Electric ship battery health state estimation method based on multi-target crowd-sourcing algorithm and deep learning neural network

Abstract

The invention discloses an electric ship battery health state estimation method based on a multi-target crowd-sourcing algorithm and a deep learning neural network, and the method comprises the steps: generating a battery data set, and reading all data of a battery in the data set; preprocessing the battery data; determining an optimization target: minimizing a root mean square error and a mean absolute error; setting a parameter range of decision variables; the parameter configuration of each particle is used for operating a Transform network model; iteration is carried out through a multi-objective particle swarm optimization (MOPSO) algorithm, and a global optimal solution is updated; the global optimal solution with the lowest RMSE and MAE is used as the optimal parameter configuration of the Transform model; and outputting the battery SOH estimation value. According to the method, the powerful time sequence data processing capability and the long-term dependency relationship capturing capability of the Transformer deep learning neural network are utilized and combined with the MOPSO optimization algorithm, so that the overall neural network model parameters are optimized, the time-related scale dependency relationship of the battery SOH is enhanced, and accurate estimation of the SOH of the pure electric ship power battery is realized.

Description

Technical Field

[0001] The present invention belongs to the field of ship power systems and relates to a technology for estimating the health status of batteries in electric ships, and specifically to a method for estimating the health status of batteries in electric ships based on a multi-objective swarm intelligence algorithm and a deep learning neural network. Background Art

[0002] In recent years, with growing environmental awareness and the intensifying energy crisis, pure electric vessels have gradually gained attention as a clean and efficient mode of water transportation. A core component of a pure electric vessel is its battery system, and the battery's health directly impacts the vessel's operational efficiency and safety. However, batteries gradually age over long-term use, reducing their capacity and performance. Therefore, accurately assessing battery health is crucial to ensuring safe vessel operation and extending battery life.

[0003] A battery's state of health (SOH) refers to the ratio of the amount of electricity a battery can charge or discharge under certain conditions to its nominal capacity. A new battery's SOH is 100%. As battery performance deteriorates with cycling, the SOH decreases. According to IEEE standards, when the SOH drops to 80%, the battery is considered aged and should be replaced promptly to avoid failure. Lithium-ion battery SOH, a key parameter for state estimation and fault detection in battery management systems, is crucial for improving battery life and ensuring system safety.

[0004] Existing battery health status estimation methods mostly rely on simplified physical models or empirical formulas, which often cannot accurately reflect the actual working status of the battery, especially in complex offshore environments. Summary of the Invention

[0005] Purpose of the invention: In order to overcome the shortcomings of the existing technology, especially the problem of inaccurate evaluation under environmental conditions such as insufficient total data and poor data quality, a method for estimating the health status of electric ship batteries based on a multi-objective swarm intelligence algorithm and a deep learning neural network is provided. The powerful time series data processing capability of the Transformer deep learning neural network and its ability to capture long-term dependencies are utilized, and combined with the MOPSO optimization algorithm to optimize the overall neural network model parameters, enhance the scale dependence of the battery SOH on time, improve the overall stability of the model, and realize accurate estimation of the SOH of pure electric ship power batteries.

[0006] Technical solution: To achieve the above objectives, the present invention provides a method for estimating the health status of electric ship batteries based on a multi-objective swarm intelligence algorithm and a deep learning neural network, comprising the following steps:

[0007] S1: Generate a battery data set from the sensor and read all the battery data in the data set;

[0008] S2: Preprocess the battery data and divide it into training set and test set;

[0009] S3: Determine the optimization objective: minimize the root mean square error and mean absolute error, which are two indicators used to evaluate the accuracy of SOH estimation;

[0010] S4: Set the parameter range of the decision variable, and the position of the particle is randomly generated within the defined parameter range;

[0011] S5: Initialize a group of particles, run the Transformer network model using the parameter configuration of each particle, and obtain the SOH estimate and the corresponding RMSE and MAE;

[0012] S6: The process of step S5 is iterated through the multi-objective particle swarm algorithm MOPSO. According to the update formula, the speed and position of each particle are updated, and the particle with the lowest RMSE and MAE among all particles is found to update the global optimal solution.

[0013] S7: After completing all iterations, the global optimal solution with the lowest RMSE and MAE is used as the optimal parameter configuration of the Transformer model;

[0014] S8: Train the Transformer model in step S7 using the training set and the test set;

[0015] S9: Output the battery SOH estimation value through the trained Transformer model.

