Battery pack thermal runaway early warning method and device

Abstract

The invention discloses a battery pack thermal runaway early warning method and device, and the method comprises the steps: collecting the pressure data, multiple pieces of state signal data, temperature data and multiple pieces of video data of a to-be-early-warned battery pack, carrying out the frame image extraction of each piece of video data, and obtaining a first frame image and a second frame image, performing image semantic recognition on the first frame of image by using the trained first target recognition model to obtain a first recognition result, and performing character labeling processing on the second frame of image by using the trained second target recognition model to obtain a second recognition result, according to the method, the pressure data, the multiple pieces of state signal data, the temperature data, the first recognition result and the second recognition result are judged, multiple judgment results are obtained, when any judgment result meets the corresponding early warning condition, early warning information is sent out, any abnormal signal can be accurately detected through the method, and early warning prompt is immediately triggered.

Description

Technical Field

[0001] The present invention relates to the technical field of battery pack thermal runaway early warning, and particularly to a method and device for battery pack thermal runaway early warning. Background Art

[0002] With the wide application of lithium-ion batteries in fields such as electric vehicles, energy storage systems, and portable electronic devices, the safety and reliability issues of lithium batteries have attracted increasing attention. In recent years, the frequent occurrence of lithium battery fires and explosions has not only caused huge economic losses but also posed a serious threat to public safety. Therefore, how to effectively identify and predict lithium battery thermal runaway events and control them in a timely manner has become a key problem that urgently needs to be solved in the lithium battery detection industry.

[0003] Currently, the lithium battery detection industry mainly relies on traditional fault diagnosis methods and limited empirical judgments. These methods have many limitations when faced with complex and variable lithium battery thermal runaway events. Traditional thermal runaway triggering methods usually rely on specific and single thermal runaway triggering signals, such as battery pack fire and smoke, voltage abnormality, temperature abnormality, etc. The detection and analysis of these signals often require a large amount of manual intervention and complex feature engineering processing. In addition, traditional methods often show low accuracy and generalization ability when dealing with large-scale multi-type data, and it is difficult to adapt to the rapid change of lithium battery technology iteration and application scenarios. Summary of the Invention

[0004] To solve the above technical problems, embodiments of the present invention provide a method and device for battery pack thermal runaway early warning to solve the technical problem that the accuracy of the existing technology in data processing for battery thermal runaway using traditional methods is low, resulting in low early warning accuracy.

[0005] The first aspect of the embodiments of the present invention provides a method for battery pack thermal runaway early warning, and the method includes:

[0006] Collect relevant data of the battery pack to be warned, where the relevant data includes pressure data, multiple status signal data, temperature data, and multiple video data;

[0007] Extract frame images from each video data to obtain a first frame image and a second frame image. Use the trained first target recognition model to perform image semantic recognition on the first frame image to obtain a first recognition result. Use the trained second target recognition model to perform character annotation processing on the second frame image to obtain a second recognition result, where the first target recognition model is obtained by training the DeepLab V3+ network using a first sample data set, and the second target recognition model is obtained by training the CRNN network using a second sample data set;

[0008] Judge the pressure data, multiple status signal data, temperature data, the first recognition result and the second recognition result respectively to obtain multiple judgment results. When any judgment result reaches the corresponding warning condition, a warning message is issued.

[0009] In a possible implementation manner of the first aspect, use the trained first target recognition model to perform image semantic recognition on the first frame of image, and obtain the first recognition result, including:

[0010] Input the first frame of image into the feature extraction network layer of the first target recognition model to perform feature extraction, and obtain the first feature extraction result;

[0011] Use the atrous spatial pyramid pooling module of the trained first target recognition model to perform multi-scale feature extraction and fusion on the first feature extraction result, and obtain the second feature extraction result;

[0012] Use the output layer to process the second feature extraction result to obtain the first recognition result.

[0013] In a possible implementation manner of the first aspect, the first target recognition model is obtained by training the DeepLab V3+ network using the first sample data set, including:

[0014] Obtain the first sample data set, and divide the first sample data set into a first training set and a first test set according to a preset ratio;

[0015] Perform contrast processing and brightness processing on each sample image in the first training set to obtain the processed first training set, add Gaussian noise to the processed first training set to obtain the enhanced first training set, and use the Laplace edge detection method to perform edge detection on the enhanced first training set to obtain the finally processed first training set;

[0016] Input the finally processed first training set into the DeepLab V3+ network for training to obtain the initial first target recognition model, and then use the sample images in the first test set to adjust the initial first target recognition model to obtain the trained first target recognition model.

[0017] In a possible implementation manner of the first aspect, use the sample images in the first test set to adjust the initial first target recognition model to obtain the trained first target recognition model, including:

[0018] Input the sample images of the first test set into the initial first target recognition model for image recognition to obtain the first test recognition result, and obtain the current intersection over union ratio according to the first test recognition result and the first ground truth result, where the calculation formula of the current intersection over union ratio is:

[0019]

[0020] Wherein, TP is the true positive sample in the first test recognition result, FP is the false positive sample in the first test recognition result, and FN is the false negative sample in the first test recognition result;

[0021] After the current intersection over union ratio meets the first preset condition, the number of network layers in the initial first target recognition model is reduced by one layer to obtain the trained first target recognition model. If the current intersection over union ratio does not meet the preset condition, continue to obtain the first sample data set to train the DeepLab V3+ network until the trained first target recognition model is obtained. Among them, the first preset condition is that the current intersection over union ratio is less than the intersection over union ratio calculated during the previous training of the DeepLab V3+ network.

