Sodium battery residual life prediction method for underwater robot application scene

Abstract

An underwater robot application scene-oriented sodium battery residual life prediction method comprises the following steps of: obtaining normalized features after input features pass through a double-path feature normalization module; processing the input features through a multi-local sensing path module, and fusing and splicing the processed input features with the normalized features to obtain spliced features; performing 1 * 1 convolution operation on the input features, adding the input features and the splicing features element by element, and performing processing through a lightweight residual connection module to obtain multi-scale features; enabling the coder-decoder to predict an initial capacity prediction value according to characteristics rich in global time sequence information in the multi-scale characteristic data; performing element-by-element addition on the initial capacity prediction value and the normalized feature, and calculating a final sodium battery residual life prediction value; the coder-decoder can capture the long-range dependency relationship of the battery capacity sequence through a self-attention mechanism, the gradient disappearance problem of the recurrent neural network in the long sequence is solved, and the prediction precision is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of sodium ion battery health status estimation, and in particular to a sodium battery remaining life prediction method for underwater robot application scenarios. Background Art

[0002] The widespread application of underwater robots in deep-sea exploration, seabed resource development, military security, and other fields places extremely high demands on the performance of sodium batteries. Due to the complex deep-sea environment, sodium batteries must not only possess high energy density to support long-term operations but also meet special design requirements such as pressure resistance, waterproofness, and corrosion resistance. However, sodium batteries inevitably experience aging phenomena such as capacity decay and increased internal resistance during repeated charge and discharge cycles, directly impacting the endurance and mission reliability of underwater robots.

[0003] Traditional remaining useful life (RUL) prediction methods rely on complete charge and discharge data, failing to identify favorable aging characteristics. They struggle to adapt to the dynamic and fragmented data collection scenarios of deep-sea operations and are unable to effectively model the nonlinear characteristics of battery degradation. Furthermore, the high maintenance costs of underwater robots necessitate high-precision RUL prediction technology to optimize battery replacement cycles, reduce operational risks, and improve mission success rates. Summary of the Invention

[0004] In response to the needs of the prior art, the present invention provides a sodium battery remaining life prediction method for underwater robot application scenarios, aiming to solve the problems of extracting favorable features of battery aging and modeling long-term dependencies, and to improve prediction accuracy.

[0005] A method for predicting the remaining life of sodium batteries for underwater robot applications includes the following steps:

[0006] Step 1: Build and obtain a trained sodium battery remaining life prediction network, which includes a dual-path feature normalization module, a multi-local perception path module, a lightweight residual connection module, and an encoder-decoder;

[0007] Step 2: Obtain sodium battery data and extract the label entropy of voltage data in the sodium battery data as the input feature of the sodium battery remaining life prediction network;

[0008] Step 3: After the input features pass through the dual-path feature normalization module, normalized features are obtained;

[0009] Step 4: Process the input features through the multi-local perception path module and fuse and splice them with the normalized features to obtain the spliced ​​features;

[0010] Step 5: The input features are convolved with the concatenated features and then added element-by-element. The addition result is processed by a lightweight residual connection module to obtain multi-scale features.

[0011] Step 6: The encoder-decoder predicts an initial capacity prediction value based on the features rich in global temporal information in the multi-scale feature data;

[0012] Step 7: Add the initial capacity prediction value and the normalized feature element by element to obtain the correction value. According to the time step and sampling interval of the correction value, calculate and output the remaining service life of the battery from the current moment, that is, the final remaining service life prediction value of the sodium battery.

[0013] Further, the encoder in the encoder-decoder includes a linear transformation module, a scaled dot product calculation module, a gated sparse mask generation module, a weighted calculation module, a multi-head attention splicing and linear mapping module, a first hybrid normalization module, a sparse feedforward network and a second hybrid normalization module; step 6 is specifically:

[0014] Step 6.1: Use the linear transformation module to transform the multi-scale feature data into query matrix, key matrix, and value matrix respectively;

[0015] Step 6.2: Calculate the attention weight of each moment to the remaining time steps through the scaled dot product calculation module to obtain the original weighted relationship matrix;

[0016] Step 6.3: The original weighted relationship matrix is ​​introduced into the gated sparse mask generation module, and the gated score matrix is ​​calculated based on the matching relationship between the query matrix and the key matrix;