[0016] Furthermore, the pre-processing in step S2 includes:

[0017] The second property of the normal distribution is used to handle abnormal data, that is, about 95% of the data is within two standard deviations of the mean; the upper threshold th_max and the lower threshold th_min are obtained by adding or subtracting twice the standard deviation of the mean of the data segment, and the formulas are:

[0018]

[0019] Among them, th_max is the upper threshold; th_min is the lower threshold; mean is the mean; σ is the standard deviation;

[0020] The mean is defined as:

[0021]

[0022] Where, mean is the mean; n is the number of data points; is the normalization coefficient of the summation result, ensuring that the average value rather than the sum is calculated; x i is the i-th data point; ∑ represents the sum.

[0023] Furthermore, the decision variables in step S4 are the learning rate, hidden layer dimension, dropout rate and number of layers of the Transformer neural network model; the position of particle i is represented by x i =[lr i ,hd i ,dr i ,nl i ], where lr i represents the learning rate; hd i represents the hidden layer dimension; dr i represents the drop rate; nl i Represents the number of layers in the network model.

[0024] Furthermore, the definition of the SOH estimation value in step S5 is:

[0025]

[0026] Among them, C real is the actual maximum available capacity of the battery; C rated It is the nominal rated capacity of the battery, which is a fixed value;

[0027]

[0028] Among them, y i is the true value; It is an estimate.

[0029] Furthermore, the Transformer network model in step S5 includes an encoder and a decoder, and both the encoder and the decoder include a multi-head attention layer, a feedforward network layer, and a residual connection and normalization layer, and the multi-head attention layer is provided with a self-attention mechanism.

[0030] Furthermore, the self-attention mechanism of the multi-head attention layer is specifically as follows:

[0031] The self-attention mechanism can pay weighted attention to different positions in the sequence, calculate the correlation between each position and other positions, and capture long-range dependencies in the sequence. The main purpose of the self-attention mechanism is to generate a new representation for each element in the sequence, which consists of weighted contextual information of the entire sequence.

[0032] Input sequence X = [x1, x2, ..., x n ], where xi is the embedding representation of the i-th element in the input sequence;

[0033] First, the query, key, and value representations of the element are calculated: the query in the self-attention represents the battery level data to be predicted, the key represents the battery level data at different time points within the data sequence, and the value is weighted to generate the output, that is, the label of the true SOH;

[0034] Three different weight matrices are used:

[0035]

[0036] in, is the weight matrix to be learned; X is the input sequence; Q i , K i 、V i is the query, key, and value after the transformation of the i-th input, i.e., the battery data and its output at the i-th time point;

[0037] Next, to obtain the contextual information of the i-th element in the sequence, the dot product of the query and all keys is calculated, and then the weight score is obtained through the softmax function:

[0038]

[0039] Among them, At is the attention mechanism; Q, K, V are query, key, and value; softmax is the activation function; d k is the dimension of the key vector;

[0040] The core idea of ​​the multi-head attention layer is to allow the model to capture different aspects of the sequence simultaneously in multiple representation subspaces. The operation of the multi-head attention layer is expressed as:

[0041] M(Q,K,V)=C(h1,h2,...,h i )W O

[0042] Among them, M is multi-head attention; Q, K, V are query, key, and value; W O is the weight matrix; C is the operation of connecting vectors; h i refers to the i-th head in the long position, h i =At(Q i ,K i ,V i );

[0043] Finally, all header information is output through a linear transformation.

[0044] Furthermore, a feedforward neural network (FFN) is a fully connected, position-specific network. This means that while the same feedforward neural network shares parameters across all positions, it performs separate operations for each position in the sequence. Each Transformer encoder and decoder block contains one such layer.

[0045] The feedforward network layer contains two layers of linear transformations, with a nonlinear activation function between the two layers, called the rectified linear unit RLU, specifically:

[0046] FFN(y)=W2·RLU(W1·y+b1)+b2

[0047] Among them, FFN(y) is the final output of the feedforward neural network; z1 is the vector after the first linear transformation; y is the data input to the feedforward neural network; W1 is the weight matrix of this layer, which controls the input battery power data and maps it to a higher-dimensional feature space; b1 and b2 are bias terms; W2 is the weight matrix of the second layer; z2 is the result obtained after applying the RLU function to z1.