[0022] In a possible implementation manner of the first aspect, using the second target recognition model to perform character annotation processing on the second frame of image to obtain the second recognition result, including:

[0023] Using the convolutional layer of the second target recognition model to extract features from the second frame of image to obtain the thermal imaging extraction result;

[0024] Using the recurrent layer of the second target recognition model to process the extraction result of the second frame of image to obtain serialized features, where the recurrent layer is constructed based on the LSTM network;

[0025] Using the transcription layer of the second target recognition model to perform transformation processing on the serialized features to obtain the second recognition result, where the transcription layer is constructed by connecting the connectionist temporal classification algorithm.

[0026] In a possible implementation manner of the first aspect, the second target recognition model is obtained by training the CRNN network using the second sample data set, including:

[0027] Obtain the second sample data set, extract the target regions of each sample image in the second sample data set to obtain multiple target images, and perform grayscale processing, normalization processing, and denoising processing on each target image to obtain the processed second sample data set;

[0028] Divide the processed second sample data set into a second training set and a second test set according to a preset ratio, input the sample data in the second training set into the CRNN network for training to obtain the initial second target recognition model, and use the sample data in the second test set to adjust the initial second target recognition model to obtain the trained second target recognition model.

[0029] In a possible implementation of the first aspect, the initial second target recognition model is adjusted using the sample data in the second test set to obtain a trained second target recognition model, including:

[0030] Input the sample data in the second test set into the initial second target recognition model for image recognition to obtain a second test recognition result. According to the second test recognition result and the second true result, the current character error rate is obtained. The calculation formula of the current character error rate is:

[0031]

[0032] In the formula, N is the length of the original character string in the true result, S is the number of replaced characters, D is the number of deleted characters, and I is the number of inserted characters;

[0033] After the current character error rate meets the second preset condition, the number of network layers in the initial second target recognition model is reduced by one layer to obtain a trained second target recognition model, and the initial second target recognition model is determined as the trained second target recognition model. If the character error rate does not meet the preset condition, continue to obtain the second sample data set to train the CRNN network until a trained second target recognition model is obtained. The second preset condition is that the current character error rate is less than the character error rate calculated during the previous training of the CRNN network.

[0034] In a possible implementation of the first aspect, the pressure data, multiple state signal data, temperature data, the first recognition result, and the second recognition result are respectively judged to obtain multiple judgment results. When any judgment result reaches the corresponding warning condition, a warning message is issued, including:

[0035] Judge whether the pressure data is greater than the pressure threshold. If it is greater, issue a warning message of abnormal pressure;

[0036] Judge whether each state signal data is greater than the corresponding state signal threshold. If it is greater, issue a warning message of abnormal device state;

[0037] According to the temperature data, calculate the rising rate of the battery pack to be warned within a preset time period, and judge whether the rising rate is greater than the preset rate threshold. If it is greater, issue a warning message of abnormal internal temperature;

[0038] Judge whether the battery pack to be warned is on fire according to the first recognition result. If it is on fire, issue a warning message of battery pack fire. Judge whether the surface temperature of the battery pack to be warned is greater than the preset temperature threshold according to the second recognition result. If it is greater, issue a warning message of abnormal external temperature.

[0039] To solve the same technical problem, a second aspect of the embodiments of the present invention provides a battery pack thermal runaway warning device, including a collection module, an identification module, and a warning module, where,

[0040] The collection module is used to collect relevant data of the battery pack to be warned, where the relevant data includes pressure data, a plurality of status signal data, temperature data, and a plurality of video data;

[0041] The identification module is used to extract frame images from each video data to obtain a first frame image and a second frame image, perform image semantic recognition on the first frame image using a trained first target recognition model to obtain a first recognition result, and perform character annotation processing on the second frame image using a trained second target recognition model to obtain a second recognition result, where the first target recognition model is obtained by training the DeepLab V3+ network using a first sample data set, and the second target recognition model is obtained by training the CRNN network using a second sample data set;

[0042] The warning module is used to respectively judge the pressure data, a plurality of status signal data, temperature data, the first recognition result, and the second recognition result to obtain a plurality of judgment results. When any judgment result reaches the corresponding warning condition, a warning message is sent.

[0043] In a possible implementation manner of the second aspect, the identification module includes a first extraction unit, a second extraction unit, and a processing unit, where,

[0044] The first extraction unit is used to input the first frame image into the feature extraction network layer of the first target recognition model for feature extraction to obtain a first feature extraction result;

[0045] The second extraction unit is used to perform multi-scale feature extraction and fusion on the first feature extraction result using the atrous spatial pyramid pooling module of the trained first target recognition model to obtain a second feature extraction result;

[0046] The processing unit is used to process the second feature extraction result using the output layer to obtain a first recognition result.

