Three-level alarm identification method and device for automobile battery

Abstract

The invention provides a three-level alarm identification method and device for an automobile battery. The method comprises the following steps: firstly, receiving real-time signal data uploaded by a target vehicle; then, performing data cleaning on the real-time signal data to obtain an input parameter set; finally, the input parameter set is input into a third-level alarm recognition model, whether the battery of the target vehicle is in a third-level alarm state or not is judged through the third-level alarm recognition model, and the third-level alarm recognition model is a bidirectional long-short-term memory structure sequence model obtained through pre-training. Therefore, the three-level alarm identification model is obtained by pre-training the bidirectional long and short time memory structure sequence model, and the time sequence characteristics of the signal data before and after the alarm moment can be considered, so that the three-level alarm identification model can accurately and effectively identify whether the vehicle battery is in the three-level alarm state. Therefore, the fault of the vehicle battery can be analyzed and processed in time, and the use risk of the new energy battery is reduced.

Description

Technical Field

[0001] The present application relates to the technical field of driving safety of new energy vehicles, and in particular to a three-level alarm identification method and device for automobile batteries. Background Art

[0002] With the improvement of environmental protection awareness and the increasingly serious energy crisis, new energy vehicles as a low-carbon, environmentally friendly means of transportation are gradually gaining attention and praise worldwide. New energy vehicles use unconventional automotive fuels as a power source, and integrate advanced technologies in vehicle power control and driving, with the characteristics of low pollution, low energy consumption and green environmental protection. At the same time, new energy vehicles usually use "electricity" as power, and the safety issues caused by electric vehicle failures cannot be ignored.

[0003] Battery thermal management is of great significance in pure electric vehicles, hybrid electric vehicles and other power systems that use dynamic batteries as power sources. Battery thermal management mainly includes: air cooling, liquid cooling, thermoelectric cooling, heat pipe cooling and phase change material thermal management. According to the published standards, the fault alarm levels of electric vehicles can be divided into level one, level two and level three, among which the level three alarm is the most serious fault. The three-level alarm status can include battery high temperature alarm, single cell overvoltage alarm, on-board energy storage device type overvoltage alarm, on-board energy storage device type overcharge alarm and insulation alarm. However, the current existing technology cannot accurately identify whether the on-board battery of a new energy vehicle is in the three-level alarm state, resulting in safety risks in the use of on-board batteries of new energy vehicles. Summary of the invention

[0004] In view of this, the present application provides a three-level alarm identification method and device for a vehicle battery, aiming to accurately identify whether the vehicle battery of a new energy vehicle is in a three-level alarm state.

[0005] In a first aspect, the present application provides a three-level alarm identification method for a vehicle battery, the method comprising:

[0006] Receive real-time signal data uploaded by the target vehicle;

[0007] Performing data cleaning on the real-time signal data to obtain an input parameter set;

[0008] The input parameter set is input into a three-level alarm recognition model so as to use the three-level alarm recognition model to determine whether the battery of the target vehicle is in a three-level alarm state. The three-level alarm recognition model is a pre-trained bidirectional long short-term memory structure sequence model.

[0009] Optionally, the performing data cleaning on the real-time signal data to obtain an input parameter set includes:

[0010] Determine whether the real-time signal data has data missing or data anomaly;

[0011] If there is data missing, determine whether the missing value of the real-time signal data exceeds a preset missing range;

[0012] If exceeded, the real-time signal data is discarded;

[0013] If it does not exceed, the real-time signal data is supplemented by using the average value interpolation method;

[0014] If there is data anomaly, determine whether the abnormal value of the real-time signal data exceeds a preset abnormal range;

[0015] If exceeded, the real-time signal data is discarded;

[0016] If not, the abnormal value of the real-time signal data is replaced by an average interpolation method.

[0017] Optionally, the three-level alarm recognition model is trained in the following manner:

[0018] Acquire multiple groups of original signal data to establish a historical database, wherein the original signal data is vehicle signal data within an alarm time range, wherein the alarm time range is a preset time range of the alarm moment when the vehicle battery is in a level 3 alarm state, and the vehicle battery is of the same type as the battery of the target vehicle;

[0019] Determine a training data set according to the historical database;

[0020] Establish a bidirectional long and short-term memory structure model;

[0021] Iteratively optimizing the bidirectional long short-term memory structure model according to the training data set for multiple times;

[0022] A bidirectional long short-term memory structure model that meets the requirements is selected from multiple optimization results as the three-level alarm recognition model.

[0023] Optionally, acquiring multiple groups of original signal data to establish a historical database includes:

[0024] Preprocessing the multiple groups of original signal data, wherein the preprocessing includes cleaning the original signal data with missing data and abnormal data;

[0025] A historical database is established based on the preprocessed raw signal data.

