TDNN and LLM fused lithium battery electrolyte ultrasonic quantitative detection method

Abstract

The invention discloses a TDNN and LLM fused lithium battery electrolyte ultrasonic quantitative detection method, and relates to the field of lithium ion battery health management, and the method comprises the steps: obtaining original ultrasonic waveform data, and carrying out the preprocessing; based on a pre-trained time delay neural network TDNN, extracting a high-dimensional depth feature vector; mapping to a text embedding space of a large language model LLM through a linear projection layer, generating a text prototype and constructing a complete prompt sequence; reasoning to obtain an electrolyte content predicted value based on a pre-trained large language model LLM; calculating a global estimated value, and carrying out physical correction on the global estimated value by utilizing the infiltration area ratio to obtain the corrected electrolyte content; and outputting the corrected electrolyte content and the two-dimensional distribution diagram of the electrolyte in the battery. According to the method, deep features of ultrasonic signals are learned through a cascade architecture of TDNN and LLM, reasoning is carried out, physical correction is supplemented, and high-sensitivity and reliable quantitative detection of the content of the lithium battery electrolyte is achieved.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery health management, and more specifically to an ultrasonic quantitative detection method for lithium battery electrolyte that integrates TDNN and LLM. Background Technology

[0002] In the field of lithium-ion battery health management, quantitative detection of electrolyte is a crucial step in ensuring battery performance and safety. In the existing technology, ultrasonic-based non-destructive testing methods have attracted attention due to their low cost and sensitivity to internal structures. However, current mainstream methods usually rely on manually extracting a few traditional acoustic features, such as peak amplitude or time of flight, from the raw ultrasonic signal and attempting to establish a simple statistical correlation with electrolyte volume. This artificial feature engineering severely limits the model's perception capability because the manually selected macroscopic features are difficult to capture the complex, nonlinear acoustic response patterns generated by minute changes in electrolyte within the porous electrode of the battery. This results in the model being extremely insensitive to small changes in electrolyte content, and the prediction accuracy drops sharply in the critical low content range.

[0003] A further drawback is that existing models based on traditional features usually implicitly assume a uniform distribution of electrolyte. However, in reality, electrolyte reduction often manifests first as local shrinkage and uneven distribution of the wetting area. This dynamic change in spatial distribution cannot be effectively characterized by a few global statistical features. As a result, when the battery is unevenly wetted, the prediction error of the model will increase significantly, and the generalization ability will be seriously insufficient.

[0004] In addition, the technical approaches used to achieve detection also have limitations. On the one hand, some high-precision non-destructive technologies, such as X-ray or neutron imaging, are difficult to deploy online in industrial settings due to expensive equipment and complex processes. On the other hand, traditional single-element ultrasonic C-scanning technology usually requires water immersion coupling, which poses safety risks, is cumbersome, and has low scanning efficiency. These factors have hindered the large-scale application of high-precision ultrasonic quantitative detection on battery production lines.

[0005] Therefore, how to design an ultrasonic quantitative detection method for lithium battery electrolyte that integrates TDNN and LLM, capable of highly sensitively sensing minute changes in the electrolyte and robustly handling uneven wetting conditions, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the purpose of this technical solution is to provide an ultrasonic quantitative detection method for lithium battery electrolytes, so as to solve the problems of existing technologies that rely on manual feature extraction, resulting in insensitivity to small changes, inability to effectively handle uneven wetting, poor model generalization ability, and complex deployment. This method aims to achieve a non-destructive testing solution with higher accuracy, better reliability, and more suitable for industrial online applications.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A quantitative ultrasonic detection method for lithium battery electrolytes integrating TDNN and LLM includes the following steps: S1. Acquire the raw ultrasonic waveform data under full-area scanning of the lithium battery; S2. Preprocess the original ultrasonic waveform data to obtain a preprocessed waveform sequence; S3. Input the preprocessed waveform sequence into a pre-trained time-delay neural network (TDNN) to extract high-dimensional deep feature vectors; S4. Map the high-dimensional deep feature vector to the text embedding space of the large language model LLM through a linear projection layer to generate a text prototype, and concatenate prefix prompts to construct a complete prompt sequence; S5. The pre-trained large language model LLM with the complete prompt sequence input parameters frozen is used to perform inference to obtain the predicted electrolyte content of each local region. S6. Calculate the global estimate based on the predicted electrolyte content of each local area, and use the wetting area ratio to physically correct the global estimate to obtain the corrected electrolyte content. S7. Output the corrected electrolyte content and the two-dimensional distribution map of the electrolyte inside the battery.