[0047] The technical solution of the present invention has the following advantages:

[0048] The battery pack thermal runaway warning method provided by the embodiment of the present invention collects the pressure data, multiple state signal data, temperature data and multiple video data of the battery pack to be warned, extracts frame images from each video data to obtain a first frame image and a second frame image, uses the trained first target recognition model to perform image semantic recognition on the first frame image to obtain a first recognition result, uses the trained second target recognition model to perform character annotation processing on the second frame image to obtain a second recognition result, respectively judges the pressure data, multiple state signal data, temperature data, first recognition result and second recognition result to obtain multiple judgment results, and when any judgment result reaches the corresponding warning condition, a warning message is sent. The above method monitors the inside of the battery through the designed logical algorithm and monitors the outside of the battery pack using the trained target recognition model, and can accurately detect any abnormal signal and immediately trigger a warning prompt. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0050] Figure 1 It is a flowchart of the battery pack thermal runaway warning method in the embodiment of the present invention;

[0051] Figure 2 It is a training flowchart of the first target recognition model and the second target recognition model of the battery pack thermal runaway warning method in the embodiment of the present invention;

[0052] Figure 3 It is an effect diagram of sample image semantic segmentation processing of the battery pack thermal runaway warning method in the embodiment of the present invention;

[0053] Figure 4 It is a graph of the change in accuracy of the battery pack thermal runaway warning method in the embodiment of the present invention;

[0054] Figure 5 It is a graph of the change in loss of the battery pack thermal runaway warning method in the embodiment of the present invention;

[0055] Figure 6 It is a schematic diagram of the intersection over union related parameters of the battery pack thermal runaway warning method in the embodiment of the present invention;

[0056] Figure 7 It is a schematic diagram of the network architecture of the CRNN model of the battery pack thermal runaway warning method in the embodiment of the present invention;

[0057] Figure 8 The character annotation processing effect diagram of the battery pack thermal runaway warning method in the embodiment of the present invention;

[0058] Figure 9 The enlarged effect diagram of the character annotation of the battery pack thermal runaway warning method in the embodiment of the present invention;

[0059] Figure 10 The structural block diagram of the battery pack thermal runaway warning device in the embodiment of the present invention. Detailed implementation manners

[0060] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0061] In the description of the present invention, it should be noted that the terms "first", "second", and "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0062] The battery pack thermal runaway warning method provided by the embodiment of the present invention is as Figure 1 shown, Figure 1 The flowchart of the battery pack thermal runaway warning method, including steps S101 to S104, and the specific steps are as follows:

[0063] S101. Collect relevant data of the battery pack to be warned, where the relevant data includes pressure data, multiple status signal data, temperature data, and multiple video data.

[0064] In this embodiment, the warning end of the battery pack can be divided into two parts: logical algorithm and deep learning. The logical algorithm mainly includes a data collector (Gantner), a BMS, and a pressure sensor. Deep learning mainly includes a high-definition camera and a thermal imager; the control end only represents a programmable DC power source. For the logical algorithm part, the internal relevant data of the battery to be detected is collected through the pressure sensor, BMS, and data collector; in the deep learning part, the video data of the battery pack to be warned is collected by using a high-definition camera and a thermal imager.

[0065] S102. Extract frame images from each video data to obtain a first frame image and a second frame image. Use the trained first target recognition model to perform image semantic recognition on the first frame image to obtain a first recognition result. Use the trained second target recognition model to perform character annotation processing on the second frame image to obtain a second recognition result. Among them, the first target recognition model is obtained by training the DeepLab V3+ network using a first sample data set, and the second target recognition model is obtained by training the CRNN network using a second sample data set.

[0066] In this embodiment, after extracting the frame images from the video data, for the high-definition camera and the thermal imager, due to their different image feature processing methods, the high-definition camera mainly performs image semantic segmentation processing on the collected images, while the thermal imager mainly performs character annotation processing on the collected images. For the convolutional neural network models of both, a targeted selection is made. For the images collected by the high-definition camera, the DeepLab V3+ network is selected for deep learning and training to obtain the first target recognition model; for the thermal imager, the CRNN (Convolutional Recurrent Neural Network) model is selected for training to obtain the second target recognition model, as Figure 2 shown, Figure 2 is the model training process.

[0067] During the actual thermal runaway test, use a computer terminal to connect the high-definition camera and the thermal imager through TCP / IP communication or USB, and cyclically obtain video frames through the terminal. Then, adjust the frame images to the format required by the model through operations such as cropping and scaling. Next, input the processed pictures into the first target recognition model in real time for image semantic recognition. When it is detected that there is a fire and smoke, remind relevant personnel through a message pop-up window or a sound alarm, such as "Battery pack is smoking", "Battery pack is on fire"; at the same time, perform a threshold judgment on the temperature of the thermal imager recognized by the second target recognition model through an algorithm. If it exceeds the set threshold, remind relevant personnel through a message pop-up window or a sound alarm, such as the temperature anomaly warning information of "Thermal imager temperature is abnormal".

[0068] In one embodiment, using the trained first target recognition model to perform image semantic recognition on the first frame image to obtain a first recognition result includes:

[0069] Input the first frame image into the feature extraction network layer of the first target recognition model for feature extraction to obtain a first feature extraction result;

[0070] Use the atrous spatial pyramid pooling module of the trained first target recognition model to perform multi-scale feature extraction and fusion on the first feature extraction result to obtain a second feature extraction result;

[0071] The output layer is used to process the second feature extraction result to obtain the first recognition result.

