input node, feature value 2 to the second input node, and so on. In other words, each input node is configured to input the corresponding feature value.

[0055] The above input layer, hidden layer, and output layer are each connected through links having specific weights, and at this time, the safe-DNN model generation device (100) already knows the result (actual SOX) according to the characteristic value of the battery input to the input node during the learning process, so that the error between the SOX learning result (i.e., the learning result output from the output layer) according to the characteristic value of the battery during the learning process and the actual SOX can be reduced by updating (adjusting) the weight of the link.

[0056] The above weight refers to a learning parameter, and the learning is performed by updating the weight (i.e., updating the learning parameter).

[0057] Updating the above weights is performed through the backpropagation method, which reduces errors by backpropagating them through the deep neural network.

[0058] Traditional deep neural network training focuses solely on reducing output layer errors. Applying this directly to estimating a battery's SOX can lead to estimates of SOX that are significantly higher than the actual battery's SOX. This can lead to the assumption that the battery is not significantly deteriorated, even when it is significantly deteriorated.

[0059] For example, while driving an electric vehicle, if the actual SOC is 25% but is estimated to be 50%, the user may perceive that there is still a long time left until the battery is discharged, and continue driving, but the battery may suddenly be discharged, causing the electric vehicle to suddenly stop while driving.

[0060] Ultimately, in order to estimate the SOX of a battery, it is necessary to conservatively bias the actual SOX of the battery, including the risk that can be considered, to safely estimate the SOX.

[0061] Accordingly, the safe-DNN model generation device (100) of the present invention sets a target value for biased SOX from the estimation result of SOX through a pre-learning DNN model, and performs learning by updating weights when the learning result for the first learning data input in the process of generating a safe-DNN model (i.e., learning process) is equal to or less than the set target value, and generates a safe-DNN model by not updating the weights when it exceeds the target value.

[0062] The above target value is intended to enable learning to be performed with a conservative bias toward the actual SOX of the battery within a range that can tolerate risk.

[0063] Setting the above target value will be described in detail with reference to FIGS. 2 to 4.

[0064] At this time, when machine learning is completed, the corresponding deep neural network becomes a safe-DNN model. Here, the deep neural network illustrated in Figure 1 illustrates an artificial intelligence learning network based on ANN, but the deep neural network in the present invention is not limited to this and may be composed of various artificial intelligence learning networks such as a deep convolutional neural network (DCNN), a transformer, and a temporal convolutional neural network (TCNN).

[0065] The above safe-DNN model is configured to output SOX and the probability of each SOX when the characteristic value of an actual battery is input, and the safe-DNN model generation device (100) estimates the SOX of the corresponding battery by selecting the SOX with the highest probability.

[0066] The above safe-DNN model generation device (100) can be configured in the form of a server on the cloud, and can be configured to receive the characteristic values ​​of a battery measured from a battery-equipped application (e.g., an electric vehicle), estimate the SOX of the battery, and provide the same to the application.

[0067] However, the safe-DNN model generation device (100) of the present invention generates only a safe-DNN model, and it is also possible to provide a SOX estimation device for estimating SOX including the generated safe-DNN model in the form of an application program so as to directly estimate SOX locally.

[0068] FIG. 2 is a diagram illustrating a method for setting a target value in a learning process according to one embodiment of the present invention.

[0069] As illustrated in FIG. 2, a safe-DNN model generation device (100) according to one embodiment of the present invention first generates a plurality of pre-learning DNN models in order to set a target value in the learning process.

[0070] The above pre-trained DNN model is created by learning the first training data using a general machine learning method.

[0071] In addition, the safe-DNN model generation device (100) calculates an error for the actual SOX of the battery from the estimation results of multiple pre-learning DNN models, and determines the upper limit (MAX) and lower limit (MIN) of the error from the distribution of the calculated error.

[0072] Here, the upper bound refers to the maximum error in the positive (+) direction, and the lower bound refers to the maximum error in the negative (-) direction. The upper and lower bounds can be set within an acceptable range.