[0048] The feedforward network layer determines the model's ability to convert input features in battery SOH estimation, enabling the model to better learn battery power information, extract implicit information such as the power change pattern in the sequence, and provide SOH estimation results that are more stable and close to the true value.

[0049] Furthermore, in the residual connection and normalization layer:

[0050] Residual connections are designed to avoid vanishing or exploding gradients in deep networks. The output of each sublayer in the Transformer (such as a self-attention or feedforward neural network) is summed with the input entering that sublayer, forming a "short-circuit" connection. This allows gradients to flow directly back to the original input during training, accelerating convergence and improving model training stability. Normalization stabilizes network activations and allows for higher learning rates.

[0051] Given input a and sublayer function SubLayer, the residual connection is expressed as:

[0052] b=a+SubLayer(a)

[0053] Among them, a is the input; b is the output; SubLayer(a) is the sublayer function;

[0054] For the output b obtained from the residual connection, the normalization is expressed as:

[0055]

[0056] Among them, μ b is the mean of vector b; is the variance of the vector b; γ and β are learnable scaling and offset parameters; ò is a small constant to prevent division by zero;

[0057] Residual connection and normalization layer formula:

[0058] c=Norm(a+SubLayer(a))

[0059] Among them, c is the output of residual connection and normalization; Norm is layer normalization; SubLayer is the sublayer operation applied to input a; a is the previous output.

[0060] Furthermore, the Multi-Objective Particle Swarm Optimization (MOPSO) algorithm in step S6 is an intelligent optimization algorithm based on the Particle Swarm Optimization (PSO) algorithm. MOPSO searches the solution space by simulating the social behavior of bird flocks or fish schools. Each particle represents a potential solution and updates its position by following the optimal solution.

[0061] In MOPSO, each particle has two properties: position and velocity, and moves in the solution space to find the optimal solution. Each particle updates its velocity and position based on its own experience (individual optimal solution) and the experience of the group (global optimal solution).

[0062] The MOPSO velocity and position update formulas are as follows:

[0063]

[0064] in, is the velocity of particle i at the next moment t+1; w is the inertia weight; c1 and c2 are learning factors; r1 and r2 are random numbers in the range [0,1]; pbest i is the best individual position found so far by particle i; gbest is the global best position found in the current group; is the position of particle i at the current time t; is the new position of particle i at the next moment t+1.

[0065] Furthermore, in step S6, MOPSO updates the velocity and position of the particle through iteration, aiming to find the decision variable combination that minimizes RMSE and MAE. In each iteration, the particle is updated according to its own pbest iAnd the gbest of the group is used to update the position, and the Transformer network model is run to evaluate the RMSE and MAE values ​​of the new position. If particle i is at the new position x i The RMSE and MAE values ​​are better than its historical best position pbest i If the value of is good, update pbest i The new x for the current position i ; Among the new positions of all particles in the group, the position with the best RMSE and MAE values ​​is selected as the new global optimal position gbest. In the iterative process, the above steps are repeated continuously, and each particle is updated according to the pbest. i and gbest adjust their own positions, while continuously evaluating and adjusting these positions based on the objective function values ​​returned by the Transformer model.

[0066] In view of the above, the present invention has the following significant features:

[0067] 1. A multi-objective particle swarm algorithm automatically optimizes the parameters and structure of the Transformer network to find the model configuration that best suits the battery health status. This ensures that the model can run efficiently even when processing complex data, reducing model estimation errors and improving computational efficiency.

[0068] 2. By more accurately predicting and monitoring the health status of batteries, it helps improve the energy management strategy of pure electric ships, extend battery life, reduce unexpected failures, and improve the overall efficiency, safety and reliability of the ship's energy system.

[0069] 3. Under the conditions of insufficient total data and poor quality, the multi-objective swarm intelligence algorithm and deep learning neural network method can enhance the robustness of the model, enhance its adaptability to data quantity and quality restrictions, and improve the accuracy of the model in estimating the battery health status under restricted conditions.

[0070] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0071] 1. The solution of the present invention innovatively adopts the MOPSO optimization algorithm to explore the optimal network architecture and hyperparameter configuration, significantly improving the performance of the Transformer network model. On the basis of capturing long-term dependencies, it improves the accuracy of the model in estimating battery SOH and its computational efficiency.