[0072] In this embodiment, the first target recognition model is trained through the DeepLab V3+ network. This model combines the lightweight network Xception with depthwise separable convolutions as the backbone network for initial feature extraction, which helps to distinguish the background from targets such as smoke and flames. The core formula of depthwise separable convolution is:

[0073]

[0074]

[0075] SePConv(W p ,W d ,y) (i,j) PointwiseConv(W,y) (i,j) (W p ,DepthwiseConv (i,j) (V

[0076] In the formula, K and L are the sizes of the convolution kernel in the horizontal and vertical directions respectively, y is the input feature map, (i, j) are the positions of the pixel points of the input feature map respectively, and m represents the channel index of the input feature map.

[0077] At the same time, the DeepLab V3+ network uses the ASPP (Atrous Spatial Pyramid Pooling) atrous spatial pyramid pooling module to capture multi-scale context information, which helps to process smoke and flames of different scales. By inputting the frame image into the feature extraction network layer of the first target recognition model for feature extraction, and then using the atrous spatial pyramid pooling module to perform multi-scale feature extraction and fusion on the result of the feature extraction, and further using the output layer to process and output the fused result, the first recognition result is obtained.

[0078] In one embodiment, the first target recognition model is obtained by training the DeepLab V3+ network using the first sample data set, including:

[0079] Obtain the first sample data set, and divide the first sample data set into a first training set and a first test set according to a preset ratio;

[0080] Perform contrast processing and brightness processing on each sample image in the first training set to obtain the processed first training set. Add Gaussian noise to the processed first training set to obtain the enhanced first training set. Use the Laplace edge detection method to perform edge detection on the enhanced first training set to obtain the finally processed first training set;

[0081] Input the finally processed first training set into the DeepLab V3+ network for training. After obtaining the initial first target recognition model, use the sample images in the first test set to adjust the initial first target recognition model to obtain the trained first target recognition model.

[0082] In this embodiment, select the previously recorded videos of battery pack thermal runaway for frame-by-frame processing. To make the model training effect closer to human eye recognition, for the selected videos, for the resolution, battery pack type, recording angle, and the phenomenon of battery pack smoking and catching fire, as Figure 3 shown, perform semantic segmentation processing on a certain frame image intercepted from the video, and perform the same processing on other images. Considering that the camera resolution is low and the visibility in the test environment is poor, etc., it is necessary to perform contrast and brightness processing on the images before training; add Gaussian noise to enhance the model's robustness to noise; use the Laplace edge detection technology to highlight the edges of the smoke and fire areas, thereby improving the accuracy and robustness of the model detection. The specific training process is as follows: Assume that there are a total of 1000 photos for image semantic segmentation processing. Use 80% of the images as the first training set to train Deeplab V3+, and 20% of the images are used for testing. Considering that the proportion of the fire and smoke in the whole image is small, as can be seen from the above figure, the main body of the image is basically the test environment, followed by the battery pack. Therefore, class weights are used during training to improve this balance, that is, the test environment with the largest proportion has the lowest weight.

[0083] It should be noted that when dividing the first sample data set into the first training set and the first test set according to the preset proportion, the proportion can be selected according to actual needs. The proportion of the training set and the test set is not limited to 8:2.

[0084] When using 60% of the images to train the Deeplab V3+ network, the learning depth can be deepened by adding network layers. The Deeplab V3+ network is generally composed of multiple network layers. Each network layer is mainly composed of 8 learning layers: two-dimensional convolutional layer (ConvolutionalLayer), batch normalization layer (Batch Normalization), clipped rectified linear unit layer (ClippedReluLayer), two-dimensional grouped convolutional layer (GroupedConvolutional Layer, depthwise separable convolution), batch normalization layer, clipped rectified linear unit layer, two-dimensional convolutional layer, and batch normalization layer.

[0085] Meanwhile, the settings of some parameters during network training also have a great impact on the training results. For example, the type of solver, generally there are three types: sgdm, adam, and rmsprop; InitialLearnRate is the initial learning rate for training the network. If the initial learning rate is too high, the training may fall into a sub-optimal result or diverge; MaxEpochs is the maximum number of epochs used for training. An epoch is a complete pass of the training algorithm over the entire training set; MiniBatchSize is the mini-batch size used for each training iteration. An iteration is a step taken in the gradient descent algorithm to minimize the loss function using mini-batches. From Figure 4 and Figure 5 it can be seen that within a certain range, as the number of training rounds and the number of iterations per round increase, the accuracy rate gradually improves and the loss amount gradually decreases. After obtaining the initially trained first target recognition model, the remaining 20% test set is substituted into the trained initial first target recognition model for performance verification.

[0086] In one embodiment, the initial first target recognition model is adjusted using the sample images in the first test set to obtain a trained first target recognition model, including:

[0087] Input the sample images of the first test set into the initial first target recognition model for image recognition to obtain the first test recognition result. According to the first test recognition result and the first true result, obtain the current intersection over union ratio. The calculation formula for the current intersection over union ratio is:

[0088]

[0089] In the formula, TP is the true positive sample in the first test recognition result, FP is the false positive sample in the first test recognition result, and FN is the false negative sample in the first test recognition result;

[0090] After the current intersection over union ratio meets the first preset condition, reduce the number of network layers in the initial first target recognition model by one layer to obtain the trained first target recognition model. If the current intersection over union ratio does not meet the preset condition, continue to obtain the first sample data set to train the DeepLab V3+ network until the trained first target recognition model is obtained. The first preset condition is that the current intersection over union ratio is less than the intersection over union ratio calculated during the previous training of the DeepLab V3+ network.