[0073] The estimation results of the above multiple pre-training DNN models can be obtained by inputting each feature value of the battery into multiple pre-training DNN models.

[0074] In addition, the safe-DNN model generation device (100) sets a target value for biased SOX from the estimation results of the plurality of pre-learning DNN models during the learning process of generating a safe-DNN model by inputting learning data.

[0075] At this time, the safe-DNN model generation device (100) sets a target value by adding an error (MAX) corresponding to an upper limit to the actual SOX of the battery according to the first learning data (more specifically, the characteristic value of the battery input as learning data) when the first learning data is input during the learning process of generating the safe-DNN. In other words, the target value is newly set for each first learning data input during the learning process of generating the safe-DNN.

[0076] For example, if the actual SOX of the battery for the characteristic value of the battery input as the first learning data is 0.7 and MAX is 0.1, the target value of the corresponding learning data (the characteristic value of the battery) becomes 0.8.

[0077] At this time, the safe-DNN model generation device (100) updates the weights of the safe-DNN model (more specifically, the deep neural network) only when the learning result of the process of generating the safe-DNN model (i.e., the learning result for the input learning data) is equal to or less than the set target value (i.e., the set target value for the input learning data).

[0078] Through this, the safe-DNN model generation device (100) generates a safe-DNN model that can safely estimate the SOX of the battery by conservatively biasing the actual SOX of the battery within an allowable range.

[0079] Meanwhile, the safe-DNN model generation device (100) can determine the upper and lower limits for the learning results according to the learning data (i.e., the characteristic values ​​of the battery) based on the error distribution according to the estimation results of a plurality of pre-learning DNN models in the process of generating a safe-DNN model. At this time, the actual characteristic values ​​of the battery are input into the safe-DNN model to estimate the SOX of the battery, and the learning result has a value between the upper and lower limits. However, if the learning result is outside the range composed of the upper and lower limits corresponding to the actual characteristic values ​​of the battery, the safe-DNN model generation device (100) can be configured to output a safety alarm notifying that a problem has occurred in the battery.

[0080] FIG. 3 is a diagram illustrating a method for setting a target value according to another embodiment of the present invention, and FIG. 4 is a diagram illustrating a learning result of a safe-DNN model generated by setting a target value according to another embodiment of the present invention.

[0081] As illustrated in FIG. 3, a method for setting a target value according to another embodiment of the present invention is as described with reference to FIG. 2, in which a safe-DNN model generation device (100) determines an upper limit and a lower limit of an error from the estimation results of a plurality of pre-learning DNN models.

[0082] In addition, the safe-DNN model generation device (100) sets a target value for biased SOX from the estimation results of the plurality of pre-learning DNN models in the process of generating a safe-DNN model by inputting first learning data (learning process).

[0083] At this time, the safe-DNN model generation device (100) sets a target value by subtracting (i.e., actual SOX ?? MAX) an error (MAX) corresponding to an upper limit of the actual SOX of the battery according to the first learning data (more specifically, the characteristic value of the battery input as the first learning data) when the first learning data is input during the learning process of generating the safe-DNN. That is, the target value is newly set for each first learning data input during the learning process of generating the safe-DNN.

[0084] For example, if the actual SOX of the battery for the characteristic value of the battery input as the first learning data is 0.7 and MAX is 0.1, the target value of the corresponding learning data (the characteristic value of the battery) becomes 0.6.

[0085] At this time, the safe-DNN model generation device (100) updates the weights of the safe-DNN model (more specifically, the deep neural network) only when the learning result of the process of generating the safe-DNN model (i.e., the learning result for the input learning data) is equal to or less than the set target value (i.e., the set target value for the input learning data).

[0086] Through this, the safe-DNN model generation device (100) generates a safe-DNN model that can safely estimate the SOX of the battery by being biased conservatively compared to the actual SOX of the battery.

[0087] Meanwhile, as described with reference to FIGS. 1 and 2, the target value can be set to the actual SOX + MAX of the battery or the actual SOX ?? MAX of the battery, but can also be set to the actual SOX of the battery (in this case, the actual SOX of the battery labeled in the first learning data is used as is).

