A method and apparatus for predicting the evolution of defect events in lithium-ion batteries

Abstract

This invention relates to a method and apparatus for predicting the evolution of defect events in lithium-ion batteries, comprising: acquiring lithium-ion battery defect data and preprocessing the lithium-ion battery defect data to obtain preprocessed data; using a BERT-BiLSTM-CRF model to extract defect events from the preprocessed data to obtain defect event extraction results; based on the defect event extraction results, using a defect event relationship extraction model to extract the relationships between various defect events to obtain defect event relationship extraction results; wherein, the defect event relationship extraction model is obtained by improving a cue learning framework; based on the defect event extraction results and the defect event relationship extraction results, constructing a lithium-ion battery defect event graph, and performing evolution prediction based on the lithium-ion battery defect event graph, thereby preventing the occurrence of defect events.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery technology, and in particular to a method and apparatus for predicting the evolution of defect events in lithium-ion batteries. Background Technology

[0002] Lithium-ion batteries are characterized by being green and low-carbon, and having a long cycle life. However, defects and accidents can occur in lithium-ion batteries during production, transportation, and use, including separator perforation, loss of active lithium, and electrode powder shedding, which can lead to battery performance degradation and even thermal runaway causing safety accidents.

[0003] Therefore, in order to prevent defect accidents, there is an urgent need for a method to predict the evolution of defect events in lithium-ion batteries. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method and apparatus for predicting the evolution of defect events in lithium-ion batteries, which can prevent the occurrence of defect events.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0008] In a first aspect, embodiments of the present invention provide a method for predicting the evolution of defect events in lithium-ion batteries, comprising: acquiring lithium-ion battery defect data and preprocessing the lithium-ion battery defect data to obtain preprocessed data; using a BERT-BiLSTM-CRF model to extract defect events from the preprocessed data to obtain defect event extraction results; wherein the BERT-BiLSTM-CRF model includes a BERT feature extraction layer, a multi-granularity feature extraction layer, an attention mechanism layer, a dynamic weight fusion layer, a sequence modeling layer, and a classification decoding layer, the BERT feature extraction layer being connected to the multi-granularity feature extraction layer, ... The attention mechanism layer and the dynamic weight fusion layer are connected, and the dynamic weight fusion layer is connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the sequence modeling layer, respectively. The sequence modeling layer is also connected to the classification and decoding layer. Based on the defect event extraction results, the relationship between each defect event is extracted using a defect event relationship extraction model, resulting in defect event relationship extraction results. The defect event relationship extraction model is an improvement on the cue learning framework. Based on the defect event extraction results and the defect event relationship extraction results, a lithium-ion battery defect event graph is constructed, and evolution prediction is performed based on the lithium-ion battery defect event graph.

[0009] In one possible embodiment, the multi-granularity feature extraction layer includes a multi-scale convolution module and an output concatenation module. The multi-scale convolution module is used to scan the word vector sequence output by the BERT feature extraction layer in parallel using convolution kernels of different widths to obtain multiple convolution results. The output concatenation module is used to concatenate the multiple convolution results along the feature dimension to obtain multi-scale fused features.

[0010] In one possible embodiment, the attention mechanism layer includes a word-level attention module and a sentence-level attention module. The word-level attention module is used to calculate the association weights between words in the lithium-ion battery defect data, and the sentence-level attention module is used to perform a weighted average of all word vectors in the entire sentence in the lithium-ion battery defect data to generate a global sentence vector.

[0011] In one possible embodiment, the sequence modeling layer comprises three sequentially connected bidirectional LSTM modules.

[0012] In one possible embodiment, the defect event relation extraction model includes an input layer, a pre-trained language model layer, a parallel answer prediction layer, and a fusion discriminant layer connected in sequence. The input layer is used to construct a prompt template based on the defect event extraction results, comprising two defect events and two relation prediction mask bits for the relation to be determined. The pre-trained language model layer is used to parse the contextual semantics of the prompt template and extract the relation representation vectors for the mask positions. The parallel answer prediction layer is used to independently calculate the candidate relation probability distribution for each mask position. The fusion discriminant layer is used to independently select the optimal relation label for each mask position.

[0013] Secondly, embodiments of the present invention provide a lithium-ion battery defect event evolution prediction device, comprising:

[0014] The acquisition module is used to acquire lithium-ion battery defect data and preprocess the lithium-ion battery defect data to obtain preprocessed data.