[0091] In this embodiment, after obtaining the trained initial first target recognition model, the remaining 20% test set is substituted into the trained initial first target recognition model for performance verification. The verification metric can be measured by the intersection over union (IoU). The IoU is often used to measure the accuracy of the location information of the prediction results in the target detection task. The schematic diagram of the relevant parameters of the IoU is shown in Figure 6. The prediction result refers to the first test recognition result, and the actual image refers to the first ground truth result. The calculation method of the IoU is the size of the intersection of the two sets of the ground truth and the prediction divided by the size of the union. The calculation formula is as follows:

[0092]

[0093] In the formula, TP is the true positive sample in the first test recognition result, FP is the false positive sample in the first test recognition result, and FN is the false negative sample in the first test recognition result.

[0094] It should be noted that the first test recognition result is obtained by inputting the sample images of the first test set into the initial first target recognition model for image recognition.

[0095] In the initial stage of training the model, gradually increase the number of network layers, and at the same time observe the verification accuracy at the end of model training. When the number of network layers increases to a certain number, the verification accuracy will show a downward trend. This trend is overfitting (the prerequisite is that other training parameters are the same). When the verification accuracy of the model shows a downward trend, the model with the number of network layers reduced by one is the optimal one.

[0096] In one embodiment, the second target recognition model is used to perform character annotation processing on the second frame of image to obtain a second recognition result, including:

[0097] Use the convolutional layer of the second target recognition model to extract features from the second frame of image to obtain a thermal imaging extraction result;

[0098] Use the recurrent layer of the second target recognition model to process the extraction result of the second frame of image to obtain serialized features, where the recurrent layer is constructed based on the LSTM network;

[0099] Use the transcription layer of the second target recognition model to perform transformation processing on the serialized features to obtain a second recognition result, where the transcription layer is constructed by connecting the connectionist temporal classification algorithm.

[0100] In this embodiment, for the thermal imager, the CRNN (Convolutional Recurrent Neural Network) model is selected for training to obtain the second target recognition model. The network architecture of the CRNN model is as Figure 7As shown in the figure, the CRNN model mainly consists of three parts: the Convolutional Layer, the Recurrent Layer, and the Transcription Layer. The Convolutional Layer is responsible for extracting the features of the input image, and the Recurrent Layer serializes these features to process time series data. Finally, the serialized features are converted into corresponding characters or words through the Transcription Layer. Among them, the transcription is usually completed by CTC (Connectionist Temporal Classification). CTC uses the principle of probability theory to solve the sequence learning problem of unaligned data, and its goal is to maximize the conditional probability. The objective function of CTC is:

[0101]

[0102] In the formula, p(l|x) is the probability of the target label sequence, and P(π|x) is the probability of the set of all paths that can map to the target label sequence.

[0103] In one embodiment, the second target recognition model is obtained by training the CRNN network using the second sample dataset, including:

[0104] Obtain the second sample dataset, extract the target regions of each sample image in the second sample dataset to obtain multiple target images, and perform grayscale processing, normalization processing, and denoising processing on each target image to obtain the processed second sample dataset;

[0105] Divide the processed second sample dataset into a second training set and a second test set according to a preset ratio. Input the sample data in the second training set into the CRNN network for training to obtain an initial second target recognition model, and use the sample data in the second test set to adjust the initial second target recognition model to obtain a trained second target recognition model.

[0106] In this embodiment, select the previously recorded thermal imager video for frame-by-frame processing. To make the model training effect closer to human eye recognition, the selected video should have a wider range of resolution, recording angle, ambient brightness, and thermal imager display temperature range. For example, Figure 8 As shown in the figure, character annotation processing is performed on a certain frame image intercepted from the video, and the same processing is performed on other images. A detailed enlarged view of the character annotation of the temperature value is shown in Figure 9. Different temperature values can be selected for the above character annotation. For example, label the number 3 as "char_3" and the decimal point as "char_period" (the annotation naming rule does not allow direct use of numbers).

[0107] Considering that the resolution of the thermal imager is relatively low and the visibility of the test environment is poor, etc., it is necessary to perform feature preprocessing on the images, such as extracting the region of interest (ROI) containing temperature values, removing irrelevant background information; performing grayscale conversion, normalization, and denoising, etc.

[0108] Suppose there are a total of 1000 photos for image character annotation processing. 80% of the images are used to train the CRNN, and 20% of the images are used for testing. Given that the proportion of temperature display in the whole image is small and fixed, class weights are used during training to improve this balance, that is, the test environment with the largest proportion has the lowest weight.