[0008] Preferably, in step S1, a dry-coupled linear array sensor is used to scan the entire area of ​​the lithium battery, and the original ultrasonic waveform data in the time domain is recorded at each scanning point at a sampling frequency of 50MHz.

[0009] Preferably, S2 includes: The original ultrasonic waveform data is subjected to reversible instance normalization processing, through... Eliminate distribution offset caused by sensors or environment, where x is the original waveform data point, μ is the mean of the waveform data at that scan point, σ is the standard deviation of the waveform data at that scan point, and x′ is the normalized data point; The preprocessed waveform sequence is formed by extracting key segments containing direct echoes and at least one key interface echo from the normalized waveform.

[0010] Preferably, in S3, the time-delay neural network TDNN adopts a feedforward neural network with a time-delay structure, which contains multiple hidden layers, uses the Sigmoid function as the activation function of the hidden layers, and outputs a high-dimensional deep feature vector with a dimension of 256. Its hidden layer has a time delay window d=3, and the input at time t is [ , , ],in This is the input vector of the preprocessed waveform sequence at time t. Preferably, S4 includes: The high-dimensional deep feature vector is mapped to the LLM text embedding space through a trainable linear projection layer to obtain the initial text embedding vector. ; The initial text embedding vector Decoded into a text prototype representing a local infiltration state; The text prototype is concatenated with a preset prefix prompt, which includes task instructions, battery model information, and historical infiltration area change trends.

[0011] Preferably, in S5, the large language model LLM is a pre-trained GPT-2 model based on the Transformer architecture, with a hidden layer dimension of 768, an attention head of 12, and the model parameters are kept frozen during inference. The final hidden state output of the LLM is passed through a fully connected regression output layer to obtain the predicted electrolyte content values ​​for each local region. , where i represents the i-th local region, and the activation function of the regression output layer is a linear function.

[0012] Preferably, S6 includes: S61. Calculate the global estimate based on the predicted electrolyte content of each local region. The average of the median and mode of all local predicted values ​​is used to combat outlier interference. S62. Calculate the actual infiltration area from the current maximum ultrasonic amplitude image of the battery using an image processing algorithm. ; S63. Obtain the standard wetting area of ​​the same type of battery under the optimal electrolyte injection volume. ; S64. Calculate the wetted area ratio And a global valuation Scaling is performed to obtain the corrected electrolyte content. .

[0013] Preferably, S62 includes: Based on the original ultrasonic waveform data, the waveform corresponding to each spatial scanning point is calculated within the time window. , The maximum absolute amplitude within the range is determined, and a two-dimensional ultrasound maximum amplitude image I is generated based on the spatial position of each scanning point. An adaptive thresholding algorithm based on local image statistical features was used to segment the ultrasound maximum amplitude image I to obtain a binary image B; in which, the region with a pixel value of 1 was initially identified as the infiltrated region, and the region with a pixel value of 0 was identified as the non-infiltrated region. Morphological processing is performed on the segmented binary image B to obtain the optimized binary image. ; The largest connected region with a pixel value of 1 is identified, its pixel count N is calculated, and the actual infiltration area is obtained by combining the image spatial calibration coefficient λ. .

[0014] Preferably, the segmentation threshold of the adaptive threshold algorithm , , This represents the mean and standard deviation of pixel grayscale values ​​within the local neighborhood Ω of each pixel.

[0015] Preferably, S7 includes: The corrected electrolyte content is output as the overall quantitative content of the battery electrolyte. The predicted electrolyte content of each local area is mapped according to the spatial coordinates of its corresponding scanning point and reconstructed into a two-dimensional data matrix of electrolyte content corresponding to the battery spatial area. The two-dimensional data matrix is ​​subjected to pseudo-color rendering to generate a two-dimensional distribution map of the electrolyte inside the battery, where different colors are mapped to different electrolyte content ranges.