[0017] Step 6.4: The original weighted relationship matrix and the gated score matrix are weighted and output through the weighted calculation module to obtain the sparse attention weight;

[0018] Step 6.5: Introduce the sparse attention weights into the multi-head attention concatenation and linear mapping module to output the multi-head self-attention representation;

[0019] Step 6.6: The multi-head self-attention representation is sequentially passed through the first linear layer, ReLU activation, and another linear layer to obtain the intermediate feature response;

[0020] Step 6.7: Add the intermediate feature response and the multi-head self-attention representation element by element and send them to the BN normalization layer;

[0021] Step 6.8: The output of the batch normalization layer is processed by layer normalization and channel normalization respectively, and then the results of layer normalization and channel normalization are weighted fused to obtain the encoder output in the encoder-decoder;

[0022] Step 6.9: Feed the encoder output into the decoder in the encoder-decoder. The decoder includes a lightweight uncertainty-aware regression head and a sliding window recursive prediction module, and also includes the following steps:

[0023] Step 6.9.1: Map the global feature vector output by the encoder into a capacity prediction value through a lightweight uncertainty-aware regression head;

[0024] Step 6.9.2: The sliding window recursive prediction module uses a sliding window mechanism based on a recursive prediction strategy to continuously update the initial capacity prediction value until the initial capacity prediction value decays to the set termination threshold, and the battery is determined to have reached the end of life state;

[0025] Further, the multi-head attention splicing and linear mapping module includes a linear transformation layer, a scaled dot product attention operation module and a linear transformation layer. Step 6.5 specifically includes the following steps:

[0026] Step 6.5.1: Use several linear transformation layers to transform the input sparse attention weight map into multiple sets of corresponding Q vectors, K vectors, and V vectors, where each set of Q vectors, K vectors, and V vectors constitutes an attention head;

[0027] Step 6.5.2: Each attention head receives the sparse attention weight as a weighting factor independently, and performs the scaled dot product attention operation in its own feature subspace through the scaled dot product attention operation module to complete the attention aggregation and output the context vector;

[0028] Step 6.5.3: Concatenate the context vectors output by each attention head according to the channel dimension and integrate them through a shared linear transformation layer to output the final multi-head self-attention representation.

[0029] Further: The lightweight uncertainty-aware regression head includes two parallel linear mapping sub-branches, which are used to predict the capacity value at the next moment respectively. and the logarithmic variance corresponding to the capacity value in, , ; and are the regression weight matrix and the uncertainty prediction weight matrix respectively, and are the regression bias term and the uncertainty bias term, respectively.

[0030] Furthermore, one branch of the dual-path feature normalization module includes a parallel convolution layer, a ReLU activation function, a BN normalization layer, a Sigmoid activation function layer, and a parallel convolution layer; the other branch includes a parallel convolution layer, a ReLU activation function, an IN normalization layer, a Sigmoid activation function layer, and a parallel convolution layer. The outputs of the two branches are concatenated along the channel dimension and serve as the output of the dual-path feature normalization module.

[0031] Furthermore, the multi-local perception path module includes three parallel convolution branches distinguished by the size of the convolution kernel. The input features are sequentially passed through the first parallel convolution layer, the ReLU activation function and the BN normalization layer in the parallel convolution branch, and then fused with the normalized features and Sigmoid activation is performed; the result of the Sigmoid activation is passed through the second parallel convolution layer as the output of the parallel convolution branch; the outputs of the three parallel convolution branches are spliced ​​in the channel dimension and used as the output of the multi-local perception path module.

[0032] Furthermore, the lightweight residual connection module includes global average pooling, fully connected layer, ReLU activation function, fully connected layer and Sigmoid activation function layer. Among them, the full time series average response of the addition result is first extracted by global average pooling, and then the fully connected layer is used to model the nonlinear interaction between health indicators. Finally, it is multiplied with the addition result after ReLU activation function, fully connected layer and Sigmoid activation in sequence, and the multiplication result is used as the multi-scale feature.

[0033] Further: label entropy Defined as:

[0034]

[0035] in, is the length of the voltage sequence, is the pattern dimension and the window length of the time series used to calculate the label entropy; is the similarity tolerance threshold; is the approximate entropy, Number of pattern matches The average value for all values ​​of i;

[0036]

[0037]

[0038]

[0039] in, belong , the number of pattern matches is approximate Ratio to the number normalized to the number of individual pattern matches.