[0015] The defect event extraction module is used to extract defect events from the preprocessed data using the BERT-BiLSTM-CRF model to obtain the defect event extraction results. The BERT-BiLSTM-CRF model includes a BERT feature extraction layer, a multi-granularity feature extraction layer, an attention mechanism layer, a dynamic weight fusion layer, a sequence modeling layer, and a classification decoding layer. The BERT feature extraction layer is connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the dynamic weight fusion layer, and the dynamic weight fusion layer is connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the sequence modeling layer, and the sequence modeling layer is also connected to the classification decoding layer.

[0016] The relation extraction module is used to extract the relationships between various defect events based on the defect event extraction results using the defect event relation extraction model, thereby obtaining the defect event relation extraction results; wherein, the defect event relation extraction model is obtained by improving the prompting learning framework.

[0017] The evolution prediction module is used to construct a lithium-ion battery defect event graph based on the defect event extraction results and defect event relationship extraction results, and to perform evolution prediction based on the lithium-ion battery defect event graph.

[0018] In one possible embodiment, the multi-granularity feature extraction layer includes a multi-scale convolution module and an output concatenation module. The multi-scale convolution module is used to scan the word vector sequence output by the BERT feature extraction layer in parallel using convolution kernels of different widths to obtain multiple convolution results. The output concatenation module is used to concatenate the multiple convolution results along the feature dimension to obtain multi-scale fused features.

[0019] In one possible embodiment, the attention mechanism layer includes a word-level attention module and a sentence-level attention module. The word-level attention module is used to calculate the association weights between words in the lithium-ion battery defect data, and the sentence-level attention module is used to perform a weighted average of all word vectors in the entire sentence in the lithium-ion battery defect data to generate a global sentence vector.

[0020] In one possible embodiment, the sequence modeling layer comprises three sequentially connected bidirectional LSTM modules.

[0021] In one possible embodiment, the defect event relation extraction model includes an input layer, a pre-trained language model layer, a parallel answer prediction layer, and a fusion discriminant layer connected in sequence. The input layer is used to construct a prompt template based on the defect event extraction results, comprising two defect events and two relation prediction mask bits for the relation to be determined. The pre-trained language model layer is used to parse the contextual semantics of the prompt template and extract the relation representation vectors for the mask positions. The parallel answer prediction layer is used to independently calculate the candidate relation probability distribution for each mask position. The fusion discriminant layer is used to independently select the optimal relation label for each mask position.

[0022] Thirdly, embodiments of this application also provide an electronic device, the electronic device comprising: a memory for storing a computer program; and a processor for executing the computer program stored in the memory, wherein when the computer program is executed, the processor is used to execute a lithium-ion battery defect event evolution prediction method as described above.

[0023] Fourthly, embodiments of this application also provide a computer-readable storage medium including computer instructions that, when executed on an electronic device, cause the electronic device to perform a lithium-ion battery defect event evolution prediction method as described above.

[0024] (III) Beneficial Effects

[0025] The beneficial effects of this invention are:

[0026] This application provides a method and apparatus for predicting the evolution of lithium-ion battery defect events. It acquires lithium-ion battery defect data, preprocesses the data to obtain preprocessed data, uses a BERT-BiLSTM-CRF model to extract defect events from the preprocessed data, obtains defect event extraction results, uses a defect event relationship extraction model to extract the relationships between various defect events based on the defect event extraction results, obtains defect event relationship extraction results, and constructs a lithium-ion battery defect event graph based on the defect event extraction results and defect event relationship extraction results. Evolution prediction is then performed based on the lithium-ion battery defect event graph, thereby preventing the occurrence of defect events.

[0027] To make the above-mentioned objectives, features and advantages to be achieved by the embodiments of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A flowchart of a lithium-ion battery defect event evolution prediction method provided in an embodiment of this application is shown;

[0030] Figure 2 This paper shows a schematic diagram of the structure of a BERT-BiLSTM-CRF model provided in an embodiment of this application;

[0031] Figure 3 This illustration shows a structural schematic diagram of a defect event relationship extraction model provided in an embodiment of this application;

[0032] Figure 4 A structural block diagram of a lithium-ion battery defect event evolution prediction device provided in an embodiment of this application is shown. Detailed Implementation

[0033] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] To prevent the occurrence of defect accidents, embodiments of this application provide a method and apparatus for predicting the evolution of lithium-ion battery defect events. The method involves acquiring lithium-ion battery defect data, preprocessing the data to obtain preprocessed data, using a BERT-BiLSTM-CRF model to extract defect events from the preprocessed data, obtaining defect event extraction results, using a defect event relationship extraction model to extract the relationships between various defect events based on the defect event extraction results, obtaining defect event relationship extraction results, constructing a lithium-ion battery defect event graph based on the defect event extraction results and defect event relationship extraction results, and performing evolution prediction based on the lithium-ion battery defect event graph, thereby preventing the occurrence of defect events.