[0026] Optionally, determining the training data set according to the historical database includes:

[0027] Determine the authenticity of each set of original signal data in the historical database, wherein the authenticity is used to indicate whether each set of original signal data actually triggers a third-level alarm state of the vehicle battery;

[0028] Determine the label corresponding to each group of original signal data according to the authenticity result corresponding to each group of original signal data;

[0029] Obtaining a positive sample data set and a negative sample data set according to the labels corresponding to each group of original signal data;

[0030] The training data set is obtained according to the positive sample data set and the negative sample data set.

[0031] Optionally, the iterative optimization of the bidirectional long short-term memory structure model based on the training data set for multiple times includes:

[0032] According to the training data set, the input parameters of the bidirectional long short-term memory structure model are obtained;

[0033] Determine optimization parameters, including optimizer type, number of hidden states, number of model convolutional layers, number of single iterations, and number of samples in a single iteration;

[0034] The input parameters are input into the bidirectional long short-term memory structure model so as to optimize the bidirectional long short-term memory structure model multiple times according to the optimization parameters.

[0035] Optionally, each group of original signal data in the historical database includes first signal data and a plurality of second signal data, and obtaining the input parameters of the bidirectional long short-term memory structure model according to the training data set includes:

[0036] Obtaining a reference sequence according to the labels corresponding to each group of original signal data in the historical database;

[0037] Obtaining a comparison series according to the first signal data, wherein the first signal data is original signal data corresponding to when the vehicle battery is in a level 3 alarm state;

[0038] Determine, according to the comparison sequence and the reference sequence, the correlation between the plurality of second signal data and the first signal data, the plurality of second signal data being original signal data corresponding to other moments within the alarm time range except the alarm moment;

[0039] The input parameter is obtained according to the correlation between the plurality of second signal data and the first signal data.

[0040] In a second aspect, the present application provides a three-level alarm identification device for a car battery, the device comprising: a signal data receiving module, a data cleaning module and a three-level alarm identification module;

[0041] The signal data receiving module is used to receive real-time signal data uploaded by the target vehicle;

[0042] The data cleaning module is used to clean the real-time signal data to obtain an input parameter set;

[0043] The three-level alarm recognition module is used to input the input parameter set into the three-level alarm recognition model so as to use the three-level alarm recognition model to identify whether the battery of the target vehicle is in the three-level alarm state. The three-level alarm recognition model is a pre-trained bidirectional long short-term memory structure sequence model.

[0044] Optionally, the data cleaning module is specifically used to:

[0045] Determine whether the real-time signal data has data missing or data anomaly;

[0046] If there is data missing, determine whether the missing value of the real-time signal data exceeds a preset missing range; if so, remove the real-time signal data; if not, use the average value interpolation method to supplement the real-time signal data;

[0047] If there is data anomaly, determine whether the abnormal value of the real-time signal data exceeds a preset abnormal range; if so, remove the real-time signal data; if not, use average interpolation to replace the abnormal value of the real-time signal data.

[0048] Optionally, the device further comprises a three-level alarm recognition model training module;

[0049] The three-level alarm recognition model training module is used to obtain multiple groups of original signal data to establish a historical database, wherein the original signal data is vehicle signal data within an alarm time range, and the alarm time range is a preset time range in which the alarm moment occurs when the vehicle battery is in a three-level alarm state, and the vehicle battery is of the same battery type as the target vehicle; determine a training data set based on the historical database; establish a bidirectional long-short-term memory structure model; iteratively optimize the bidirectional long-short-term memory structure model multiple times based on the training data set; and select a bidirectional long-short-term memory structure model that meets the requirements from multiple optimization results as the three-level alarm recognition model.

[0050] The present application provides a three-level alarm identification method for automobile batteries. When executing the method, first receive the real-time signal data uploaded by the target vehicle. Then, the real-time signal data is cleaned to obtain an input parameter set. Finally, the input parameter set is input into a three-level alarm identification model, and the three-level alarm identification model is used to determine whether the battery of the target vehicle is in a three-level alarm state. The three-level alarm identification model is a pre-trained bidirectional long-short-term memory structure sequence model. In this way, by pre-training the bidirectional long-short-term memory structure sequence model, a three-level alarm identification model is obtained, which can take into account the time series characteristics of the signal data before and after the alarm moment, so that the three-level alarm identification model can accurately and effectively identify whether the vehicle battery is in a three-level alarm state, thereby improving the accuracy of identifying the vehicle battery status. In this way, the faults of the vehicle battery can be analyzed and processed in a timely manner, reducing the risk of using new energy batteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 A method flow chart of a three-level alarm identification method for a vehicle battery provided in an embodiment of the present application;

[0053] Figure 2 Another method flow chart of the three-level alarm identification method for a car battery provided in an embodiment of the present application;

[0054] Figure 3 A data flow diagram of a bidirectional long short-term memory structure model provided in an embodiment of the present application;

[0055] Figure 4 A schematic diagram of the calculation process of the forget gate in the bidirectional long short-term memory structure model provided in an embodiment of the present application;