[0088] Also, as shown in Fig. 4, when the target value is set to the actual SOX + MAX of the battery or the actual SOX of the battery, it is common to have a value lower than the estimated result of the pre-learning DNN model.

[0089] At this time, the safe-DNN model generation device (100) preferably estimates the SOX of the battery using the generated safe-DNN model, but may be configured to estimate the SOX of the battery using one of a plurality of pre-trained DNN models. In this case, if the SOX estimated using the safe-DNN model exceeds the SOX estimated using the pre-trained DNN model, a safety alarm may be output. In addition, the safe-DNN model generation device (100) may provide the learning result of the safe-DNN model or the estimation result of the pre-trained DNN model to the user, or may provide the average result to the user, depending on the status.

[0090] FIG. 5 is a diagram illustrating a safe-DNN model according to another embodiment of the present invention.

[0091] As illustrated in FIG. 5, a safe-DNN according to one embodiment of the present invention is configured to include a lower model composed of a plurality of pre-learning DNN models and an upper model that ultimately estimates the SOX of a battery based on the output results of the lower models.

[0092] To this end, the safe-DNN model generation device (100) generates multiple pre-learning DNN models using the feature values ​​of the battery as first learning data, and configures the multiple pre-learning DNN models as sub-models.

[0093] In addition, the safe-DNN model generation device (100) generates an upper model using the estimation results (SOX) of multiple pre-learning DNN models composed of lower models as second learning data.

[0094] The above second learning data is generated by labeling the actual SOX of the battery in a SOX set composed of SOX estimated according to the characteristic values ​​of the battery in multiple DNN models configured as sub-models.

[0095] In addition, the safe-DNN model generation device (100) sets a target value for biased SOX from the estimation results of multiple pre-learning DNN models composed of the lower models during the learning process of generating an upper model by inputting second learning data.

[0096] At this time, the safe-DNN model generation device (100) sets a target value according to the second learning data (more specifically, the SOX of the lower model input as learning data) when the second learning data is input during the learning process of generating the upper model. In other words, the target value is newly set for each second learning data input during the learning process of generating the upper model.

[0097] Setting the above target value is performed in the same manner as the method described with reference to FIGS. 2 to 4, except that the data (feature values ​​and SOX) that constitute the learning data are different, so a detailed description will be omitted.

[0098] FIG. 6 is a diagram illustrating a safe-DNN model according to another embodiment of the present invention.

[0099] As illustrated in FIG. 6, a safe-DNN model according to another embodiment of the present invention is configured to include a plurality of pre-learning DNN models and an estimation module.

[0100] In order to create the above safe-DNN model, the safe-DNN model creation device (100) creates multiple pre-learning DNN models using the feature values ​​of the battery as first learning data.

[0101] The above multiple pre-trained DNN models output SOX (estimated result) when the characteristic values ​​of an actual battery are input, and the above SOX safety module finally estimates SOX for an actual battery by selecting the minimum value among the estimated results of the multiple pre-trained DNN models.

[0102] At this time, multiple pre-training DNN models are constructed by increasing the number of pre-training DNN models until one of the outputs (estimated results) of the pre-training DNN models always has a value less than the actual SOX.

[0103] Fig. 7 is a block diagram showing the configuration of a safe-DNN model generation device according to one embodiment of the present invention.

[0104] As illustrated in FIG. 7, a safe-DNN model generation device (100) according to one embodiment of the present invention is configured to include a battery feature value receiving unit (110), a SOX estimation unit (120), a SOX estimation result providing unit (130), a safe alarm output unit (140), a pre-learning DNN model generation unit (150), an error calculation unit (160), an error limit determination unit (170), and a safe-DNN model generation unit (180).

[0105] The above-described pre-training DNN model generation unit (150) generates multiple pre-training DNN models using the battery's feature values ​​as the first learning data. As described above, the first learning data is generated by labeling the battery's feature values ​​with SOX.