[0109] The training of the CRNN model improves the training results by improving the network architecture. For example, at the convolutional layer end, the number of convolutional layers is increased to improve the model's ability to extract spatial features, or like the DeepLab V3+ network, adding depthwise separable convolutional layers or normalization layers accelerates the training process and improves the model's performance; for the recurrent layer, more complex recurrent units such as long short-term memory units (LSTM) are used. The LSTM contains a memory unit and three multiplicative gates, namely the input, output, and forget gates. The special design of the LSTM can capture the long-term dependencies that often occur in picture-based sequences; for the transcription layer, by adding a conditional random field (CRF) layer, the dependencies between labels are captured to improve the accuracy of sequence labels. Especially for the continuous temperature value recognition task, the CRF layer can effectively improve the model's performance. In addition, for the addition of the above learning layers and units, the quantity needs to be controlled according to the trend of the validation accuracy to prevent overfitting.

[0110] The setting of the training parameters for the CRNN network model is the same as that in training the DeepLab V3+ network above. For example, the type of solver, generally there are three types: sgdm, adam, and rmsprop; InitialLearnRate is the initial learning rate for training the network. If the initial learning rate is too high, the training may fall into suboptimal results or diverge; MaxEpochs is the maximum number of epochs for training. An epoch is a complete pass of the training algorithm over the entire training set; MiniBatchSize is the mini-batch size for each training iteration. An iteration is a step taken in the gradient descent algorithm to minimize the loss function using mini-batches. After obtaining the initial second target recognition model through the above training, then the remaining 20% test set is substituted into the initial second target recognition model for performance verification.

[0111] In one embodiment, the initial second target recognition model is adjusted using the sample data in the second test set to obtain a trained second target recognition model, including:

[0112] Input the sample data in the second test set into the initial second target recognition model for image recognition to obtain the second test recognition result. According to the second test recognition result and the second ground truth result, calculate the current character error rate. The calculation formula for the current character error rate is as follows:

[0113]

[0114] In the formula, N is the length of the original string in the ground truth result, S is the number of replaced characters, D is the number of deleted characters, and I is the number of extra inserted characters;

[0115] After the current character error rate meets the second preset condition, reduce the number of network layers in the initial second target recognition model by one layer to obtain the trained second target recognition model, and determine the initial second target recognition model as the trained second target recognition model. If the character error rate does not meet the preset condition, continue to obtain the second sample data set to train the CRNN network until the trained second target recognition model is obtained. The second preset condition is that the current character error rate is less than the character error rate calculated during the previous training of the CRNN network.

[0116] In this embodiment, after obtaining the initial second target recognition model through the above training, then substitute the remaining 20% test set into the initial second target recognition model for performance verification. The verification index can be measured by the character error rate (CER). In the character recognition task, CER is used to measure the edit distance between the recognized characters and the correct answers. The edit distance refers to the minimum number of edit operations required to change one string into another string. The edit operations include insertion, deletion, and replacement, thereby reflecting the error rate of the model at the character level. The CER calculation formula is as follows:

[0117]

[0118] In the formula, N is the length of the original string in the ground truth result, S is the number of replaced characters, D is the number of deleted characters, and I is the number of extra inserted characters.

[0119] When training the CRNN network, for the addition of learning layers and units, the quantity needs to be controlled according to the trend of the verification accuracy to prevent overfitting. That is, when the number of network layers increases to a certain number, the verification accuracy will show a downward trend. This trend is overfitting (the prerequisite is that other training parameters are the same). When the model verification accuracy shows a downward trend, the model with one less network layer is the optimal one.

[0120] S103. Judge the pressure data, multiple status signal data, temperature data, the first recognition result, and the second recognition result respectively to obtain multiple judgment results. When any judgment result reaches the corresponding warning condition, a warning message is sent.

[0121] In this embodiment, connect and communicate according to the device address, baud rate, stop bit, etc. specified by the Modbus protocol of the pressure sensor. Read the pressure value of the battery pack in real time through the register address and register type where the pressure signal is located. Edit the corresponding algorithm according to the pressure threshold for opening the valve of the battery pack, and update the pressure value of the pressure sensor in real time during the test. If the specified pressure threshold is exceeded, relevant personnel are reminded by means of a message pop-up window or a sound alarm, such as "The internal pressure of the battery pack is abnormal".

[0122] Establish a communication connection through the specified CAN protocol type and baud rate of the battery pack, load the corresponding DBC file to parse the signals, edit the corresponding logic algorithm according to the BMS signal threshold, and update it in real time during the test. If the threshold provided by the customer is exceeded, relevant personnel are reminded by means of a message pop-up window or a sound alarm, such as "Abnormal temperature of cell 1", "Insulation failure". Among them, the BMS signals include but are not limited to single voltage, total voltage, cell temperature, insulation detection status, relay status, etc.

[0123] Establish a communication connection through the specified CAN protocol type and baud rate of Gantner, load the corresponding DBC file to parse the signals, and perform a logic algorithm on the collected voltage and temperature. For example, as specified in GB / T 36276-2023 6.7.4.2, "Set the determination condition for thermal runaway to occur when three consecutive temperature rise rate values are all ≥ 3°C / s or there is a fire or explosion". Corresponding algorithms can be set for this. When the condition is met, relevant personnel are reminded by means of a message pop-up window or a sound alarm, such as "The temperature rise rate ≥ 3°C / s for 3 consecutive seconds".