[0056] Figure 5 A schematic diagram of the calculation process of the input gate in the bidirectional long short-term memory structure model provided in an embodiment of the present application;

[0057] Figure 6 A schematic structural diagram of a three-level alarm identification device for a car battery provided in an embodiment of the present application. DETAILED DESCRIPTION

[0058] As mentioned above, according to the GB / T 32960 standard, the fault alarm levels of electric vehicles can be divided into level 1, level 2 and level 3, among which level 3 alarm is the most serious fault. In order to meet the requirements of the standard, it is necessary not only to analyze and process the alarm status of the on-board battery in a timely manner, but also to accurately identify whether the on-board battery of the electric vehicle is in the level 3 alarm state. However, there is no method in the prior art that can accurately and effectively identify whether the on-board battery of a new energy vehicle is in the level 3 alarm state.

[0059] In view of this, the present application provides a three-level alarm identification method and device for automobile batteries. The method includes: first receiving real-time signal data uploaded by the target vehicle. Then, the real-time signal data is cleaned to obtain an input parameter set. Finally, the input parameter set is input into a three-level alarm identification model, and the three-level alarm identification model is used to determine whether the battery of the target vehicle is in a three-level alarm state. The three-level alarm identification model is a pre-trained bidirectional long-short-term memory structure sequence model. In this way, by pre-training the bidirectional long-short-term memory structure sequence model, a three-level alarm identification model is obtained, which can take into account the time series characteristics of the signal data before and after the alarm moment, so that the three-level alarm identification model can accurately and effectively identify whether the vehicle battery is in a three-level alarm state, thereby improving the accuracy of identifying the vehicle battery state. In this way, the fault of the vehicle battery can be analyzed and processed in time, and the problem of thermal dissipation or thermal runaway caused by the battery working under too high or too low temperature can be solved in time, so as to improve the overall performance of the battery and reduce the use risk of new energy batteries.

[0060] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0061] The method provided in the embodiment of the present application can be applied to the new energy monitoring background of a car company, so as to promptly determine whether the vehicle's on-board battery is in a third-level alarm state after receiving a signal uploaded by the vehicle in real time.

[0062] See also Figure 1 , Figure 1 A method flow chart of a three-level alarm identification method for a car battery provided in an embodiment of the present application includes:

[0063] S101: receiving real-time signal data uploaded by a target vehicle.

[0064] Among them, the real-time signal data may include multiple signals uploaded by the target vehicle at the current moment, which are used for information exchange within the vehicle and data transmission between the vehicle and the vehicle company's backend.

[0065] Optionally, the TBOX communication module of the target vehicle can be used for real-time acquisition and uploading. Other methods can also be used to obtain real-time signal data, which will not affect the normal implementation of the embodiment of the present application.

[0066] S102: Clean the real-time signal data to obtain an input parameter set.

[0067] The input parameter set may include alarm information uploaded by multiple target vehicles, and the alarm information is stored in the input parameter set in the form of a multi-dimensional matrix so as to serve as input parameters of the three-level alarm recognition model.

[0068] Data cleaning is the process of re-examining and verifying data to find and correct identifiable errors in real-time signal data. It is usually done automatically by a computer. The purpose is to delete duplicate information, correct existing errors, and provide data consistency.

[0069] Optionally, the real-time signal data can be cleaned in the following manner: First, determine whether the real-time signal data has data missing or data anomaly. If there is data missing, determine whether the missing value of the real-time signal data exceeds the preset missing range. If it exceeds, the real-time signal data is removed; if it does not exceed, the real-time signal data is supplemented by the average value interpolation method. If there is data anomaly, determine whether the abnormal value of the real-time signal data exceeds the preset abnormal range. If it exceeds, the real-time signal data is removed; if it does not exceed, the abnormal value of the real-time signal data is replaced by the average interpolation method.

[0070] Among them, the average interpolation method is a data processing method in statistics, which uses the mean to fill in the missing part of the data, or uses the mean to replace the abnormal part of the data. Of course, other methods can also be used to correct the real-time signal data, which will not affect the normal implementation of the embodiment of the present application.

[0071] Optionally, in order to eliminate the dimensional influence between indicators, data standardization is required to solve the comparability between data indicators. The real-time signal data can also be normalized and the input parameter set is generated so that the real-time signal data after data cleaning is limited to a certain range (such as [0,1] or [-1,1]), thereby eliminating the adverse effects caused by singular sample data.

[0072] S103: Inputting the input parameter set into the three-level alarm recognition model, and using the three-level alarm recognition model to determine whether the battery of the target vehicle is in the three-level alarm state.

[0073] Among them, the three-level alarm recognition model is a pre-trained bidirectional long short-term memory structure sequence model.

[0074] Among them, the Long Short Term Memory (LSTM) model is a type of neural network model, which is usually used to process time series data. It can capture long-term dependencies in text, thereby improving the accuracy of the model.