[0072] 2. The solution of the present invention takes the minimization of RMSE and MAE dual errors as the optimization goal, which can maintain the accuracy of the network model in estimating SOH when the total amount of input data is insufficient or the quality is poor, maintain the stability of the output results, and increase the reliability and robustness of the model.

[0073] 3. The solution of the present invention combines the optimization algorithm and the neural network model to accurately estimate the battery SOH. It can adapt to different types of batteries and usage patterns, increase the generalization ability of the model, and can be extended to multiple types of ships, extending battery life and reducing maintenance costs, thereby improving the safety and reliability of ship systems.

[0074] 4. The solution of the present invention introduces an optimization algorithm based on the Transformer neural network model to help the model be more adaptable and flexible when facing complex data sets, dynamically adjust the learning process, reduce the model's dependence on specific data features, and reduce the risk of overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 is a flow chart of the method of the present invention;

[0076] Figure 2 This is the Transformer deep learning neural network model diagram;

[0077] Figure 3 The figure shows the estimated and true values ​​of battery SOH. DETAILED DESCRIPTION

[0078] The present invention is further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0079] Example 1:

[0080] like Figure 1 As shown, this embodiment provides a method for estimating the health status of an electric ship battery based on a multi-objective swarm intelligence algorithm and a deep learning neural network, comprising the following steps:

[0081] S1: Generate a battery data set from the sensor, read all the battery data in the data set, and read the battery data such as cycle stage, step stage, number of cycles, test time, capacity, etc.;

[0082] S2: Preprocess the battery data and divide it into training set and test set;

[0083] Preprocessing includes: using the second property of normal distribution to handle abnormal data, that is, about 95% of the data is within two standard deviations of the mean; the upper threshold th_max and the lower threshold th_min: are obtained by adding or subtracting twice the standard deviation of the mean of the data segment, and the formulas are:

[0084]

[0085] Among them, th_max is the upper threshold; th_min is the lower threshold; mean is the mean; σ is the standard deviation;

[0086] The mean is defined as:

[0087]

[0088] Where, mean is the mean; n is the number of data points; is the normalization coefficient of the summation result, ensuring that the average value rather than the sum is calculated; x i is the i-th data point; ∑ represents the sum.

[0089] S3: Determine the optimization objective: minimize the root mean square error (RMSE) and the mean absolute error (MAE), which are used to evaluate the accuracy of the SOH estimation.

[0090]

[0091] Among them, y i is the true value; It is an estimate.

[0092] S4: Set the parameter range of the decision variable, and the position of the particle is randomly generated within the defined parameter range;

[0093] The decision variables are the learning rate, hidden layer dimension, dropout rate, and number of layers of the Transformer neural network model; the position of particle i is represented by x i =[lr i ,hd i ,dr i ,nl i ], where lr i represents the learning rate; hd i represents the hidden layer dimension; dr i represents the drop rate; nl i Represents the number of layers in the network model.

[0094] S5: Initialize a group of particles, run the Transformer network model using the parameter configuration of each particle, and obtain the SOH estimate and the corresponding RMSE and MAE;

[0095] The definition of the SOH estimate is:

[0096]

[0097] Among them, C real is the actual maximum available capacity of the battery; C rated It is the nominal rated capacity of the battery, which is a fixed value;

[0098] like Figure 2 As shown in the figure, the Transformer network model includes an encoder and a decoder. Both the encoder and the decoder include a multi-head attention layer, a feedforward network layer, a residual connection and a normalization layer. The multi-head attention layer is equipped with a self-attention mechanism.

[0099] The self-attention mechanism of the multi-head attention layer is specifically:

[0100] The self-attention mechanism can pay weighted attention to different positions in the sequence, calculate the correlation between each position and other positions, and capture long-range dependencies in the sequence. The main purpose of the self-attention mechanism is to generate a new representation for each element in the sequence, which consists of weighted contextual information of the entire sequence.