[0016] As can be seen from the above technical solution, compared with the prior art, the technical solution of the present invention has the following beneficial effects: 1. This method automatically extracts high-dimensional depth features directly from the original ultrasonic waveform using a time-delay neural network (TDNN), replacing the traditional method that relies on manually designed features. This enables the model to capture nonlinear temporal patterns related to the microscopic wetting process of the electrolyte, thus improving its ability to perceive minute changes in electrolyte content.

[0017] 2. A physical correction mechanism based on the measured infiltration area of ​​ultrasonic images is introduced, which combines data-driven prediction with electrochemical physical laws. The global estimate is scaled by the infiltration area ratio, which effectively avoids the unreasonable output of the pure data model in the case of non-uniform infiltration, and improves the physical consistency and generalization ability of the results.

[0018] 3. By mapping the deep features extracted by TDNN to the semantic space of the large language model LLM and constructing a prompt sequence containing task instructions and model information, the inference capability of LLM is used to achieve adaptive analysis of different battery structures, forming a detection process that does not require complex feature engineering, which is convenient for deployment and application in industrial online inspection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 The flowchart illustrates a quantitative ultrasonic detection method for lithium battery electrolyte that integrates TDNN and LLM, as provided in this embodiment of the invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, this embodiment provides an ultrasonic quantitative detection method for lithium battery electrolyte that integrates TDNN and LLM, including the following steps. S1. Acquire the raw ultrasonic waveform data under full-area scanning of the lithium battery; S2. Preprocess the original ultrasonic waveform data to obtain a preprocessed waveform sequence; S3. Input the preprocessed waveform sequence into a pre-trained time-delay neural network (TDNN) to extract high-dimensional deep feature vectors; S4. Map the high-dimensional deep feature vector to the text embedding space of the large language model LLM through a linear projection layer to generate a text prototype, and concatenate prefix prompts to construct a complete prompt sequence; S5. The pre-trained large language model LLM with the complete prompt sequence input parameters frozen is used to perform inference to obtain the predicted electrolyte content of each local region. S6. Calculate the global estimate based on the predicted electrolyte content of each local area, and use the wetting area ratio to physically correct the global estimate to obtain the corrected electrolyte content. S7. Output the corrected electrolyte content and the two-dimensional distribution map of the electrolyte inside the battery.

[0023] It automatically extracts deep features from the original ultrasonic signal through a time-delay neural network (TDNN) and combines the semantic reasoning capability of a large language model (LLM) to achieve high sensitivity to subtle changes in electrolyte content. At the same time, it introduces a physical correction mechanism based on the wetting area to improve the prediction reliability and physical rationality of the model under uneven wetting conditions. Finally, the end-to-end process design reduces the reliance on manual feature engineering and enhances the applicability of the method in different battery models and industrial online inspection scenarios.

[0024] The following provides a further detailed explanation of each step in the above method; In this embodiment, S1, the original ultrasonic waveform data under full-area scanning of the lithium battery is obtained. A dry-coupled linear array sensor is used to scan the entire area of ​​the lithium battery, and the original ultrasonic waveform data in the time domain is recorded at each scanning point at a sampling frequency of 50MHz.

[0025] In this embodiment S2, the original ultrasonic waveform data is preprocessed to obtain a preprocessed waveform sequence; including: The original ultrasonic waveform data is subjected to reversible instance normalization processing, through... Eliminate distribution offset caused by sensors or environment, where x is the original waveform data point, μ is the mean of the waveform data at that scan point, σ is the standard deviation of the waveform data at that scan point, and x′ is the normalized data point; The preprocessed waveform sequence is formed by extracting key segments containing direct echoes and at least one key interface echo from the normalized waveform.

[0026] In this embodiment, S3, the preprocessed waveform sequence is input into a pre-trained time-delay neural network (TDNN) to extract high-dimensional deep feature vectors. Among them, the Time Delay Neural Network (TDNN) adopts a feedforward neural network with a time delay structure, which contains multiple hidden layers and uses the Sigmoid function as the activation function of the hidden layers, outputting a high-dimensional deep feature vector with a dimension of 256. Its hidden layer has a time delay window d=3, and the input at time t is [ , , ],in This is the input vector of the preprocessed waveform sequence at time t; In this step, the TDNN performs a deep analysis of the preprocessed ultrasonic wave sequence through its time delay structure. The input of the hidden layer of this network at time t includes not only the input vector at the current time. It also includes the input vectors from the first two time steps. and This time delay window (d=3) design enables the network to perceive the dynamic change pattern of the waveform within a local time range. Each hidden layer uses the Sigmoid activation function to perform nonlinear transformation on the weighted input, gradually building a hierarchical understanding of the waveform's temporal characteristics. As the signal propagates forward through multiple hidden layers, the network extracts and combines features layer by layer through cascaded nonlinear transformations. Lower layers mainly capture local details and simple patterns in the waveform, while higher layers integrate these basic features into complex abstract features that can characterize the electrolyte wetting state. Finally, the network outputs a 256-dimensional deep feature vector, which condenses the essential information in the original waveform that is closely related to the electrolyte content and distribution. Its information density and discriminative power far exceed the artificially designed acoustic features in traditional methods.