[0035] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0036] Please see Figure 1 , Figure 1 A flowchart of a lithium-ion battery defect event evolution prediction method provided in an embodiment of this application is shown. It should be understood that this lithium-ion battery defect event evolution prediction method can be executed by a lithium-ion battery defect event evolution prediction device, and the specific device can be configured according to actual needs; this embodiment is not limited thereto. For example, the lithium-ion battery defect event evolution prediction device can be a computer, or a server, etc. Specifically, the lithium-ion battery defect event evolution prediction method includes:

[0037] Step S110: Obtain lithium-ion battery defect data and preprocess the lithium-ion battery defect data to obtain preprocessed data.

[0038] Specifically, the first step is to obtain raw data on lithium-ion battery defects using web crawling technology, and then preprocess the raw data, including data cleaning, custom dictionary creation, and stop word removal. This may include deleting invalid, duplicate, and abnormal data from the text data; establishing structural nouns, defect names, and logical relationship names for the lithium-ion battery defect domain; and removing punctuation marks, spaces, and interjections from the text.

[0039] In addition, it can also annotate defect events for data processing. Defect event annotation includes event arguments, event trigger words, and event types. Among them, event arguments include the event subject and object; event trigger words are the verbs that cause the event; event types include three types: battery material level defects, battery pack level defects, and battery management system level defects. Defect event annotation uses the BIO annotation method, where B represents the start of the event, I represents the middle or end of the event, and O represents irrelevant words.

[0040] In addition, a BERT tokenizer (e.g., Tokenizer) can be used to segment the original defect description text (e.g., “negative electrode lithium plating induces thermal runaway”, etc.) to generate token sequences that conform to the pre-trained vocabulary (e.g., “activity”, “lithium”, “loss”, “cause”, “capacity”, “attenuation”, etc.), and generate two key masks: a padding mask and an attention mask.

[0041] Step S120: Use the BERT-BiLSTM-CRF model to extract defect events from the preprocessed data to obtain the defect event extraction results.

[0042] It should be understood that the specific structure of the BERT-BiLSTM-CRF model can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0043] Optionally, such as Figure 2 As shown, the BERT-BiLSTM-CRF model includes a BERT feature extraction layer, a multi-granularity feature extraction layer, an attention mechanism layer, a dynamic weight fusion layer, a sequence modeling layer, and a classification decoding layer. The BERT feature extraction layer is connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the dynamic weight fusion layer, respectively. The dynamic weight fusion layer is connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the sequence modeling layer, respectively. The sequence modeling layer is also connected to the classification decoding layer.

[0044] The BERT feature extraction layer is used to input the token sequence into a 12-layer Transformer architecture pre-trained BERT model. Each layer captures global inter-word dependencies (such as the long-distance association between "lithium plating" and "thermal runaway") through a multi-head self-attention mechanism, and extracts non-linear features through a feedforward network. Finally, it can output a 768-dimensional context vector for each token.

[0045] The multi-granularity feature extraction layer includes a multi-scale convolution module and an output concatenation module. The multi-scale convolution module is used to scan the word vector sequence output by the BERT feature extraction layer in parallel using convolution kernels of different widths to obtain multiple convolution results. The output concatenation module is used to concatenate the multiple convolution results along the feature dimension to obtain multi-scale fused features.