[0040] Further: Approximate quantity To find the vector and others vectors After the distance between them, count the ones that meet the conditions < The number of vectors, belong ,and ≠ .

[0041] The beneficial effects of the present invention are as follows: the Transformer-based time series modeling method can capture the long-range dependencies of battery capacity sequences through the self-attention mechanism, overcome the gradient vanishing problem of recurrent neural networks in long sequences, and significantly improve the prediction accuracy; at the same time, feature engineering optimization enhances the model's ability to characterize aging characteristics by quantifying the complexity changes of the voltage sequence, reducing dependence on high-cost measurement equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flow chart of the present invention;

[0043] Figure 2 It is a structural block diagram of the present invention;

[0044] Figure 3 It is a structural block diagram of the encoder in the codec;

[0045] Figure 4 This is the structural block diagram of the decoder in the encoder-decoder. DETAILED DESCRIPTION

[0046] The present invention will be described in detail below with reference to the accompanying drawings. The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention. The directional terms such as left, center, right, top, and bottom in the embodiments of the present invention are merely relative concepts or are based on the normal use state of the product, and should not be considered as restrictive.

[0047] A method for predicting the remaining life of sodium batteries for underwater robot applications includes the following steps:

[0048] Step 1: Build and obtain a trained sodium battery remaining life prediction network, which includes a dual-path feature normalization module, a multi-local perception path module, a lightweight residual connection module, and an encoder-decoder;

[0049] Step 2: Obtain sodium battery data and extract the label entropy of voltage data in the sodium battery data as the input feature of the sodium battery remaining life prediction network; in the voltage data, for a voltage sequence of length N , which is formed by window length segmentation and reconstruction indivual dimensional vector, The vectors are:

[0050]

[0051] It can be called the pattern dimension, which is the window length of the time series used to calculate the label entropy; The value is 2; each sequence is equal to all The distance between sequences is defined as the maximum absolute difference between corresponding scalar elements:

[0052]

[0053] Find the vector and others vectors ( ≠ ) and count the distances between them that meet the conditions < The number of vectors is denoted as ,in, is the similarity tolerance threshold; when When the value of is large, more information will be lost, and when When the value of is small, the statistical characteristics of the system cannot be estimated ideally; the same as the approximate entropy, the present invention will The value of is 0.2SD(x), where SD(x) is the standard deviation of the voltage data;

[0054] if If the corresponding endpoints are within the similarity tolerance threshold, then the eigenvector and The pattern is Lower approximation; the ratio of the approximate number to the number normalized for a single pattern match is recorded as the pattern match number. At this time, the total number of pattern matches is :

[0055]

[0056] Find the average number of pattern matches for all values ​​of i:

[0057]

[0058] Increasing the dimension by 1 becomes , repeat the above steps to obtain the approximate entropy;

[0059]

[0060] Then the label entropy is defined as:

[0061]

[0062] The sample entropy developed in this paper is a non-parametric, noise-resistant, and working condition-independent feature. It does not require complex models or high-cost equipment, is compatible with dynamic and incomplete data, and is highly adaptable to cutting-edge technologies such as machine learning and quantum computing. It is particularly suitable for battery RUL prediction of robots in complex underwater scenarios. Label entropy is an independent feature. Voltage data is input, and the voltage data label entropy is extracted as a feature input into the SOH estimation network.

[0063] Step 3: After the input features pass through the dual-path feature normalization module, normalized features are obtained. One branch of the dual-path feature normalization module includes a parallel convolution layer, a ReLU activation function, a BN normalization layer, a Sigmoid activation function layer, and a parallel convolution layer. The other branch includes a parallel convolution layer, a ReLU activation function, an IN normalization layer, a Sigmoid activation function layer, and a parallel convolution layer. The outputs of the two branches are concatenated along the channel dimension and serve as the output of the dual-path feature normalization module.

[0064] Step 4: The input features are processed by the multi-local perception path module and fused and spliced ​​with the normalized features to obtain spliced ​​features. The multi-local perception path module includes three parallel convolution branches distinguished by the size of the convolution kernel. The input features are sequentially passed through the first parallel convolution layer, the ReLU activation function, and the BN normalization layer in the parallel convolution branch, and then fused with the normalized features and Sigmoid activation is performed. The result of the Sigmoid activation is passed through the second parallel convolution layer as the output of the parallel convolution branch. The outputs of the three parallel convolution branches are spliced ​​in the channel dimension and used as the output of the multi-local perception path module.