[0027] In this embodiment, S4, the high-dimensional deep feature vector is mapped to the text embedding space of the Large Language Model (LLM) through a linear projection layer to generate a text prototype, and prefix suggestions are concatenated to construct a complete suggestion sequence; including: The high-dimensional deep feature vector is mapped to the LLM text embedding space through a trainable linear projection layer to obtain the initial text embedding vector. ; The initial text embedding vector Decoded into a text prototype representing a local infiltration state; The text prototype is concatenated with a preset prefix prompt, which includes task instructions, battery model information, and historical infiltration area change trends.

[0028] In this step, the linear projection layer is a trainable fully connected neural network layer that linearly transforms the high-dimensional deep feature vector output by the TDNN through a weight matrix and a bias vector, mapping it from the original acoustic feature space to the text embedding space inherent in the LLM. This achieves modal alignment between numerical acoustic features and symbolic language models, laying the foundation for subsequent use of the semantic reasoning capabilities of the LLM.

[0029] In this embodiment S5, the pre-trained large language model LLM with the complete prompt sequence input parameters frozen is used to perform inference to obtain the predicted electrolyte content of each local region. Among them, the large language model LLM is a pre-trained GPT-2 model based on the Transformer architecture, with a hidden layer dimension of 768, an attention head of 12, and the model parameters are kept frozen during inference. The final hidden state output of the LLM is passed through a fully connected regression output layer to obtain the predicted electrolyte content for each local region. , where i represents the i-th local region, and the activation function of the regression output layer is a linear function.

[0030] In this step, the pre-trained LLM with frozen parameters receives a complete sequence consisting of text prototypes and prefix cues. The model utilizes the multi-layer self-attention mechanism in its Transformer architecture to perform interactive computation on each element in the input sequence. Twelve attention heads work in parallel, enabling the model to capture the complex relationships between deep acoustic features and contextual information such as task instructions and battery models from different semantic dimensions. Through this mechanism, the LLM transfers its language understanding and logical reasoning capabilities, which it has acquired through pre-training on massive amounts of text data, to the parsing of acoustic feature sequences, thereby achieving a deep semantic understanding of electrolyte distribution patterns. Furthermore, the final hidden state output of LLM is fed into a specially designed fully connected regression output layer. This layer maps the 768-dimensional hidden state vector of LLM output into a single numerical prediction. Its linear activation function ensures the continuity of the output value and directly corresponds to the electrolyte content estimation of each local region. This allows the semantic reasoning ability of LLM to be effectively transformed into specific quantitative predictions, while maintaining the model's generalization performance across battery models. Accurate content inversion can be achieved without fine-tuning parameters for specific tasks.

[0031] In this embodiment, S6, a global estimate is calculated based on the predicted electrolyte content of each local region, and the global estimate is physically corrected using the wetting area ratio to obtain the corrected electrolyte content; including: S61. Calculate the global estimate based on the predicted electrolyte content of each local region. The average of the median and mode of all local predicted values ​​is used to combat outlier interference. S62. Calculate the actual infiltration area from the current maximum ultrasonic amplitude image of the battery using an image processing algorithm. ; S63. Obtain the standard wetting area of ​​the same type of battery under the optimal electrolyte injection volume. ; S64. Calculate the wetted area ratio And a global valuation Scaling is performed to obtain the corrected electrolyte content. .