[0046] For example, the multi-scale convolution module is used to parallelly scan the word vector sequence output by the BERT feature extraction layer using three convolution kernels with widths of 3, 5, and 7, obtaining multiple convolution results. Among them, these three convolution kernels with different widths correspond to different windows. The convolution kernel with a width of 3 corresponds to a small window used to capture adjacent word combinations (such as "analysis" + "lithium", etc.), the convolution kernel with a width of 5 corresponds to a medium window used to identify phrase-level structures (such as "thermal runaway"), and the convolution kernel with a width of 7 corresponds to a large window used to capture cross-word length patterns (such as "negative active material shedding", etc.);

[0047] The attention mechanism layer includes a word-level attention module and a sentence-level attention module. The word-level attention module is used to calculate the association weights between words in the lithium-ion battery defect data, and the word-level attention module learns the semantic associations in different subspaces in parallel through a multi-head mechanism; the sentence-level attention module is used to perform weighted averaging on all word vectors of the entire sentence in the lithium-ion battery defect data to generate a global sentence vector. And the sentence-level attention module can aggregate the information of each token in the sequence through weighted averaging to capture the overall semantics of the sentence.

[0048] Among them, the word-level attention module used to capture the semantic dependence between words can dynamically learn the correlation between words using the Query-Key-Value mechanism and reconstruct the semantic representation of each word through weight assignment; the attention module used to generate the concentrated representation of the overall semantics of the sentence can dynamically calculate the correlation weights between elements through the Query-Key-Value mechanism and perform weighted aggregation of the Value information to achieve semantic enhancement, that is, use the vector dot product similarity to generate attention weights and reconstruct the key features through weighted fusion;

[0049] The dynamic weight fusion layer is used to adaptively fuse the outputs of the BERT feature extraction layer, the multi-granularity feature extraction layer, and the attention mechanism layer through a learnable weight vector, so that the model can autonomously learn the importance of different features and avoid the limitations of artificially preset weights.

[0050] For example, the dynamic weight fusion layer stacks the above three outputs in alignment with Tokens to form a three-dimensional tensor. And a trainable weight matrix is introduced, and adaptive weights for the three types of features are generated through Softmax (such as emphasizing multi-granularity features for "dendrite growth" and emphasizing word attention features for "causing"), and the weighted sum is used to generate the fused features;

[0051] This sequence modeling layer uses a three-layer bidirectional LSTM connected sequentially to process and fuse features, obtaining a final vector for each word. The first layer of bidirectional LSTM learns word-level patterns (e.g., "expansion" is often the starting word of a phenomenon), the second layer constructs phrase structures (e.g., "internal short circuit" is a complete phenomenon entity), and the third layer captures long-distance dependencies (e.g., the cross-sentence association between "overcharging" and "fire").

[0052] This classification decoding layer may include a classification module and a CRF (Conditional Random Field) module. The classification module is used to convert the sequence features output by the sequence modeling layer into label scores, that is, to map each word vector to the label space. The CRF module is used to decode the optimal sequence that meets the semantic constraints by utilizing the label transition probabilities, that is, to output the globally optimal label sequence by learning the transition rules between labels. For example, activity (B-defect cause) lithium (I-defect cause) loss (I-defect cause) leading to (O) capacity (B-defect phenomenon) decay (I-defect phenomenon).

[0053] In other words, the classification module independently predicts the label scores of each word, and the CRF module decodes the globally optimal sequence based on the label transition probability constraints. The label score defines the original prediction confidence of a single word, and the label transition probability defines the order rule between labels (e.g., B-cause must be followed by I-cause). The CRF module solves for the globally optimal path by adding the two together (label score + transition probability).

[0054] Step S130: Based on the defect event extraction results, the relationship between each defect event is extracted using the defect event relationship extraction model to obtain the defect event relationship extraction results. The defect event relationship extraction model is an improvement upon the cue learning framework.

[0055] It should be understood that the specific structure of the defect event relationship extraction model can also be set according to actual needs, and the embodiments of this application are not limited thereto.

[0056] Optionally, such as Figure 3 As shown, the defect event relation extraction model comprises an input layer, a pre-trained language model layer, a parallel answer prediction layer, and a fusion discriminant layer connected in sequence. The input layer constructs a prompt template based on the defect event extraction results, including two defect events and two relation prediction mask positions for the relation to be determined. The pre-trained language model layer parses the contextual semantics of the prompt template and extracts the relation representation vectors for the mask positions. The parallel answer prediction layer independently calculates the candidate relation probability distribution for each mask position, and the fusion discriminant layer independently selects the optimal relation label for each mask position.

[0057] It should be understood that the specific structure of the prompt template can be set according to actual needs, as long as it includes two defect events of the relationship to be determined, the relationship prediction bit MASK, and soft and hard prompt words. The embodiments of this application are not limited to this.