[0106] The above error calculation unit (160) calculates the error for the estimation results of the multiple pre-trained DNN models generated, and the error limit determination unit (170) determines the upper and lower limits based on the distribution of the calculated errors. The calculation of the error and the determination of the upper and lower limits have been described with reference to FIG. 2, and therefore are omitted here.

[0107] The above safe-DNN model generation unit (180) is configured to generate a safe-DNN model by learning the first learning data, and includes a learning data input unit (181), a target value setting unit (182), a learning result evaluation unit (183), and a weight update unit (184).

[0108] The above learning data input unit (181) inputs the first learning data into a deep neural network (i.e., safe-DNN model).

[0109] The target value setting unit (182) sets a target value according to the input first learning data. The target value is set by adding or subtracting an error corresponding to the upper limit determined by the error limit determination unit (170) to the actual SOX. In this case, the target value can be set to the actual SOX.

[0110] Setting the above target value is described with reference to FIGS. 2 and 3, so it is omitted here.

[0111] The above learning result evaluation unit (183) evaluates whether the learning result for the input first learning data exceeds the target value set for the first learning data during the learning process for creating a safe-DNN model.

[0112] The weight update unit (184) updates the weights of the deep neural network when the learning result evaluated by the learning result evaluation unit (183) is equal to or less than the target value. The weight update unit (184) updates the weights to reduce the error between the learning result for the first learning data input through the backpropagation method and the target value.

[0113] The above battery characteristic value receiving unit (110) receives battery characteristic values ​​from an application (e.g., an electric vehicle) provided by the user.

[0114] The above SOX estimation unit (120) arranges the received battery feature values ​​to fit the input of the safe-DNN model and inputs the battery feature values ​​as an input vector to the safe-DNN model to estimate SOX.

[0115] The SOX estimation result provision unit (130) provides the estimated SOX to an application provided by the user. Here, the estimation can be performed locally on the application device or on a cloud server and provided to the local device.

[0116] The above safe alarm output unit (140) outputs a safe alarm according to the estimated SOX, and outputting the safe alarm is performed when it is determined that an abnormal condition has occurred in the battery. Since this has been described with reference to FIGS. 2 to 4, further detailed description will be omitted.

[0117] Here, a SOX estimation device that extracts feature values ​​from the measurement values ​​received from the battery and makes them into an input vector of a safe-DNN model, and estimates safe SOX through this, including a SOX estimation unit (120), a SOX estimation result provision unit (130), and a safe alarm output unit (140), is more preferably configured separately and provided in a device or cloud server on the application side. In the present invention, for convenience, the battery SOX estimation device is only described as being integrated into the safe-DNN model generation device (100).

[0118] Fig. 8 is a block diagram showing the configuration of a safe-DNN model generation device according to another embodiment of the present invention.

[0119] As illustrated in FIG. 8, a safe-DNN model generation device (100a) according to another embodiment of the present invention is configured to include a battery feature value receiving unit (110a), a SOX estimation unit (120a), a SOX estimation result providing unit (130a), a safe alarm output unit (140a), and a safe-DNN model generation unit (150a).

[0120] The above battery characteristic value receiving unit (110a), SOX estimation unit (120a), SOX estimation result providing unit (130a), and safe alarm output unit (140a) have the same configuration as the battery characteristic value receiving unit (110), SOX estimation unit (120), SOX estimation result providing unit (130), and safe alarm output unit (140) described with reference to FIG. 7, so a detailed description thereof will be omitted.

[0121] The above safe-DNN model generation unit (150a) is for generating a safe-DNN model that includes a lower model composed of multiple pre-learning DNN models, unlike the safe-DNN model described in FIG. 7, and an upper model generated using the estimation results of the lower models, and is configured to include a pre-learning DNN model generation unit (151a), an error calculation unit (152a), an error limit determination unit (153a), and an upper model generation unit.

[0122] The above pre-learning DNN model generation unit (151a) learns the first learning data and generates multiple pre-learning DNN models to form a lower model.