[0065] Step 5: After the input features undergo a 1×1 convolution operation, they are element-wise added to the concatenated features. The added results are processed by a lightweight residual connection module to obtain multi-scale features. The lightweight residual connection module includes global average pooling, a fully connected layer, a ReLU activation function, a fully connected layer, and a Sigmoid activation function layer. First, global average pooling is used to extract the full temporal average response of the added result. Then, a fully connected layer is used to model the nonlinear interaction between health indicators. Finally, the ReLU activation function, the fully connected layer, and the Sigmoid activation are sequentially applied and multiplied with the added result. The multiplication result is used as the multi-scale feature.

[0066] Step 6: The encoder-decoder predicts an initial capacity prediction value based on features rich in global temporal information in the multi-scale feature data. Before inputting the multi-scale feature data into the encoder-decoder, a learnable positional encoding is added to the multi-scale feature data, and the multi-scale feature data is normalized.

[0067] The encoder in the encoder-decoder includes a linear transformation module, a scaled dot product calculation module, a gated sparse mask generation module, a weighted calculation module, a multi-head attention splicing and linear mapping module, a first hybrid normalization module, a sparse feedforward network, and a second hybrid normalization module; step 6 is specifically as follows:

[0068] Step 6.1: Use the linear transformation module to transform the multi-scale feature data into query matrix, key matrix, and value matrix respectively;

[0069] Step 6.2: Calculate the attention weight of each moment to the remaining time steps through the scaled dot product calculation module to obtain the original weighted relationship matrix;

[0070] Step 6.3: The original weighted relationship matrix is ​​introduced into the gated sparse mask generation module, and the gated score matrix is ​​calculated based on the matching relationship between the query matrix and the key matrix. The gated sparse mask generation module implements the above functions through a lightweight feedforward network and a sigmoid activation function.

[0071] Step 6.4: The original weighted relationship matrix and the gated score matrix are weighted and output through the weighted calculation module to obtain the sparse attention weight;

[0072] Step 6.5: Introduce the sparse attention weights into the multi-head attention splicing and linear mapping module. The multi-head attention splicing and linear mapping module includes a linear transformation layer, a scaled dot product attention operation module, and a linear transformation layer. The specific steps are as follows:

[0073] Step 6.5.1: Use several linear transformation layers to transform the input sparse attention weights into multiple sets of corresponding Q vectors, K vectors, and V vectors. Each set of Q vectors, K vectors, and V vectors constitutes an attention head, and the whole group forms a multi-head attention group, which is used to model temporal dependencies in their respective subspaces.

[0074] Step 6.5.2: Each attention head receives the sparse attention weight as a weighting factor independently, and performs the scaled dot product attention operation in its own feature subspace through the scaled dot product attention operation module to achieve feature aggregation and selective enhancement of specific time series segments, complete attention aggregation and output the context vector;

[0075] Step 6.5.3: Concatenate the context vectors output by each attention head according to the channel dimension and integrate them through a shared linear transformation layer to output the final multi-head self-attention representation;

[0076] Step 6.6: The multi-head self-attention representation is sequentially passed through the first linear layer, ReLU activation, and another linear layer to obtain the intermediate feature response;

[0077] Step 6.7: Add the intermediate feature response and the multi-head self-attention representation element by element and send them to the BN normalization layer;

[0078] Step 6.8: The output of the batch normalization layer is processed by layer normalization and channel normalization respectively, and then the results of layer normalization and channel normalization are weighted fused to obtain the encoder output in the encoder-decoder;

[0079] Step 6.9: Feed the encoder output into the decoder in the encoder-decoder. The decoder includes a lightweight uncertainty-aware regression head and a sliding window recursive prediction module, and also includes the following steps:

[0080] Step 6.9.1: Map the global feature vector output by the encoder to the initial capacity prediction value through a lightweight uncertainty-aware regression head; the lightweight uncertainty-aware regression head includes two parallel linear mapping sub-branches, which are used to predict the capacity value at the next moment respectively. and the logarithmic variance corresponding to the capacity value , to achieve the joint output of capacity point estimation and uncertainty quantification; among them, , ; and are the regression weight matrix and the uncertainty prediction weight matrix respectively, and They are the regression bias term and the uncertainty bias term, respectively. The regression weight matrix and the uncertainty prediction weight matrix are randomly assigned values ​​using the standard parameter initialization method during the model initialization phase. During training, they are continuously optimized using the negative log-likelihood loss function based on the true capacity label and prediction error. The regression bias term and the uncertainty bias term are also initialized to zero or small random values ​​and are optimized synchronously with the weight matrix during training. The role of this linear mapping sub-branch is to map the multi-dimensional feature expression into a single capacity prediction value, realizing a local to global numerical regression process.