[0124] After the convolutional neural network model is trained, for the actual thermal runaway test, only a computer terminal is needed to connect the high-definition camera and the thermal imager through TCP / IP communication or USB, and the video frames are obtained cyclically through the terminal. Then, the frame pictures are adjusted to the format required by the model through operations such as cropping and scaling, and the processed pictures are input into the first target recognition model in real time for image semantic recognition. When fire and smoke are detected, relevant personnel are reminded by means of a message pop-up window or a sound alarm, such as "The battery pack is smoking", "The battery pack is on fire"; at the same time, the second recognition result recognized by the second target recognition model, that is, the temperature of the thermal imager, is judged by an algorithm according to the threshold. If the set threshold is exceeded, relevant personnel are reminded by means of a message pop-up window or a sound alarm, such as "Abnormal temperature of the thermal imager". Relevant personnel assist in the judgment through the obtained warning information and send an instruction to the DC power supply to stop heating and terminate the test according to the specified thermal runaway trigger conditions.

[0125] The battery pack thermal runaway warning device provided by the embodiment of the present invention, as Figure 10 shown, Figure 10 is a device block diagram of the battery pack thermal runaway warning device 1000, including an acquisition module 1001, an identification module 1002, and a warning module 1003. Among them,

[0126] The acquisition module 1001 is used to acquire relevant data of the battery pack to be warned. Among them, the relevant data includes pressure data, multiple status signal data, temperature data, and multiple video data;

[0127] The identification module 1002 is used to extract frame images from each video data to obtain a first frame image and a second frame image, perform image semantic recognition on the first frame image using a trained first target recognition model to obtain a first recognition result, and perform character annotation processing on the second frame image using a trained second target recognition model to obtain a second recognition result. Among them, the first target recognition model is obtained by training the DeepLab V3+ network using a first sample data set, and the second target recognition model is obtained by training the CRNN network using a second sample data set;

[0128] The warning module 1003 is used to respectively judge the pressure data, multiple status signal data, temperature data, the first recognition result, and the second recognition result to obtain multiple judgment results. When any judgment result reaches the corresponding warning condition, a warning message is sent.

[0129] In one embodiment, the identification module includes a first extraction unit, a second extraction unit, and a processing unit. Among them,

[0130] The first extraction unit is used to input the first frame image into the feature extraction network layer of the first target recognition model for feature extraction to obtain a first feature extraction result;

[0131] The second extraction unit is used to perform multi-scale feature extraction and fusion on the first feature extraction result using the atrous spatial pyramid pooling module of the trained first target recognition model to obtain a second feature extraction result;

[0132] The processing unit is used to process the second feature extraction result using the output layer to obtain a first recognition result.

[0133] The specific implementation manner of this battery pack thermal runaway warning device is basically the same as the specific embodiment of the above battery pack thermal runaway warning method, and will not be elaborated here.

[0134] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0135] The specific embodiments described above have further elaborated on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. In particular, it is pointed out that for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A battery pack thermal runaway early warning method, characterized in that: include: Collect relevant data of the battery pack to be warned, wherein the relevant data includes pressure data, multiple state signal data, temperature data and multiple video data; Extracting frame images from each of the video data to obtain a first frame image and a second frame image, performing image semantic recognition on the first frame image using a trained first target recognition model to obtain a first recognition result, and performing character annotation processing on the second frame image using a trained second target recognition model to obtain a second recognition result, wherein the first target recognition model is obtained by training a DeepLab V3+ network using a first sample data set, and the second target recognition model is obtained by training a CRNN network using a second sample data set; The pressure data, the multiple state signal data, the temperature data, the first recognition result and the second recognition result are judged respectively to obtain multiple judgment results. When any judgment result reaches the corresponding warning condition, a warning message is issued.

2. The battery pack thermal runaway warning method according to claim 1, characterized in that: The using the trained first target recognition model to perform image semantic recognition on the first frame of image to obtain a first recognition result includes: Inputting the first frame image into the feature extraction network layer of the first target recognition model to perform feature extraction, thereby obtaining a first feature extraction result; Using the atrous space pyramid pooling module of the trained first target recognition model to perform multi-scale feature extraction and fusion on the first feature extraction result to obtain a second feature extraction result; The second feature extraction result is processed using the output layer to obtain a first recognition result.

3. The battery pack thermal runaway warning method according to claim 1, characterized in that: The first target recognition model is obtained by training the DeepLab V3+ network using the first sample data set, including: Obtain a first sample data set, and divide the first sample data set into a first training set and a first test set according to a preset ratio; Performing contrast processing and brightness processing on each sample image in the first training set to obtain a processed first training set, adding Gaussian noise to the processed first training set to obtain an enhanced first training set, and performing edge detection on the enhanced first training set using the Laplace edge detection method to obtain a final processed first training set; The finally processed first training set is input into the DeepLab V3+ network for training to obtain an initial first target recognition model, and then the initial first target recognition model is adjusted using sample images in the first test set to obtain a trained first target recognition model.