[0101] Input sequence X = [x1, x2, ..., x n ], where x i is the embedding representation of the i-th element in the input sequence;

[0102] First, the query, key, and value representations of the element are calculated: the query in the self-attention represents the battery level data to be predicted, the key represents the battery level data at different time points within the data sequence, and the value is weighted to generate the output, that is, the label of the true SOH;

[0103] Three different weight matrices are used:

[0104]

[0105] in, is the weight matrix to be learned; X is the input sequence; Q i , K i 、V i is the query, key, and value after the transformation of the i-th input, i.e., the battery data and its output at the i-th time point;

[0106] Next, to obtain the contextual information of the i-th element in the sequence, the dot product of the query and all keys is calculated, and then the weight score is obtained through the softmax function:

[0107]

[0108] Among them, At is the attention mechanism; Q, K, V are query, key, and value; softmax is the activation function; d k is the dimension of the key vector;

[0109] The core idea of ​​multi-head attention in the multi-head attention layer is to allow the model to capture different aspects of the sequence simultaneously in multiple representation subspaces. The operation of the multi-head attention layer is expressed as:

[0110] M(Q,K,V)=C(h1,h2,...,h i )W O

[0111] Among them, M is multi-head attention; Q, K, V are query, key, and value; W O is the weight matrix; C is the operation of connecting vectors; h i refers to the i-th head in the long position, h i =At(Q i ,K i ,V i );

[0112] Finally, all header information is output through a linear transformation.

[0113] A feedforward neural network (FFN) is a fully connected, position-specific network. This means that while the same feedforward neural network shares parameters across all positions, each position in the sequence is operated on independently. Each Transformer's encoder and decoder block contains one such layer.

[0114] The feedforward network layer contains two layers of linear transformations, with a nonlinear activation function between the two layers, called the rectified linear unit RLU, specifically:

[0115] FFN(y)=W2·RLU(W1·y+b1)+b2

[0116] Among them, FFN(y) is the final output of the feedforward neural network; z1 is the vector after the first linear transformation; y is the data input to the feedforward neural network; W1 is the weight matrix of this layer, which controls the input battery power data and maps it to a higher-dimensional feature space; b1 and b2 are bias terms; W2 is the weight matrix of the second layer; z2 is the result obtained after applying the RLU function to z1.

[0117] The feedforward network layer determines the model's ability to convert input features in battery SOH estimation, enabling the model to better learn battery power information, extract implicit information such as the power change pattern in the sequence, and provide SOH estimation results that are more stable and close to the true value.

[0118] In the residual connection and normalization layer:

[0119] Residual connections are designed to avoid vanishing or exploding gradients in deep networks. The output of each sublayer in the Transformer (such as a self-attention or feedforward neural network) is summed with the input entering that sublayer, forming a "short-circuit" connection. This allows gradients to flow directly back to the original input during training, accelerating convergence and improving model training stability. Normalization stabilizes network activations and allows for higher learning rates.

[0120] Given input a and sublayer function SubLayer, the residual connection is expressed as:

[0121] b=a+SubLayer(a)

[0122] Among them, a is the input; b is the output; SubLayer(a) is the sublayer function;

[0123] For the output b obtained from the residual connection, the normalization is expressed as:

[0124]

[0125] Among them, μ b is the mean of vector b; is the variance of the vector b; γ and β are learnable scaling and offset parameters; ò is a small constant to prevent division by zero;

[0126] Residual connection and normalization layer formula:

[0127] c=Norm(a+SubLayer(a))

[0128] Among them, c is the output of residual connection and normalization; Norm is layer normalization; SubLayer is the sublayer operation applied to input a; a is the previous output.

[0129] S6: The process of step S5 is iterated through the multi-objective particle swarm algorithm MOPSO. According to the update formula, the speed and position of each particle are updated, and the particle with the lowest RMSE and MAE among all particles is found to update the global optimal solution.

[0130] The Multi-Objective Particle Swarm Optimization (MOPSO) algorithm is an intelligent optimization algorithm based on the Particle Swarm Optimization (PSO) algorithm. MOPSO searches the solution space by simulating the social behavior of bird flocks or fish schools. Each particle represents a potential solution and updates its position by following the optimal solution.

[0131] In MOPSO, each particle has two properties: position and velocity, and moves in the solution space to find the optimal solution. Each particle updates its velocity and position based on its own experience (individual optimal solution) and the experience of the group (global optimal solution).

[0132] The MOPSO velocity and position update formulas are as follows:

[0133]

[0134] in, is the velocity of particle i at the next moment t+1; w is the inertia weight; c1 and c2 are learning factors; r1 and r2 are random numbers in the range [0,1]; pbest i is the best individual position found so far by particle i; gbest is the global best position found in the current group; is the position of particle i at the current time t; is the new position of particle i at the next moment t+1.