[0058] For example, the structure of this prompt template can be: e 1+[ ST 1]+ leads to+[ MASK 1]+[ ST 2] + premise + [ MASK 2]+[ ST 3]+ e 2; among which, e 1 and e 2 represents two defective events whose relationship needs to be determined; ST 1. ST 2 and ST 3 represents the first vector placeholder, the second vector placeholder, and the third vector placeholder, respectively, which are all software prompt words that can be trained vectors; "cause" and "premise" can be hard prompt words to enhance professional semantic guidance capabilities; MASK 1 and MASK 2 represents the first relation prediction bit and the second relation prediction bit, respectively, which can be used to predict complex relation categories such as causal relationships and triggering conditions between events.

[0059] It should also be understood that the specific structure of the pre-trained language model layer can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0060] For example, this pre-trained language model layer includes sequentially connected embedding modules and multiple... Transformer The module processes the input sequence layer by layer and outputs a context-aware semantic representation. Multiple of these modules... Transformer Each in the module Transformer Each module includes a multi-head self-attention sub-layer, a feedforward neural network sub-layer, residual connections, and layer normalization.

[0061] It should also be understood that the specific structure of the parallel answer prediction layer can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0062] For example, this parallel answer prediction layer includes processing MASK 1. The main relation channel and the channel used for processing MASK 2. Auxiliary relation channels. Both the primary relation channel and the auxiliary relation channel are used to calculate the representation of the corresponding... MASK A similarity score is given to the degree of matching between the semantic vector of a location and the candidate relation word vector, and this score is used to represent... MASK The score vector of the numerical set of positional semantic matching with all candidate relations is used for... Softmax Normalization.

[0063] It should also be understood that the specific structure of the fusion discrimination layer can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0064] For example, the fusion discriminant layer may also include two relation channels and an integrated output layer connected to the two relation channels. Each relation channel is used to normalize each score vector and select the maximum score item. The integrated output layer is used to integrate the data representing the main causal relationship and possible conditional dependency relationship between the events from the outputs of the two channels to obtain the relation label pair of the two defect events.

[0065] Step S140: Based on the defect event extraction results and defect event relationship extraction results, construct a lithium-ion battery defect event graph and perform evolution prediction based on the lithium-ion battery defect event graph.

[0066] Specifically, based on information extraction, a JSON file can be constructed to import this information into a graph database. After constructing the JSON file, the Pandas library is used to read the data by passing in the filename. Finally, in the Py2neo library, a Graph object is used to connect to Neo4j, and the Create method is used to add Nodes to the database. The Neo4j database provides a visualization interface, allowing users to view event attributes by clicking on nodes and perform more complex queries using Cypher statements. This lithium-ion battery defect knowledge graph can intuitively display the cause, process, and result of lithium-ion battery defect events, as well as the relationships between the subjects and objects involved in the events.

[0067] Based on the constructed event graph of lithium-ion battery defects, predictions can be made for any newly generated event instance.

[0068] In summary, by utilizing the above technical solutions, this application obtains lithium-ion battery defect data, preprocesses the lithium-ion battery defect data to obtain preprocessed data, uses the BERT-BiLSTM-CRF model to extract defect events from the preprocessed data to obtain defect event extraction results, uses a defect event relationship extraction model to extract the relationships between various defect events based on the defect event extraction results to obtain defect event relationship extraction results, and constructs a lithium-ion battery defect event graph based on the defect event extraction results and defect event relationship extraction results, and performs evolution prediction based on the lithium-ion battery defect event graph, thereby preventing the occurrence of defect events.

[0069] It should be understood that the above-described method for predicting the evolution of lithium-ion battery defect events is merely exemplary, and those skilled in the art can make various modifications based on the above method, and the modified solutions also fall within the protection scope of this application.

[0070] Please see Figure 4 , Figure 4 A structural block diagram of a lithium-ion battery defect event evolution prediction device 400 provided in an embodiment of this application is shown. It should be understood that the lithium-ion battery defect event evolution prediction device 400 is capable of performing the various steps in the above method embodiments. The specific functions of the lithium-ion battery defect event evolution prediction device 400 can be found in the description above; to avoid repetition, detailed descriptions are appropriately omitted here. The lithium-ion battery defect event evolution prediction device 400 includes at least one software function module that can be stored in a memory or embedded in the operating system (OS) of the lithium-ion battery defect event evolution prediction device 400 in the form of software or firmware. Specifically, the lithium-ion battery defect event evolution prediction device 400 includes:

[0071] The acquisition module 410 is used to acquire lithium-ion battery defect data and preprocess the lithium-ion battery defect data to obtain preprocessed data.