[0075] Optionally, the three-level alarm recognition model can be trained in the following manner: First, multiple groups of original signal data are obtained to establish a historical database. The original signal data is the vehicle signal data within the alarm time range, and the alarm time range is the preset time range of the alarm moment when the vehicle battery is in the three-level alarm state, and the vehicle battery is of the same battery type as the target vehicle. Then, a training sample set is determined based on the historical database. A bidirectional long-short-term memory structure model is then established, and the bidirectional long-short-term memory structure model is iteratively optimized multiple times based on the training sample set. Finally, a bidirectional long-short-term memory structure model that meets the requirements is selected from multiple optimization results as the three-level alarm recognition model.

[0076] The historical database may be a collection of battery-related signals. In order to ensure the accuracy of the training of the three-level alarm recognition model, the original signal data is the signal data uploaded to the background by different vehicles with the same battery type as the target vehicle.

[0077] Optionally, the training sample set can be obtained in the following manner: first, determine the authenticity of each group of original signal data in the historical database, and the authenticity is used to indicate whether each group of original signal data actually triggers the third-level alarm state of the vehicle battery. Then, determine the label corresponding to each group of original signal data according to the authenticity result corresponding to each group of original signal data. Then, obtain the positive sample data set and the negative sample data set according to the label corresponding to each group of original signal data. Finally, obtain the training sample set according to the positive sample data set and the negative sample data set.

[0078] Optionally, after-sales return visits can be used to determine from customers whether the vehicle signal data triggers a true third-level alarm or a false third-level alarm state. Of course, other methods can also be used to determine whether the vehicle signal data is true alarm data, which does not affect the normal implementation of the embodiments of the present application.

[0079] Optionally, acquiring multiple groups of original signal data to establish a historical database includes preprocessing the multiple groups of original signal data, wherein the preprocessing includes cleaning the original signal data with missing data and abnormal data, and establishing a historical database based on the preprocessed original signal data.

[0080] The vehicle signal data at each moment is composed of multiple different signals. If there is more than one missing signal in the vehicle signal data at a certain moment, the vehicle signal data is considered to be missing. If multiple vehicle signal data are missing, and the number of missing vehicle signal data exceeds the preset number, then this group of real alarm data or false alarm data is considered to be seriously missing data, and the group of real alarm data or false alarm data is eliminated. If the number of missing vehicle signal data exceeds the preset number, the interpolation method is used to supplement the missing signal data. When there is an abnormality in the vehicle signal data, it is consistent with the above-mentioned method for handling missing vehicle signal data, and will not be repeated here.

[0081] Optionally, the pre-established bidirectional LSTM model can be optimized by using the training sample set for multiple iterations in the following manner: According to the training data set, the input parameters of the bidirectional long short-term memory structure model are obtained. The optimization parameters are determined, and the optimization parameters include the optimizer type, the number of hidden states, the number of model convolution layers, the number of single iterations, and the number of single iteration samples. The input parameters are input into the bidirectional long short-term memory structure model so that the bidirectional long short-term memory structure model is optimized multiple times according to the optimization parameters.

[0082] Optionally, each group of original signal data in the historical database may include first signal data and multiple second signal data. The first signal data is the original signal data corresponding to when the vehicle battery is in the third-level alarm state, and the second signal data is the original signal data corresponding to other times within the alarm time range except the alarm time. In some possible implementations, the following method can be used to extract features from the original signal data to obtain input parameters: According to the labels corresponding to each group of original signal data in the historical database, a reference series is obtained. According to the first signal data, a comparison series is obtained. According to the comparison series and the reference series, the correlation between the multiple second signal data and the first signal data is determined. According to the correlation corresponding to each second signal data, the input parameter is obtained.

[0083] Among them, the correlation degree is used to measure the correlation between the signal data corresponding to each moment and whether the alarm is true or false. The reference series is a data sequence that reflects the characteristics of system behavior and is used as the evaluation standard for correlation analysis. The comparison series is a data sequence composed of factors that affect system behavior and is used as the evaluation object of correlation analysis.

[0084] The embodiment of the present application obtains a three-level alarm recognition model by pre-training a bidirectional long-short-term memory structure sequence model, which can take into account the time series characteristics of the signal data before and after the alarm moment, so that the three-level alarm recognition model can accurately and effectively determine whether the vehicle battery is in a three-level alarm state, thereby improving the accuracy of identifying the vehicle battery state. In this way, the true state of the battery can be effectively identified, which is conducive to timely determining whether a new energy battery has failed. For example, a high temperature alarm of the battery can be discovered in time, and the battery can be cooled through a cooling pipeline, thereby reducing the risk of using new energy batteries.

[0085] In the embodiment of the present application, the above Figure 1 There are many possible training methods for the three-level alarm recognition model, which are introduced below. It should be noted that the implementation methods given in the following introduction are only for exemplary purposes and do not represent all implementation methods of the embodiments of the present application.