[0135] MOPSO updates the velocity and position of particles through iterations, aiming to find the decision variable combination that minimizes RMSE and MAE. In each iteration, the particle updates the velocity and position of particles according to its own pbest. i And the gbest of the group is used to update the position, and the Transformer network model is run to evaluate the RMSE and MAE values ​​of the new position. If particle i is at the new position x i The RMSE and MAE values ​​are better than its historical best position pbest i If the value of is good, update pbest i The new x for the current position i ; Among the new positions of all particles in the group, the position with the best RMSE and MAE values ​​is selected as the new global optimal position gbest. In the iterative process, the above steps are repeated continuously, and each particle is updated according to the pbest. i and gbest adjust their own positions, while continuously evaluating and adjusting these positions based on the objective function values ​​returned by the Transformer model.

[0136] S7: After completing all iterations, the global optimal solution with the lowest RMSE and MAE is used as the optimal parameter configuration of the Transformer model;

[0137] S8: Train the Transformer model in step S7 using the training set and the test set;

[0138] S9: Output the battery SOH estimation value through the trained Transformer model.

[0139] Example 2:

[0140] In order to verify the effectiveness and effect of the method of the present invention, this example uses the battery data set provided by the Center for Advanced Lifecycle Engineering at the University of Maryland as test data to conduct experimental verification, as follows:

[0141] The main parameters used in the neural network model in this embodiment are: the inertia weight of the multi-objective particle swarm algorithm is 1.5; the individual learning factor is 1.0; the social learning factor is 1.5; there are four decision variables, namely: learning rate, hidden layer dimension, drop rate, and number of model layers; among them, the learning rate boundary is 0.0001 to 0.01, the hidden layer dimension boundary is 1 to 32, the drop rate boundary is 0.0001 to 0.01, and the number of model layers boundary is 2 to 8; the maximum number of iterations is 20.

[0142] The initial number of cycles of the Transformer neural network is 350, the number of multi-head layers is 16, and the input feature dimension is 32.

[0143] After optimization by the MOPSO algorithm, the learning rate of the Transformer neural network is 0.051;

[0144] The hidden layer dimension is 38; the dropout rate is 0.00036; and the number of model layers is 6.

[0145] The performance of the improved model is calculated, including indicators such as mean square error, to evaluate the accuracy of the model's estimation of the test set data. The prediction results are as follows: Figure 3 And obtain the comparative data shown in Table 1.

[0146] Table 1 Comparison of main indicators:

[0147]

[0148] Comparing the above results, the model with the optimization algorithm converges significantly faster and saves more computing power. Compared to models without the optimization algorithm, the multi-objective swarm intelligence algorithm and deep learning neural network achieve higher estimation accuracy and better performance. The mean square error of the network model with the optimization algorithm is 0.0562, a 28.31% reduction compared to the model without optimization. These results demonstrate that the estimation scheme provided by the present invention effectively improves the accuracy of battery SOH assessment.

Claims

1. A method for estimating the health status of electric ship batteries based on a multi-objective swarm intelligence algorithm and a deep learning neural network, characterized in that: The steps include: S1: Generate a battery data set from the sensor and read all the battery data in the data set; S2: Preprocess the battery data and divide it into training set and test set; S3: Determine the optimization objective: minimize the root mean square error and mean absolute error, which are two indicators used to evaluate the accuracy of SOH estimation; S4: Set the parameter range of the decision variable, and the position of the particle is randomly generated within the defined parameter range; S5: Initialize a group of particles, run the Transformer network model using the parameter configuration of each particle, and obtain the SOH estimate and the corresponding RMSE and MAE; S6: The process of step S5 is iterated through the multi-objective particle swarm algorithm MOPSO. According to the update formula, the speed and position of each particle are updated, and the particle with the lowest RMSE and MAE among all particles is found to update the global optimal solution. S7: After completing all iterations, the global optimal solution with the lowest RMSE and MAE is used as the optimal parameter configuration of the Transformer model; S8: Train the Transformer model in step S7 using the training set and the test set; S9: Output the battery SOH estimation value through the trained Transformer model.