[0032] Furthermore, S62 includes: Based on the original ultrasonic waveform data, the waveform corresponding to each spatial scanning point is calculated within the time window. , The maximum absolute amplitude within the range is determined, and a two-dimensional ultrasound maximum amplitude image I is generated based on the spatial position of each scanning point. An adaptive thresholding algorithm based on local image statistical features was used to segment the ultrasound maximum amplitude image I to obtain a binary image B; in which, the region with a pixel value of 1 was initially identified as the infiltrated region, and the region with a pixel value of 0 was identified as the non-infiltrated region. Morphological processing is performed on the segmented binary image B to obtain the optimized binary image. ; The largest connected region with a pixel value of 1 is identified, its pixel count N is calculated, and the actual infiltration area is obtained by combining the image spatial calibration coefficient λ. .

[0033] Furthermore, the segmentation threshold of the adaptive thresholding algorithm , , This represents the mean and standard deviation of pixel grayscale values ​​within the local neighborhood Ω of each pixel.

[0034] In this step, by combining data-driven prediction results with the physical laws of the infiltration area measured based on ultrasonic images, the global estimate is deterministically corrected using the calculable infiltration area ratio. This effectively embeds the physical prior knowledge of the electrochemical system into the system, improving the physical rationality of the quantitative detection results and the reliability in actual non-uniform infiltration scenarios.

[0035] In this embodiment S7, the corrected electrolyte content and the two-dimensional distribution map of the electrolyte inside the battery are output; including: The corrected electrolyte content is output as the overall quantitative content of the battery electrolyte. The predicted electrolyte content of each local area is mapped according to the spatial coordinates of its corresponding scanning point and reconstructed into a two-dimensional data matrix of electrolyte content corresponding to the battery spatial area. The two-dimensional data matrix is ​​subjected to pseudo-color rendering to generate a two-dimensional distribution map of the electrolyte inside the battery, where different colors are mapped to different electrolyte content ranges.

[0036] This embodiment presents a quantitative ultrasonic detection method for lithium battery electrolyte that integrates TDNN and LLM. It achieves full-area ultrasonic data acquisition without water immersion using a dry-coupled linear array sensor. After signal preprocessing through reversible instance normalization, high-dimensional deep acoustic features are extracted based on the time delay structure of TDNN. Then, the features are aligned with the modality of the LLM text embedding space through a linear projection layer. The input sequence is constructed by combining prefix prompts. Local electrolyte content inference is completed using a pre-trained LLM with parameter freezing. Finally, the correction factor is calculated based on the measured immersion area of ​​the ultrasonic image to physically constrain the global estimate, and the quantitative result and two-dimensional distribution image are output. The entire process is designed end-to-end to reduce human intervention. It ensures the detection sensitivity of minute content changes through the collaboration of TDNN and LLM, and ensures the reliability of results through physical correction mechanisms. It can meet the actual needs of quantitative analysis of electrolyte distribution and content in online testing scenarios of different types of lithium batteries.

[0037] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0038] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for ultrasonic quantitative detection of lithium battery electrolyte integrating TDNN and LLM, characterized in that, Includes the following steps: S1. Acquire the raw ultrasonic waveform data under full-area scanning of the lithium battery; S2. Preprocess the original ultrasonic waveform data to obtain a preprocessed waveform sequence; S3. Input the preprocessed waveform sequence into a pre-trained time-delay neural network (TDNN) to extract high-dimensional deep feature vectors; S4. Map the high-dimensional deep feature vector to the text embedding space of the large language model LLM through a linear projection layer to generate a text prototype, and concatenate prefix prompts to construct a complete prompt sequence; S5. The pre-trained large language model LLM with the complete prompt sequence input parameters frozen is used to perform inference to obtain the predicted electrolyte content of each local region. S6. Calculate the global estimate based on the predicted electrolyte content of each local area, and use the wetting area ratio to physically correct the global estimate to obtain the corrected electrolyte content. S7. Output the corrected electrolyte content and the two-dimensional distribution map of the electrolyte inside the battery.

2. The ultrasonic quantitative detection method for lithium battery electrolyte integrating TDNN and LLM according to claim 1, characterized in that, In step S1, a dry-coupled linear array sensor is used to scan the entire area of ​​the lithium battery, and the original ultrasonic waveform data in the time domain is recorded at each scanning point at a sampling frequency of 50MHz.