[0086] See also Figure 2 , which is another method flow chart of the three-level alarm identification method for a car battery provided in an embodiment of the present application, the method comprising:

[0087] S201: Collect a large amount of original signal data to establish a historical database.

[0088] Optionally, the vehicle battery can be determined to be in a level 3 alarm state based on the internal conditions of the vehicle battery, and the moment when the vehicle battery is in the level 3 alarm state can be determined to obtain the alarm record of the vehicle battery. Then, based on the obtained alarm record, a large number of vehicle-related original signals of the vehicle battery before and after the moment when the level 3 alarm state is triggered are obtained to establish a historical database. In some possible implementations, all original signals of the battery vehicle for a total of 61 seconds before and after the level 3 alarm is triggered can be intercepted as multiple groups of original signal data.

[0089] Each set of collected raw signal data may include raw signal data at different times. The raw signal data at each time may include multiple signals uploaded by a vehicle containing a new energy vehicle battery at that time.

[0090] Optionally, the collected original signal data may be preprocessed before establishing the historical database, or the original signal data in the historical database may be preprocessed after the historical database is obtained directly based on the collected original signal data.

[0091] The processing of the original signal data includes processes such as adding labels to the original signal data, data cleaning and feature extraction.

[0092] Optionally, a label may be added to each set of raw signal data in the following manner: determine the alarm time corresponding to each set of raw signal data according to the vehicle battery alarm record, and confirm with the user through a return visit the real state of the vehicle battery at that time, whether there is a false alarm, whether maintenance is required, and the maintenance result, etc. The label of each set of raw signal data is determined according to the return visit result corresponding to each set of raw signal data.

[0093] As an example, the vehicle signal data and corresponding labels of 30 seconds before and after the vehicle battery reaches the third-level alarm state are intercepted to form a historical database. Each group of vehicle signal data in the historical database includes a 61-second vehicle signal data set. According to the labels used to distinguish between real alarms and false alarms, the data in the historical database can be divided into real alarm data sets and false alarm data sets.

[0094] Among them, the positive sample data set can be T:

[0095]

[0096] Among them, T is the real alarm data set, x n (1) is the real alarm data of the first moment of the nth group. The superscripts (1), (2)…(t) represent the time, and the maximum value t is 61 seconds.

[0097] The negative sample dataset can be F:

[0098]

[0099] Among them, F is the false alarm data set, y n (1) is the real alarm data of the first moment of the nth group. The superscripts (1), (2)…(t) represent the time, and the maximum value t is 61 seconds.

[0100] In some possible embodiments, data cleaning may be performed in the following manner: As an example, the historical database may include vehicle signal data x n (1) . Where n is used to represent the nth group of vehicle signal data, and the subscript (1) is used to represent the time corresponding to the vehicle signal data. n (1) This vehicle signal data, x n (1) Can include And other signals, among which, The superscript (1) indicates the time, the subscript 1 indicates the first alarm time, and m indicates that the signal is the mth signal.

[0101] like When one or more signals are missing in this signal group, determine x n (1) This vehicle signal data has data missing. This signal group has 12 or more signals missing, confirm x n (1) This vehicle signal data is missing important data, so it is removed. If the missing values ​​of the signal group are less than 12, the missing signals can be supplemented by interpolation.

[0102] Optionally, in some embodiments, if the signal If missing, the value can be and If the missing signal is Since it is the value of the first second, this value is the same as the value of the next moment. Keep consistent. and The superscripts (1), (2) and (3) indicate time, the subscript 1 indicates the first alarm time, and m indicates that the signal is the mth signal.

[0103] In some possible implementations, different vehicle signal data have different valid data ranges, and other data may be used as a threshold for whether data is missing, which does not affect the normal implementation of the embodiments of the present application.

[0104] Optionally, the GRA grey correlation analysis method can be used to extract features from the original signal data at each moment, so as to extract the time series characteristics of the signal data before and after the vehicle battery is in the third-level alarm state. The method is as follows: First, determine the reference sequence and the comparison sequence. Among them, the labels corresponding to each group of original signal data are used as the reference sequence, and the data list composed of each signal in each original signal data is used as the comparison sequence. For example, when the moment when the vehicle battery is in the third-level alarm state is the 31st second, the comparison sequence can be:

[0105]

[0106]

[0107]

[0108] Alternatively, due to the different physical meanings of various factors in the system, the dimensions of the data may not be the same, which makes it difficult to compare or to obtain correct conclusions when comparing. Therefore, before feature extraction, each signal needs to be de-dimensionalized.

[0109] Optionally, to signal For example, we can use the following method to dedimensionalize the signal: First, The minimum value is centered according to the following formula:

[0110]

[0111] Finally, the correlation coefficient between the original signal at each moment and the vehicle battery triggering the third-level alarm state is calculated based on the reference sequence and the comparison sequence. The calculation formula is as follows:

[0112]

[0113] Among them, ξ i (k) represents the correlation coefficient between the kth data in the i-th comparison sequence and the kth data in the reference sequence, and ρ is the resolution coefficient, where ρ = 0.5. Finally, the correlation between the original signal at each moment and the vehicle battery triggering the third-level alarm state is calculated and sorted according to the correlation coefficient corresponding to the signal data at each moment.