[0123] The above pre-learning DNN model generation unit (151a), error calculation unit (152a), and error limit determination unit (153a) are identical components that perform the same functions as the pre-learning DNN model generation unit (150), error calculation unit (160), and error limit determination unit (170) described with reference to FIG. 7, so a detailed description thereof will be omitted.

[0124] The above upper model generation unit is configured to generate an upper model for learning second learning data according to the estimation result of the lower model and ultimately outputting the SOX of the battery as a learning result, and includes a learning data input unit (154a), a target value setting unit (155a), a learning result evaluation unit (156a), and a weight update unit (157a).

[0125] The above learning data input unit (154a) inputs the second learning data into the upper model (i.e., deep neural network).

[0126] The target value setting unit (155a) sets a target value according to the input second learning data. The target value is set by adding or subtracting an error corresponding to the upper limit determined by the error limit determination unit (153a) to the actual SOX. At this time, the target value can be set to the actual SOX. The operation of the target value setting unit (155a) is identical to the operation of the target value setting unit (182) of FIG. 7.

[0127] The above learning result evaluation unit (156a) evaluates whether the learning result for the input second learning data exceeds the target value set for the second learning data during the learning process for creating a safe-DNN model.

[0128] The weight update unit (157a) updates the weights of the safe-DNN model when the learning result evaluated by the learning result evaluation unit (156a) is equal to or less than the target value. The weight update unit (157a) updates the weights to reduce the error between the learning result for the input learning data and the target value through the backpropagation method.

[0129] Here, a SOX estimation device that extracts feature values ​​from the measurement values ​​received from the battery and makes them into an input vector of a safe-DNN model, and estimates safe SOX through this, including a SOX estimation unit (120a), a SOX estimation result provision unit (130a), and a safe alarm output unit (140a), is more preferably configured separately and provided in a device or cloud server on the application side. In the present invention, for convenience, the battery SOX estimation device is only described as being integrated into the safe-DNN model generation device (100a).

[0130] FIG. 9 is a block diagram showing the configuration of a safe-DNN model generation device according to another embodiment of the present invention.

[0131] As illustrated in FIG. 9, a safe-DNN model generation device (100b) according to another embodiment of the present invention is configured to include a battery feature value receiving unit (110b), a SOX estimation unit (120b), a SOX estimation result providing unit (130b), a safe alarm output unit (140b), and a safe-DNN model generation unit (150b).

[0132] The above battery characteristic value receiving unit (110b), SOX estimation unit (120b), SOX estimation result providing unit (130b), and safe alarm output unit (140b) have the same configuration as the battery characteristic value receiving unit (110), SOX estimation unit (120), SOX estimation result providing unit (130), and safe alarm output unit (140) described with reference to FIG. 7, so a detailed description thereof will be omitted.

[0133] The above safe-DNN model generation unit (150b) is configured to include a pre-learning DNN model generation unit (151b) and a SOX safety module generation unit (152b).

[0134] The above pre-learning DNN model generation unit (151b) has the same configuration as the pre-learning DNN model generation unit (150) of FIG. 7, and is generated by learning the first learning data.

[0135] The above SOX safety module generation unit (152b) is configured to compile the estimation results of multiple pre-learning DNN models and output the result of estimating the safe SOX of an actual battery.

[0136] The above SOX safety module generation unit (152b) is configured to input estimation results of multiple pre-learning DNN models, select the minimum value from the input estimation results, and output the SOX learning result, thereby outputting a safe SOX learning result.

[0137] Figures 7 to 9 illustrate block diagrams of a safe-DNN model generation device according to the configuration of a safe-DNN model according to one embodiment of the present invention, but can be integrated into a single safe-DNN model generation device. In this case, duplicate components performing the same function are configured as a single component, and can be configured to generate a safe-DNN model of a specific configuration as needed.

[0138] Meanwhile, each safe-DNN model generation device according to an embodiment of the present invention illustrated in FIGS. 7 to 9 may be built in an application-side device using a battery or built in a server on the cloud to generate each safe-DNN model and store it in a database or memory. The safe-DNN model generation device is also utilized as a SOX estimation device that estimates the SOX of a battery by retrieving each stored safe-DNN model from the memory or database. The SOX estimation device may be configured in a separate device from the safe-DNN model generation device.