4. The battery pack thermal runaway warning method according to claim 3, characterized in that: The step of adjusting the initial first object recognition model by using the sample images in the first test set to obtain the trained first object recognition model includes: The sample image of the first test set is input into the initial first target recognition model for image recognition to obtain a first test recognition result, and a current intersection-and-union ratio is obtained according to the first test recognition result and the first true result, wherein the calculation formula of the current intersection-and-union ratio is: Wherein, TP is a true positive sample in the first test recognition result, FP is a false positive sample in the first test recognition result, and FN is a false negative sample in the first test recognition result; When the current intersection-and-union ratio meets the first preset condition, the number of network layers in the initial first target recognition model is reduced by one layer to obtain the trained first target recognition model. If the current intersection-and-union ratio does not meet the preset condition, the first sample data set is continued to be obtained to train the DeepLab V3+ network until the trained first target recognition model is obtained, wherein the first preset condition is that the current intersection-and-union ratio is less than the intersection-and-union ratio calculated when the DeepLab V3+ network was trained last time.

5. The battery pack thermal runaway warning method according to claim 1, characterized in that: The using the second object recognition model to perform character labeling processing on the second frame image to obtain a second recognition result includes: Using the convolutional layer of the second target recognition model to perform feature extraction on the second frame image to obtain a thermal imaging extraction result; Processing the second frame image extraction result by using the recursive layer of the second target recognition model to obtain serialized features, wherein the recursive layer is constructed based on the LSTM network; The serialized features are transformed and processed by using the transcription layer of the second target recognition model to obtain a second recognition result, wherein the transcription layer is constructed by a connection time series classification algorithm.

6. The battery pack thermal runaway warning method according to claim 1, characterized in that: The second target recognition model is obtained by training the CRNN network using the second sample data set, including: Acquire a second sample data set, extract a target region from each sample image in the second sample data set to obtain a plurality of target images, perform grayscale processing, normalization processing and denoising processing on each of the target images to obtain a processed second sample data set; The processed second sample data set is divided into a second training set and a second test set according to a preset proportion, the sample data in the second training set is input into the CRNN network for training to obtain an initial second target recognition model, and the initial second target recognition model is adjusted using the sample data in the second test set to obtain a trained second target recognition model.

7. The battery pack thermal runaway warning method according to claim 6, characterized in that: The step of adjusting the initial second object recognition model by using the sample data in the second test set to obtain the trained second object recognition model includes: The sample data in the second test set is input into the initial second target recognition model for image recognition to obtain a second test recognition result. According to the second test recognition result and the second true result, the current character error rate is obtained, wherein the calculation formula of the current character error rate is: Where N is the length of the original string in the actual result, S is the number of replaced characters, D is the number of deleted characters, and I is the number of additional inserted characters; When the current character error rate meets the second preset condition, the number of network layers in the initial second target recognition model is reduced by one layer to obtain the trained second target recognition model, and the initial second target recognition model is determined to be the trained second target recognition model. If the character error rate does not meet the preset condition, continue to obtain the second sample data set to train the CRNN network until the trained second target recognition model is obtained, wherein the second preset condition is that the current character error rate is less than the character error rate calculated when the CRNN network was trained last time.

8. The battery pack thermal runaway warning method according to claim 1, characterized in that: The pressure data, the plurality of state signal data, the temperature data, the first recognition result and the second recognition result are judged respectively to obtain a plurality of judgment results, and when any judgment result reaches a corresponding warning condition, a warning message is issued, including: Determine whether the pressure data is greater than the pressure threshold, and if so, issue a warning message of abnormal pressure; Determine whether each of the state signal data is greater than the corresponding state signal threshold, and if so, issue an early warning message indicating that the device state is abnormal; Calculate the temperature rise rate of the battery pack to be warned within a preset time period according to the temperature data, and determine whether the temperature rise rate is greater than a preset rate threshold. If so, issue a warning message of abnormal internal temperature. Based on the first recognition result, it is determined whether the battery pack to be warned has caught fire. If so, a warning message of battery pack fire is issued. Based on the second recognition result, it is determined whether the surface temperature of the battery pack to be warned is greater than a preset temperature threshold. If so, a warning message of external temperature abnormality is issued.

9. A battery pack thermal runaway warning device, characterized in that: It includes a collection module, an identification module, and an early warning module, among which: The acquisition module is used to acquire relevant data of the battery pack to be warned, wherein the relevant data includes pressure data, multiple state signal data, temperature data and multiple video data; The recognition module is used to extract frame images from each of the video data to obtain a first frame image and a second frame image, perform image semantic recognition on the first frame image using a trained first target recognition model to obtain a first recognition result, and perform character annotation processing on the second frame image using a trained second target recognition model to obtain a second recognition result, wherein the first target recognition model is obtained by training the DeepLab V3+ network using the first sample data set, and the second target recognition model is obtained by training the CRNN network using the second sample data set; The early warning module is used to judge the pressure data, the multiple status signal data, the temperature data, the first recognition result and the second recognition result respectively to obtain multiple judgment results. When any judgment result reaches the corresponding early warning condition, an early warning message is issued.

10. The battery pack thermal runaway warning method according to claim 9, characterized in that: The recognition module includes a first extraction unit, a second extraction unit and a processing unit, wherein: The first extraction unit is used to input the first frame image into the feature extraction network layer of the first target recognition model to perform feature extraction, so as to obtain a first feature extraction result; The second extraction unit is used to perform multi-scale feature extraction and fusion on the first feature extraction result by using the hole space pyramid pooling module of the trained first target recognition model to obtain a second feature extraction result; The processing unit is used to process the second feature extraction result using the output layer to obtain a first recognition result.

Images