[0072] The defect event extraction module 420 is used to extract defect events from the preprocessed data using the BERT-BiLSTM-CRF model to obtain the defect event extraction results. The BERT-BiLSTM-CRF model includes a BERT feature extraction layer, a multi-granularity feature extraction layer, an attention mechanism layer, a dynamic weight fusion layer, a sequence modeling layer, and a classification decoding layer. The BERT feature extraction layer is connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the dynamic weight fusion layer, and the dynamic weight fusion layer is connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the sequence modeling layer, and the sequence modeling layer is also connected to the classification decoding layer.

[0073] The relation extraction module 430 is used to extract the relationships between various defect events based on the defect event extraction results using the defect event relation extraction model, thereby obtaining the defect event relation extraction results; wherein, the defect event relation extraction model is obtained by improving the prompting learning framework.

[0074] The evolution prediction module 440 is used to construct a lithium-ion battery defect event graph based on the defect event extraction results and defect event relationship extraction results, and to perform evolution prediction based on the lithium-ion battery defect event graph.

[0075] In one possible embodiment, the multi-granularity feature extraction layer includes a multi-scale convolution module and an output concatenation module. The multi-scale convolution module is used to scan the word vector sequence output by the BERT feature extraction layer in parallel using convolution kernels of different widths to obtain multiple convolution results. The output concatenation module is used to concatenate the multiple convolution results along the feature dimension to obtain multi-scale fused features.

[0076] In one possible embodiment, the attention mechanism layer includes a word-level attention module and a sentence-level attention module. The word-level attention module is used to calculate the association weights between words in the lithium-ion battery defect data, and the sentence-level attention module is used to perform a weighted average of all word vectors in the entire sentence in the lithium-ion battery defect data to generate a global sentence vector.

[0077] In one possible embodiment, the sequence modeling layer comprises three sequentially connected bidirectional LSTM modules.

[0078] In one possible embodiment, the defect event relation extraction model includes an input layer, a pre-trained language model layer, a parallel answer prediction layer, and a fusion discriminant layer connected in sequence. The input layer is used to construct a prompt template based on the defect event extraction results, comprising two defect events and two relation prediction mask bits for the relation to be determined. The pre-trained language model layer is used to parse the contextual semantics of the prompt template and extract the relation representation vectors for the mask positions. The parallel answer prediction layer is used to independently calculate the candidate relation probability distribution for each mask position. The fusion discriminant layer is used to independently select the optimal relation label for each mask position.

[0079] Since the apparatus described in the above embodiments of the present invention is an apparatus used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and variations of the apparatus based on the methods described in the above embodiments of the present invention, and therefore will not be described again here. All apparatuses used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0082] It should be noted that the word "a" or "an" preceding a component does not preclude the existence of multiple such components. This invention can be implemented using hardware comprising several different components and using a suitably programmed computer. Among the listed devices, several of these devices may be embodied by the same hardware. The use of terms such as "first," "second," "third," etc., is merely for convenience and does not indicate any order. These terms can be understood as part of the component names.

[0083] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0084] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the technical solution should be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0085] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the invention's technical solutions and their equivalents, then the invention should also include these modifications and variations.

Claims

1. A method for predicting the evolution of defect events in lithium-ion batteries, characterized in that, include: Obtain lithium-ion battery defect data and preprocess the lithium-ion battery defect data to obtain preprocessed data; The BERT-BiLSTM-CRF model is used to extract defect events from the preprocessed data to obtain defect event extraction results. The BERT-BiLSTM-CRF model includes a BERT feature extraction layer, a multi-granularity feature extraction layer, an attention mechanism layer, a dynamic weight fusion layer, a sequence modeling layer, and a classification decoding layer. The BERT feature extraction layer is connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the dynamic weight fusion layer, respectively. The dynamic weight fusion layer is also connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the sequence modeling layer, respectively. Furthermore, the sequence modeling layer is connected to the classification decoding layer. Based on the defect event extraction results, the relationship between each defect event is extracted using a defect event relationship extraction model to obtain defect event relationship extraction results; wherein, the defect event relationship extraction model is obtained by improving the prompting learning framework; Based on the defect event extraction results and the defect event relationship extraction results, a lithium-ion battery defect event graph is constructed, and evolution prediction is performed based on the lithium-ion battery defect event graph. The defect event relation extraction model comprises an input layer, a pre-trained language model layer, a parallel answer prediction layer, and a fusion discriminant layer connected in sequence. The input layer is used to construct a prompt template based on the defect event extraction results, including two defect events with a relation to be determined and two relation prediction mask bits. The pre-trained language model layer is used to parse the contextual semantics of the prompt template and extract the relation representation vector at each mask position. The parallel answer prediction layer is used to independently calculate the candidate relation probability distribution for each mask position. The fusion discriminant layer is used to independently select the optimal relation label for each mask position.