[0139] The above SOX estimation device configures the measured values ​​of the battery received from each battery characteristic value receiving unit (110, 110a, 110b) of FIGS. 7 to 9 as characteristic values, inputs them as input vectors to each safe-DNN model, estimates SOX through each SOX estimation unit (120, 120a, 120b), provides SOX estimation results through each SOX estimation result providing unit (130, 130a, 130b), and outputs them through each safe alarm output unit (140, 140a, 140b) when it is determined that a safety problem has occurred.

[0140] FIG. 10 is a flowchart illustrating a procedure for generating a safe-DNN model according to one embodiment of the present invention.

[0141] FIG. 10 illustrates a procedure for generating a safe-DNN model through the safe-DNN model generating device (100) described in FIG. 7. As illustrated in FIG. 10, the safe-DNN model generating device (100) according to an embodiment of the present invention performs a pre-training DNN model generating step of generating a plurality of pre-training DNN models using battery feature values ​​as first learning data (S110), and performs an error calculating step (S120) of calculating an error using the estimation results of the plurality of generated pre-training DNN models, and an error limit determining step (S130) of determining an upper limit and a lower limit according to the distribution of the calculated error to determine an error limit.

[0142] The process of generating the above multiple DNN pre-training models, calculating errors, and determining error limits has been described with reference to Fig. 2, and thus is omitted here.

[0143] Next, the safe-DNN model generation device (100) performs a safe-DNN model generation step of generating a safe-DNN model.

[0144] The above safe-DNN model creation step first performs a learning data input step of inputting the first learning data into the deep neural network (S140).

[0145] Next, the safe-DNN model generation step performs a target value setting step (S150), which sets a target value for the biased SOX of the input first training data using the determined upper bound of the error. The target value is set for each first training data, and the process of setting the target value is omitted here as it has been described with reference to Figure 2.

[0146] Next, the safe-DNN model creation step performs an evaluation step to evaluate whether the learning result of the deep neural network for the first learning data input in the process of creating the safe-DNN model (i.e., the learning process) is equal to or less than a target value set according to the first learning data, and if the learning result as a result of the evaluation is equal to or less than the target value (S160), a weight update step to update the weights of the deep neural network to minimize the error between the learning result and the target value is performed (S170).

[0147] Updating the above weights is performed via backpropagation.

[0148] Next, the safe-DNN model generation device (100) repeatedly performs steps S140 to S180 until all first learning data is input and learning is completed (S180).

[0149] At this time, the deep neural network that has completed learning becomes the safe-DNN model.

[0150] FIG. 11 is a flowchart illustrating a procedure for generating a safe-DNN model according to another embodiment of the present invention.

[0151] FIG. 11 illustrates a procedure for generating a safe-DNN model through the safe-DNN model generating device (100a) described in FIG. 8. As illustrated in FIG. 11, the safe-DNN model generating device (100) according to an embodiment of the present invention performs a pre-training DNN model generating step of generating a plurality of pre-training DNN models using battery feature values ​​as first learning data and configuring them as lower models (S210), and performs an error calculating step (S220) of calculating an error using the estimation results of the plurality of generated pre-training DNN models, and an error limit determining step (S230) of determining an upper limit and a lower limit according to the distribution of the calculated error to determine an error limit.

[0152] Next, the safe-DNN model generation device (100a) performs an upper model generation step of generating an upper model.

[0153] The above-mentioned upper model generation step first performs a learning data input step of inputting multiple pre-learning DNN estimation results as second learning data into the deep neural network for upper model generation (S240).

[0154] Next, the upper model generation step performs a target value setting step that sets a target value for the biased SOX of the input second learning data using the determined upper limit (S250). The target value setting step sets the target value in the same manner as the target value setting step described in FIG. 10.