[0114] As an example, in the above feature extraction process, 43 highly correlated signals as shown in Table 1 can be obtained, and the input parameter can be a 61*43 multidimensional matrix, 61 represents the time dimension, and 43 represents the specific signal dimension.

[0115] Table 1

[0116]

[0117]

[0118] S202: Obtain a training sample set according to the historical database.

[0119] Optionally, the data in the real alarm data set in the historical database is used as a positive sample to obtain a positive sample data set, and the data in the false alarm data set in the historical database is used as a negative sample to obtain a negative sample data set.

[0120] The positive sample data set may include multiple sets of real alarm data, each set of real alarm data contains vehicle signal data at different times. Similarly, the negative sample data set may include multiple sets of false alarm data, each set of false alarm data contains vehicle signal data at different times.

[0121] S203: Perform training optimization on the established bidirectional long short-term memory structure model according to the training sample set.

[0122] Among them, see Figure 3The bidirectional long short-term memory (LSTM) model includes a forget gate ①, an input gate ②, and an output gate ④. The input gate controls whether the new input information is added to the cell state, the forget gate controls whether the old information is retained in the cell state, and the output gate controls whether the information in the cell state is output. The functions of these gates can be automatically adjusted through the learning of the neural network, thereby realizing the modeling and prediction of sequence data.

[0123] As an example, set the size of the hidden state to 64, and assume that at time (t-1) the LSTM network is used and the output C is obtained. (t-1) and the hidden state h output at time (t-1) (t-1) , and a new signal x is added at time (t-1) (t) , the signal dimension is 1 row * 43 columns, as shown below:

[0124]

[0125] In this bidirectional LSTM model, first, the hidden state h (t-1) , output C (t-1) And the new signal x (t) Enter the forget gate ① and use the following formula for calculation:

[0126] f (t) =σ(W f ·[h (t-1) ,x (t) ]+b f )

[0127] Among them, W f is the parameter matrix of the forget gate, b f is the bias coefficient, and σ represents the sigmoid activation function.

[0128] See also Figure 4 As shown in this formula, h (t-1) and x (t) First form the matrix α. h (t-1) The dimension of is 64 rows * 1 column, which is converted to a column vector with a dimension of 43 rows * 1 column. Then the dimension of α is 107 rows and 1 column. α is the parameter matrix of the forget gate, and its number of rows is the same as h (t-1) Same as 64, number of columns is h (t-1) +x (t) is 107. f The initial value of is randomly distributed, and is continuously optimized by gradient descent during model training. The number of parameters of the forget gate is 64×(64+43+1)×2, where 1 is the bias coefficient b of the forget gate. f , the coefficient 2 is used to make the LSTM model have parameter matrices in two directions, becoming a bidirectional recurrent LSTM model.f The product with the matrix α is compressed between [0, 1] by σ (i.e., the sigmoid activation function).

[0129] In this bidirectional LSTM model, the hidden state h (t-1) , output C (t-1) And the new signal x (t) After entering the forget gate, the hidden state h (t-1) And the new signal x (t) Enter the input gate ② and use the following formula to calculate and get i (t) and

[0130] i (t) =σ(W i ·[h (t-1) ,x (t) ]+b i )

[0131]

[0132] Among them, W i is the parameter matrix of the input gate, b i is the bias coefficient, and σ represents the sigmoid activation function.

[0133] See also Figure 5 As shown, in the above calculation (t) In the formula, h (t-1) The dimension is 64 rows * 1 column, x (t) is converted to a column vector with a dimension of 43 rows * 1 column, and h (t-1) and x (t) The matrix β is composed of 107 rows and 1 column. i is the parameter matrix of the input gate, and its number of rows is the same as h (t-1) Same as 64, number of columns is h (t-1) +x (t) is 107.

[0134] Among them, W i The initial value of is randomly distributed, and can be continuously optimized by gradient descent during model training. The number of parameters of the input gate is 64×(64+43+1)×2, where 1 is the bias coefficient b of the forget gate i , the coefficient 2 is used to make the LSTM model have parameter matrices with the same dimensions in two directions, resulting in a bidirectional recurrent LSTM model. i The product with the matrix β is compressed between [0,1] by σ (i.e., the sigmoid activation function).

[0135] Optionally, the new state of the input gate The generation process of i (t) Almost the same, but The matrix product needs to be passed through the tanh activation function in the last step. The number of parameters of the input gate is also the same as the number of forget gates. and i (t) Both are 64×(64+43+1)×2.