[0081] Step 6.9.2: The sliding window recursive prediction module uses a sliding window mechanism based on a recursive prediction strategy to continuously update the initial capacity prediction value until the initial capacity prediction value decays to the set termination threshold, and the battery is determined to have reached the end of life state;

[0082] This improved encoder structure innovatively introduces a gated sparse attention mechanism to effectively filter key temporal information, improving the efficiency and accuracy of long-sequence global dependency modeling. Furthermore, a hybrid normalization strategy is employed to enhance the stability and generalization of deep feature representation. This improved decoder structure uses a lightweight uncertainty-aware regression head to achieve joint output of capacity prediction and confidence intervals, and continuously tracks the battery capacity degradation process through sliding window recursive prediction. This overall design significantly improves the accuracy and reliability of remaining life prediction for sodium batteries in underwater robots.

[0083] Step 7: Add the initial capacity prediction value and the normalized feature element by element to obtain the correction value. According to the time step and sampling interval of the correction value, calculate and output the remaining service life of the battery from the current moment, that is, the final remaining service life prediction value of the sodium battery.

[0084] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A sodium battery remaining life prediction method for underwater robot applications, characterized by: The following steps are involved: Step 1: Build and obtain a trained sodium battery remaining life prediction network, which includes a dual-path feature normalization module, a multi-local perception path module, a lightweight residual connection module, and an encoder-decoder; Step 2: Obtain sodium battery data and extract the label entropy of voltage data in the sodium battery data as the input feature of the sodium battery remaining life prediction network; Step 3: After the input features pass through the dual-path feature normalization module, normalized features are obtained; Step 4: Process the input features through the multi-local perception path module and fuse and splice them with the normalized features to obtain the spliced ​​features; Step 5: The input features are convolved with the concatenated features and then added element-by-element. The addition result is processed by a lightweight residual connection module to obtain multi-scale features. Step 6: The encoder-decoder predicts an initial capacity prediction value based on the features rich in global temporal information in the multi-scale feature data; Step 7: Add the initial capacity prediction value and the normalized feature element by element to obtain the correction value. According to the time step and sampling interval of the correction value, calculate and output the remaining service life of the battery from the current moment, that is, the final remaining service life prediction value of the sodium battery.

2. The method for predicting the remaining life of a sodium battery for underwater robot application scenarios according to claim 1 is characterized in that: The encoder in the encoder-decoder includes a linear transformation module, a scaled dot product calculation module, a gated sparse mask generation module, a weighted calculation module, a multi-head attention splicing and linear mapping module, a first hybrid normalization module, a sparse feedforward network, and a second hybrid normalization module; step 6 is specifically: Step 6.1: Use the linear transformation module to transform the multi-scale feature data into query matrix, key matrix, and value matrix respectively; Step 6.2: Calculate the attention weight of each moment to the remaining time steps through the scaled dot product calculation module to obtain the original weighted relationship matrix; Step 6.3: The original weighted relationship matrix is ​​introduced into the gated sparse mask generation module, and the gated score matrix is ​​calculated based on the matching relationship between the query matrix and the key matrix; Step 6.4: The original weighted relationship matrix and the gated score matrix are weighted and output through the weighted calculation module to obtain the sparse attention weight; Step 6.5: Introduce the sparse attention weights into the multi-head attention concatenation and linear mapping module to output the multi-head self-attention representation; Step 6.6: The multi-head self-attention representation is sequentially passed through the first linear layer, ReLU activation, and another linear layer to obtain the intermediate feature response; Step 6.7: Add the intermediate feature response and the multi-head self-attention representation element by element and send them to the BN normalization layer; Step 6.8: The output of the batch normalization layer is processed by layer normalization and channel normalization respectively, and then the results of layer normalization and channel normalization are weighted fused to obtain the encoder output in the encoder-decoder; Step 6.9: Feed the encoder output into the decoder in the encoder-decoder. The decoder includes a lightweight uncertainty-aware regression head and a sliding window recursive prediction module, and also includes the following steps: Step 6.9.1: Map the global feature vector output by the encoder into a capacity prediction value through a lightweight uncertainty-aware regression head; Step 6.9.2: The sliding window recursive prediction module uses a sliding window mechanism based on a recursive prediction strategy to continuously update the initial capacity prediction value until the initial capacity prediction value decays to the set termination threshold, and determines that the battery has reached the end of life state.