[0155] Next, the upper model creation step performs an evaluation step to evaluate whether the learning result of the deep neural network for the second learning data input in the process of creating a safe-DNN model (more specifically, the process of creating an upper model constituting the safe-DNN model) is equal to or less than a target value set according to the second learning data, and if the learning result as a result of the evaluation is equal to or less than the target value (S260), a weight update step to update the weights of the deep neural network to minimize the error between the learning result and the target value is performed (S270).

[0156] Next, the safe-DNN model generation device (100a) repeatedly performs steps S140 to S280 until all second learning data is input and learning is completed (S280).

[0157] At this time, the deep neural network that has completed learning becomes the upper model, and the safe-DNN model is composed of the upper model and lower models.

[0158] FIG. 12 is a flowchart illustrating a procedure for generating a safe-DNN model according to another embodiment of the present invention.

[0159] FIG. 12 illustrates a procedure for generating a safe-DNN model through the safe-DNN model generation device (100b) described in FIG. 9. As illustrated in FIG. 12, the safe-DNN model generation device (100b) according to one embodiment of the present invention performs a pre-training DNN model generation step of generating a plurality of pre-training DNN models using battery feature values ​​as first learning data (S310).

[0160] Here, multiple pre-trained DNN models are constructed by adding the number of pre-trained DNN models until one of the estimated results always has a value equal to or less than the actual SOX of the battery.

[0161] Next, the safe-DNN model generation device (100b) performs a SOX safety module generation step of generating a SOX safety module that compiles the respective estimation results of multiple pre-trained DNN models and ultimately learns and outputs SOX (S320). In other words, the safe-DNN model is configured to include multiple pre-trained DNN models and a SOX safety module.

[0162] FIG. 13 is a flowchart illustrating a procedure for estimating and providing SOX of a battery mounted on a user application according to one embodiment of the present invention.

[0163] As illustrated in FIG. 13, a safe-DNN model generation device (100) according to one embodiment of the present invention (same as the safe-DNN model generation devices (100a, 100b) of FIGS. 8 and 9) performs a battery feature value reception step of receiving battery feature values ​​from a user application (S410).

[0164] The above user application refers to a means equipped with a battery (e.g., an electric vehicle), and the battery characteristics refer to measurement data of the battery, including open circuit voltage, voltage, current, charge, temperature, or a combination thereof.

[0165] Next, the safe-DNN model generation device (100) performs a SOX estimation step (S420) of inputting the received battery feature values ​​into the safe-DNN model to estimate the SOX of the corresponding battery, and performs a SOX provision step of providing the estimated SOX of the battery to a user application (S430).

[0166] Meanwhile, if an abnormal state of the battery is detected as a result of estimating the SOX of the battery, the safe model generation device (100) further includes performing a safe alarm output step of outputting a safe alarm.

[0167] In the present invention, the processes of S410 to S430 are described as being performed in a safe-DNN model generation device (100, 100a, 100b), but it may be more preferable to perform them in a separate SOX estimation device.

[0168] As described above, the present invention creates a safe-DNN model that stably estimates the state of the battery by being biased conservatively compared to the actual SOX of the battery, thereby enabling the battery to be replaced or inspected in advance before damage occurs due to sudden battery deterioration or thermal runaway.

[0169] In addition, although the preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above, and various modifications can be implemented by a person having ordinary skill in the art to which the invention pertains without departing from the gist of the present invention claimed in the claims, and such modifications should not be understood individually from the technical idea or prospect of the present invention.

[0170] As described above, the present invention generates a safe-DNN model to safely estimate the SOX by conservatively biasing the actual SOX of a battery, so that when the characteristic values ​​of an actual battery are input, the SOX for the battery can be safely estimated and provided, and by providing the safely estimated SOX, the battery can be replaced or inspected, thereby preventing damage caused by sudden battery deterioration or thermal runaway in advance, and thus has industrial applicability.