[0136] See attached Figure 3 The serial number ③ in the above process is f (t) and C (t-1) Need to perform bitwise multiplication, and the output obtained by the input gate and i (t) Using the following formula, we can get C (t) :

[0137]

[0138] Optionally, the above C (t-1) C can be obtained by (t) is obtained in the same way.

[0139] C (t) Used to pass through the output gate to get h (t) , and then use the following formula to get the judgment value of whether the third-level alarm state is true:

[0140] o (t) =σ(W o ·[h (t-1) ,x (t) ]+b o )

[0141] h (t) =o t *tanh(C (t) )

[0142] Among them, the calculation process of the output gate includes: h (t-1) With x (t) After being spliced ​​into a column vector of 107 rows and 1 column, it is activated by the sigmoid function and C is activated by the tanh function. (t) Multiply bit by bit to get h (t) The number of parameters in this step is also 64×(64+43+1)×2.

[0143] Optionally, samples of 61 rows and 43 columns of the input model can be selected from the training sample set each time, and passed through a bidirectional LSTM model with one convolution layer and 64 hidden states, with a parameter number of 55296. In this process, the LSTM model passes through a one-dimensional dense convolution of (None, n, 64) and a one-dimensional dense convolution of (None, n, 32), and then enters a dropout layer with a ratio of 0.2 to clear the local network to reduce overfitting, and finally outputs a warning result.

[0144] Optionally, the number of training cycles, the number of optimization states h, the number of LSTM layers and the number of subsequent convolutional layers, as well as the optimizer type, the number of iterations, the learning rate, the number of samples per single iteration and other parameters may be predetermined to perform iterative optimization of the model.

[0145] As an example, in some possible embodiments, the optimal LSTM model is 1 layer, the number of hidden states in the first layer is 64, and the optimizer is Adam. The number of Dense layers connected after the LSTM layer is 2, with 64 and 32 hidden states respectively, the optimal learning rate is 500, the optimal number of single training samples is 50, and the optimal number of iterations is 7. In some possible implementations, other data may be used to draw different conclusions about the optimal model, which does not affect the normal implementation of the embodiments of the present application.

[0146] S204: Obtain a three-level alarm recognition model according to the optimization result.

[0147] Optionally, the model can be deployed in the new energy monitoring background of the car company to receive the signals uploaded by the vehicle in real time and determine the status of the vehicle battery.

[0148] The embodiment of the present application uses the GRA grey correlation analysis method to screen vehicle signals related to the battery triggering the third-level alarm state, thereby taking into account the time series characteristics of the vehicle signal data before and after the alarm occurs. The bidirectional LSTM sequence model established in this way can effectively identify whether the battery state is in the third-level alarm state, reducing the risk of using new energy batteries.

[0149] The above are some specific implementations of the three-level alarm identification method for automobile batteries provided in the embodiments of the present application. Based on this, the present application also provides a corresponding device. The device provided in the embodiments of the present application will be introduced from the perspective of functional modularization.

[0150] See also Figure 6 The structure diagram of the three-level alarm identification device 300 for automobile batteries is shown. The device 300 includes a signal data receiving module 301, a data cleaning module 302 and a three-level alarm identification module 303.

[0151] The signal data receiving module 301 is used to receive the real-time signal data uploaded by the target vehicle;

[0152] The data cleaning module 302 is used to clean the real-time signal data to obtain an input parameter set;

[0153] The three-level alarm recognition module 303 is used to input the input parameter set into the three-level alarm recognition model so as to use the three-level alarm recognition model to identify whether the battery of the target vehicle is in the three-level alarm state. The three-level alarm recognition model is a pre-trained bidirectional long short-term memory structure sequence model.

[0154] Optionally, the three-level alarm recognition model can be obtained by training and optimization using any of the methods described in the above embodiments.

[0155] The “first” and “second” in the names such as “first” and “second” mentioned in the embodiments of the present application are only used as name identifiers and do not represent the first or second in order.

[0156] Through the description of the above implementation methods, it can be known that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a read-only memory (ROM) / RAM, a magnetic disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in each embodiment of the present application or some parts of the embodiments.

[0157] Each embodiment in this specification is described in a progressive manner, and the same and similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0158] The above description is merely an exemplary embodiment of the present application and is not intended to limit the protection scope of the present application.

Claims

1. A three-level alarm identification method for automobile batteries, characterized in that: The method comprises: Receive real-time signal data uploaded by the target vehicle; Performing data cleaning on the real-time signal data to obtain an input parameter set; The input parameter set is input into a three-level alarm recognition model, and the three-level alarm recognition model is used to determine whether the battery of the target vehicle is in a three-level alarm state. The three-level alarm recognition model is a pre-trained bidirectional long short-term memory structure sequence model.