3. The method for predicting the remaining life of a sodium battery for underwater robot applications according to claim 2 is characterized by: The multi-head attention splicing and linear mapping module includes a linear transformation layer, a scaled dot product attention operation module, and a linear transformation layer. Step 6.5 specifically includes the following steps: Step 6.5.1: Use several linear transformation layers to transform the input sparse attention weight map into multiple sets of corresponding Q vectors, K vectors, and V vectors, where each set of Q vectors, K vectors, and V vectors constitutes an attention head; Step 6.5.2: Each attention head receives the sparse attention weight as a weighting factor independently, and performs the scaled dot product attention operation in its own feature subspace through the scaled dot product attention operation module to complete the attention aggregation and output the context vector; Step 6.5.3: Concatenate the context vectors output by each attention head according to the channel dimension and integrate them through a shared linear transformation layer to output the final multi-head self-attention representation.

4. The method for predicting the remaining life of a sodium battery for underwater robot applications according to claim 2, characterized in that: The lightweight uncertainty-aware regression head consists of two parallel linear mapping sub-branches, which are used to predict the capacity value at the next moment. and the logarithmic variance corresponding to the capacity value in, , ; and are the regression weight matrix and the uncertainty prediction weight matrix respectively, and are the regression bias term and the uncertainty bias term, respectively.

5. The method for predicting the remaining life of a sodium battery for underwater robot applications according to claim 1 is characterized in that: One branch of the dual-path feature normalization module includes a parallel convolution layer, a ReLU activation function, a BN normalization layer, a Sigmoid activation function layer, and a parallel convolution layer. The other branch includes a parallel convolution layer, a ReLU activation function, an IN normalization layer, a Sigmoid activation function layer, and a parallel convolution layer. The outputs of the two branches are concatenated along the channel dimension and used as the output of the dual-path feature normalization module.

6. The method for predicting the remaining life of a sodium battery for underwater robot applications according to claim 1 is characterized by: The multi-local perception path module includes three parallel convolution branches distinguished by the size of the convolution kernel. The input features are sequentially passed through the first parallel convolution layer, the ReLU activation function and the BN normalization layer in the parallel convolution branch, and then fused with the normalized features and Sigmoid activation is performed; the result of the Sigmoid activation is passed through the second parallel convolution layer as the output of the parallel convolution branch; the outputs of the three parallel convolution branches are spliced ​​in the channel dimension and used as the output of the multi-local perception path module.

7. The method for predicting the remaining life of a sodium battery for underwater robot applications according to claim 1 is characterized by: The lightweight residual connection module includes global average pooling, fully connected layer, ReLU activation function, fully connected layer and Sigmoid activation function layer. Among them, the full time series average response of the addition result is first extracted through global average pooling, and then the fully connected layer is used to model the nonlinear interaction between health indicators. Finally, it is multiplied with the addition result after ReLU activation function, fully connected layer and Sigmoid activation in sequence, and the multiplication result is used as the multi-scale feature.

8. The method for predicting the remaining life of a sodium battery for underwater robot applications according to claim 1 is characterized by: Tag entropy Defined as: ； in, is the length of the voltage sequence, is the pattern dimension and the window length of the time series used to calculate the label entropy; is the similarity tolerance threshold; is the approximate entropy, Number of pattern matches The average value for all values ​​of i; ； ； ； in, belong , the number of pattern matches is approximate Ratio to the number normalized to the number of individual pattern matches.

9. The method for predicting the remaining life of a sodium battery for underwater robot applications according to claim 8, characterized in that: Approximate quantity To find the vector and others vectors After the distance between them, count the ones that meet the conditions < The number of vectors, belong ,and ≠ .

Images