Claims

1. A step of creating multiple pre-trained DNN models using the characteristic values of the battery as the first learning data; and A safe-DNN model generation step for generating a safe-DNN model by training a deep neural network using the first learning data; The safe-DNN model generation step includes a target value setting step for setting a target value for SOX biased according to the input first learning data from the estimation results of the plurality of pre-training DNN models generated when first learning data is input to the deep neural network; and It includes a weight update step for updating the weights of the deep neural network according to the learning results of the process of generating the safe-DNN model and the set target value; A method for generating a safe-DNN model, characterized in that the weight update step generates the safe-DNN model that safely estimates the SOX by updating the weight only when the learning result of the process of generating the safe-DNN model is equal to or less than the set target value.

2. In claim 1, The above SOX is, A method for generating a safe-DNN model, characterized in that it includes a state of charge (SOC), a state of energy (SOE), a state of health (SOH), a state of power (SOP), or a combination thereof.

3. In claim 1, The above safe-DNN model creation method is: An error calculation step for calculating an error for the estimation result of estimating the SOX using the plurality of pre-learning DNN models generated above; and It further includes an error limit determination step of determining an error limit by determining an upper and lower limit of the error from the distribution of the errors calculated above; A method for creating a safe-DNN model, characterized in that the target value setting step sets the target value by adding the determined upper limit to the actual SOX of the battery for each of the input first learning data, or by subtracting the determined upper limit from the actual SOX of the battery.

4. In claim 1, The above target value setting step is: A method for creating a safe-DNN model, characterized in that the target value is set to the actual SOX of the battery for each of the first learning data input above.

5. A pre-training DNN model creation step in which multiple pre-training DNN models are created using the battery's feature values as the first learning data and configured as sub-models; and It includes an upper model generation step of generating an upper model by training a deep neural network using the SOX estimation results using the multiple pre-learning DNN models generated above as second learning data; The above upper model generation step is a target value setting step for setting a target value for SOX that is biased according to the input second learning data from the estimation results of the plurality of pre-learning DNN models generated when second learning data is input to the deep neural network; and It includes a weight update step for updating the weights of the deep neural network according to the learning results of the process of generating the upper model and the target value; A safe-DNN model generation method characterized in that the weight update step generates a safe-DNN model that safely estimates the SOX, including the lower model and the upper model, by updating the weight only when the learning result of the process of generating the upper model is equal to or less than the set target value.

6. In claim 5, The above SOX is, A method for generating a safe-DNN model, characterized in that it includes a state of charge (SOC), a state of energy (SOE), a state of health (SOH), a state of power (SOP), or a combination thereof.

7. In claim 5, The above safe-DNN model creation method is: An error calculation step for calculating an error for the estimation result of estimating the SOX using multiple pre-learning DNN models composed of the above sub-models; and It further includes an error limit determination step of determining an error limit by determining an upper and lower limit of the error from the distribution of the errors calculated above; A method for creating a safe-DNN model, characterized in that the target value setting step sets the target value by adding the determined upper limit to the actual SOX of the battery for each of the input second learning data, or by subtracting the determined upper limit from the actual SOX of the battery.

8. In claim 5, The above target value setting step is: A method for creating a safe-DNN model, characterized in that the target value is set to the actual SOX of the battery for each of the second learning data input above.

9. A pre-training DNN model generation step for generating multiple pre-training DNN models using the battery's feature values as the first learning data; and A SOX safety module generation step for generating a SOX safety module that finally learns the SOX by selecting the minimum value among the SOX estimation results for the battery of the plurality of pre-learning DNN models generated above; A safe-DNN model generation method characterized in that the number of pre-trained DNN models is increased so that one of the estimation results for SOX of the plurality of pre-trained DNN models is always equal to or less than the actual SOX of the battery.

10. A safe-DNN model generation device characterized in that it generates a safe-DNN model according to the safe-DNN model generation method described in any one of claims 1 to 4.

11. A safe-DNN model generation device characterized by generating a safe-DNN model according to the safe-DNN model generation method described in any one of claims 5 to 8.

12. A safe-DNN model generation device characterized by generating a safe-DNN model according to the safe-DNN model generation method described in claim 9.

Images