3. The ultrasonic quantitative detection method for lithium battery electrolyte integrating TDNN and LLM according to claim 1, characterized in that, S2 includes: The original ultrasonic waveform data is subjected to reversible instance normalization processing, through... Eliminate distribution offset caused by sensors or environment, where x is the original waveform data point, μ is the mean of the waveform data at that scan point, σ is the standard deviation of the waveform data at that scan point, and x′ is the normalized data point; The preprocessed waveform sequence is formed by extracting key segments containing direct echoes and at least one key interface echo from the normalized waveform.

4. The ultrasonic quantitative detection method for lithium battery electrolyte integrating TDNN and LLM according to claim 1, characterized in that, In S3, the time-delay neural network TDNN adopts a feedforward neural network with a time-delay structure, which contains multiple hidden layers, uses the sigmoid function as the activation function of the hidden layers, and outputs a high-dimensional deep feature vector with a dimension of 256. Its hidden layer has a time delay window d=3, and the input at time t is [ , , ],in This is the input vector of the preprocessed waveform sequence at time t.

5. The ultrasonic quantitative detection method for lithium battery electrolyte integrating TDNN and LLM according to claim 1, characterized in that, S4 includes: The high-dimensional deep feature vector is mapped to the LLM text embedding space through a trainable linear projection layer to obtain the initial text embedding vector. ; The initial text embedding vector Decoded into a text prototype representing a local infiltration state; The text prototype is concatenated with a preset prefix prompt, which includes task instructions, battery model information, and historical infiltration area change trends.

6. The ultrasonic quantitative detection method for lithium battery electrolyte integrating TDNN and LLM according to claim 1, characterized in that, In S5, the large language model LLM is a pre-trained GPT-2 model based on the Transformer architecture, with a hidden layer dimension of 768, 12 attention heads, and the model parameters are kept frozen during inference. The final hidden state output of the LLM is passed through a fully connected regression output layer to obtain the predicted electrolyte content values ​​for each local region. , where i represents the i-th local region, and the activation function of the regression output layer is a linear function.

7. The ultrasonic quantitative detection method for lithium battery electrolyte integrating TDNN and LLM according to claim 1, characterized in that, S6 includes: S61. Calculate the global estimate based on the predicted electrolyte content of each local region. The average of the median and mode of all local predicted values ​​is used to combat outlier interference. S62. Calculate the actual infiltration area from the current maximum ultrasonic amplitude image of the battery using an image processing algorithm. ; S63. Obtain the standard wetting area of ​​the same type of battery under the optimal electrolyte injection volume. ; S64. Calculate the wetted area ratio And a global valuation Scaling is performed to obtain the corrected electrolyte content. .

8. The ultrasonic quantitative detection method for lithium battery electrolyte integrating TDNN and LLM according to claim 7, characterized in that, S62 includes: Based on the original ultrasonic waveform data, the waveform corresponding to each spatial scanning point is calculated within the time window. , The maximum absolute amplitude within the range is determined, and a two-dimensional ultrasound maximum amplitude image I is generated based on the spatial position of each scanning point. An adaptive thresholding algorithm based on local image statistical features was used to segment the ultrasound maximum amplitude image I to obtain a binary image B; in which, the region with a pixel value of 1 was initially identified as the infiltrated region, and the region with a pixel value of 0 was identified as the non-infiltrated region. Morphological processing is performed on the segmented binary image B to obtain the optimized binary image. ; The largest connected region with a pixel value of 1 is identified, its pixel count N is calculated, and the actual infiltration area is obtained by combining the image spatial calibration coefficient λ. .

9. The ultrasonic quantitative detection method for lithium battery electrolyte integrating TDNN and LLM according to claim 8, characterized in that, The segmentation threshold of the adaptive threshold algorithm , , This represents the mean and standard deviation of pixel grayscale values ​​within the local neighborhood Ω of each pixel.

10. The ultrasonic quantitative detection method for lithium battery electrolyte integrating TDNN and LLM according to claim 1, characterized in that, S7 includes: The corrected electrolyte content is output as the overall quantitative content of the battery electrolyte. The predicted electrolyte content of each local area is mapped according to the spatial coordinates of its corresponding scanning point and reconstructed into a two-dimensional data matrix of electrolyte content corresponding to the battery spatial area. The two-dimensional data matrix is ​​subjected to pseudo-color rendering to generate a two-dimensional distribution map of the electrolyte inside the battery, where different colors are mapped to different electrolyte content ranges.

Images