2. The method according to claim 1, characterized in that The real-time signal data is cleaned to obtain an input parameter set, including: Determine whether the real-time signal data has data missing or data anomaly; If there is data missing, determine whether the missing value of the real-time signal data exceeds a preset missing range; If exceeded, the real-time signal data is discarded; If it does not exceed, the real-time signal data is supplemented by using the average value interpolation method; If there is data anomaly, determine whether the abnormal value of the real-time signal data exceeds a preset abnormal range; If exceeded, the real-time signal data is discarded; If not, the abnormal value of the real-time signal data is replaced by an average interpolation method.

3. The method according to claim 1, characterized in that The three-level alarm recognition model is trained in the following way: Acquire multiple groups of original signal data to establish a historical database, wherein the original signal data is vehicle signal data within an alarm time range, wherein the alarm time range is a preset time range of the alarm moment when the vehicle battery is in a level 3 alarm state, and the vehicle battery is of the same type as the battery of the target vehicle; Determine a training sample set according to the historical database; Establish a bidirectional long and short-term memory structure model; Iteratively optimizing the bidirectional long short-term memory structure model multiple times according to the training sample set; A bidirectional long short-term memory structure model that meets the requirements is selected from multiple optimization results as the three-level alarm recognition model.

4. The method according to claim 3, characterized in that The step of acquiring multiple groups of original signal data and establishing a historical database includes: Preprocessing the multiple groups of original signal data, wherein the preprocessing includes cleaning the original signal data with missing data and abnormal data; A historical database is established based on the preprocessed raw signal data.

5. The method according to claim 3, characterized in that: Determining the training sample set according to the historical database includes: Determine the authenticity of each set of original signal data in the historical database, wherein the authenticity is used to indicate whether each set of original signal data actually triggers a third-level alarm state of the vehicle battery; Determine the label corresponding to each group of original signal data according to the authenticity result corresponding to each group of original signal data; Obtaining a positive sample data set and a negative sample data set according to the labels corresponding to each group of original signal data; The training sample set is obtained according to the positive sample data set and the negative sample data set.

6. The method according to claim 3, characterized in that: The iterative optimization of the bidirectional long short-term memory structure model based on the training sample set for multiple times includes: According to the training sample set, the input parameters of the bidirectional long short-term memory structure model are obtained; Determine optimization parameters, including optimizer type, number of hidden states, number of model convolutional layers, number of single iterations, and number of samples in a single iteration; The input parameters are input into the bidirectional long short-term memory structure model so as to optimize the bidirectional long short-term memory structure model multiple times according to the optimization parameters.

7. The method according to claim 6, characterized in that Each group of original signal data in the historical database includes first signal data and multiple second signal data. The input parameters of the bidirectional long short-term memory structure model are obtained according to the training sample set, including: Obtaining a reference sequence according to the labels corresponding to each group of original signal data in the historical database; Obtaining a comparison series according to the first signal data, wherein the first signal data is original signal data corresponding to when the vehicle battery is in a level 3 alarm state; Determine, according to the comparison sequence and the reference sequence, the correlation between the plurality of second signal data and the first signal data, the plurality of second signal data being original signal data corresponding to other moments within the alarm time range except the alarm moment; The input parameter is obtained according to the correlation between the plurality of second signal data and the first signal data.

8. A three-level alarm identification device for automobile batteries, characterized in that: The device comprises: a signal data receiving module, a data cleaning module and a three-level alarm identification module; The signal data receiving module is used to receive real-time signal data uploaded by the target vehicle; The data cleaning module is used to clean the real-time signal data to obtain an input parameter set; The three-level alarm recognition module is used to input the input parameter set into a three-level alarm recognition model, and use the three-level alarm recognition model to identify whether the battery of the target vehicle is in a three-level alarm state. The three-level alarm recognition model is a pre-trained bidirectional long short-term memory structure sequence model.

9. The device according to claim 8, characterized in that The data cleaning module is specifically used for: Determine whether the real-time signal data has data missing or data anomaly; If there is data missing, determine whether the missing value of the real-time signal data exceeds a preset missing range; If it exceeds, the real-time signal data is discarded; if it does not exceed, the real-time signal data is supplemented by using the average value interpolation method; If there is data anomaly, determine whether the abnormal value of the real-time signal data exceeds a preset abnormal range; If it exceeds, the real-time signal data is discarded; if it does not exceed, the abnormal value of the real-time signal data is replaced by the average interpolation method.

10. The device according to claim 8, characterized in that The device also includes a three-level alarm recognition model training module; The three-level alarm recognition model training module is used to obtain multiple groups of original signal data to establish a historical database, wherein the original signal data is vehicle signal data within an alarm time range, and the alarm time range is a preset time range of the alarm moment when the vehicle battery is in a three-level alarm state, and the vehicle battery is of the same battery type as the target vehicle; determine a training sample set based on the historical database; establish a bidirectional long-short-term memory structure model; iteratively optimize the bidirectional long-short-term memory structure model multiple times based on the training sample set; and select a bidirectional long-short-term memory structure model that meets the requirements from multiple optimization results as the three-level alarm recognition model.

Images