2. The lithium-ion battery defect event evolution prediction method according to claim 1, characterized in that, The multi-granularity feature extraction layer includes a multi-scale convolution module and an output concatenation module. The multi-scale convolution module is used to scan the word vector sequence output by the BERT feature extraction layer in parallel using convolution kernels of different widths to obtain multiple convolution results. The output concatenation module is used to concatenate the multiple convolution results along the feature dimension to obtain multi-scale fused features.

3. The lithium-ion battery defect event evolution prediction method according to claim 2, characterized in that, The attention mechanism layer includes a word-level attention module and a sentence-level attention module. The word-level attention module is used to calculate the association weights between words in the lithium-ion battery defect data, and the sentence-level attention module is used to perform a weighted average of all word vectors in the entire sentence in the lithium-ion battery defect data to generate a global sentence vector.

4. The lithium-ion battery defect event evolution prediction method according to claim 3, characterized in that, The sequence modeling layer comprises three bidirectional LSTM modules connected in sequence.

5. A lithium-ion battery defect event evolution prediction device, characterized in that, include: The acquisition module is used to acquire lithium-ion battery defect data and preprocess the lithium-ion battery defect data to obtain preprocessed data. The defect event extraction module is used to extract defect events from the preprocessed data using the BERT-BiLSTM-CRF model to obtain defect event extraction results. The BERT-BiLSTM-CRF model includes a BERT feature extraction layer, a multi-granularity feature extraction layer, an attention mechanism layer, a dynamic weight fusion layer, a sequence modeling layer, and a classification decoding layer. The BERT feature extraction layer is connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the dynamic weight fusion layer, respectively. The dynamic weight fusion layer is also connected to the multi-granularity feature extraction layer, the attention mechanism layer, and the sequence modeling layer, respectively. Furthermore, the sequence modeling layer is connected to the classification decoding layer. The relation extraction module is used to extract the relationship between each of the defect events based on the defect event extraction results using a defect event relation extraction model, thereby obtaining defect event relation extraction results; wherein, the defect event relation extraction model is obtained by improving the prompting learning framework; An evolution prediction module is used to construct a lithium-ion battery defect event graph based on the defect event extraction results and the defect event relationship extraction results, and to perform evolution prediction based on the lithium-ion battery defect event graph. The defect event relation extraction model comprises an input layer, a pre-trained language model layer, a parallel answer prediction layer, and a fusion discriminant layer connected in sequence. The input layer is used to construct a prompt template based on the defect event extraction results, including two defect events with a relation to be determined and two relation prediction mask bits. The pre-trained language model layer is used to parse the contextual semantics of the prompt template and extract the relation representation vector at each mask position. The parallel answer prediction layer is used to independently calculate the candidate relation probability distribution for each mask position. The fusion discriminant layer is used to independently select the optimal relation label for each mask position.

6. The lithium-ion battery defect event evolution prediction device according to claim 5, characterized in that, The multi-granularity feature extraction layer includes a multi-scale convolution module and an output concatenation module. The multi-scale convolution module is used to scan the word vector sequence output by the BERT feature extraction layer in parallel using convolution kernels of different widths to obtain multiple convolution results. The output concatenation module is used to concatenate the multiple convolution results along the feature dimension to obtain multi-scale fused features.

7. The lithium-ion battery defect event evolution prediction device according to claim 6, characterized in that, The attention mechanism layer includes a word-level attention module and a sentence-level attention module. The word-level attention module is used to calculate the association weights between words in the lithium-ion battery defect data, and the sentence-level attention module is used to perform a weighted average of all word vectors in the entire sentence in the lithium-ion battery defect data to generate a global sentence vector.

8. The lithium-ion battery defect event evolution prediction device according to claim 7, characterized in that, The sequence modeling layer comprises three bidirectional LSTM modules connected in sequence.

Images