2. The method for estimating the health status of electric ship batteries based on a multi-objective swarm intelligence algorithm and a deep learning neural network according to claim 1 is characterized in that: The pre-processing in step S2 includes: The second property of the normal distribution is used to handle abnormal data; the upper threshold th_max and the lower threshold th_min are obtained by adding or subtracting twice the standard deviation of the mean of the data segment, and the formulas are: Among them, th_max is the upper threshold; th_min is the lower threshold; mean is the mean; σ is the standard deviation; The mean is defined as: Where, mean is the mean; n is the number of data points; is the normalization coefficient of the summation result, ensuring that the average value rather than the sum is calculated; x i is the i-th data point; ∑ represents the sum.

3. The method for estimating the health status of an electric ship battery based on a multi-objective swarm intelligence algorithm and a deep learning neural network according to claim 1 is characterized in that: The decision variables in step S4 are the learning rate, hidden layer dimension, dropout rate and number of layers of the Transformer neural network model; the position of particle i is represented by x i =[lr i ,hd i ,dr i ,nl i ], where lr i represents the learning rate; hd i represents the hidden layer dimension; dr i represents the drop rate; nl i Represents the number of layers in the network model.

4. The method for estimating the health status of an electric ship battery based on a multi-objective swarm intelligence algorithm and a deep learning neural network according to claim 3 is characterized in that: The definition of the SOH estimation value in step S5 is: Among them, C real is the actual maximum available capacity of the battery; C rated It is the nominal rated capacity of the battery, which is a fixed value; Among them, y i is the true value; It is an estimate.

5. The method for estimating the health status of electric ship batteries based on a multi-objective swarm intelligence algorithm and a deep learning neural network according to claim 4 is characterized in that: The Transformer network model in step S5 includes an encoder and a decoder, each of which includes a multi-head attention layer, a feedforward network layer, and a residual connection and normalization layer, and the multi-head attention layer is provided with a self-attention mechanism.

6. The method for estimating the health status of electric ship batteries based on a multi-objective swarm intelligence algorithm and a deep learning neural network according to claim 5 is characterized in that: The self-attention mechanism of the multi-head attention layer is specifically: Input sequence X = [x1, x2, ..., x n ], where x i is the embedding representation of the i-th element in the input sequence; First, the query, key, and value representations of the element are calculated: the query in the self-attention represents the battery level data to be predicted, the key represents the battery level data at different time points within the data sequence, and the value is weighted to generate the output, that is, the label of the true SOH; Three different weight matrices are used: in, is the weight matrix to be learned; X is the input sequence; Q i , K i 、V i is the query, key, and value after the transformation of the i-th input, i.e., the battery data and its output at the i-th time point; Next, to obtain the contextual information of the i-th element in the sequence, the dot product of the query and all keys is calculated, and then the weight score is obtained through the softmax function: Among them, At is the attention mechanism; Q, K, V are query, key, and value; softmax is the activation function; d k is the dimension of the key vector; The operation of the multi-head attention layer is expressed as: M(Q,K,V)=C(h1,h2,...,h i )W O Among them, M is multi-head attention; Q, K, V are query, key, and value; W O is the weight matrix; C is the operation of connecting vectors; h i Refers to the i-th head in the long position, h i =At(Q i ,K i ,V i ); Finally, all header information is output through a linear transformation.

7. The method for estimating the health status of electric ship batteries based on a multi-objective swarm intelligence algorithm and a deep learning neural network according to claim 6 is characterized in that: The feedforward network layer contains two layers of linear transformations, with a nonlinear activation function between the two layers, called the rectified linear unit RLU, specifically: FFN(y)=W2·RLU(W1·y+b1)+b2 Among them, FFN(y) is the final output of the feedforward neural network; z1 is the vector after the first linear transformation; y is the data input to the feedforward neural network; W1 is the weight matrix of this layer, which controls the input battery power data and maps it to a higher-dimensional feature space; b1 and b2 are bias terms; W2 is the weight matrix of the second layer; z2 is the result obtained after applying the RLU function to z1.

8. The method for estimating the health status of electric ship batteries based on a multi-objective swarm intelligence algorithm and a deep learning neural network according to claim 7 is characterized in that: The residual connection and normalization layer: Given input a and sublayer function SubLayer, the residual connection is expressed as: b=a+SubLayer(a) Among them, a is the input; b is the output; SubLayer(a) is the sublayer function; For the output b obtained from the residual connection, the normalization is expressed as: Among them, μ b is the mean of vector b; is the variance of the vector b; γ and β are learnable scaling and offset parameters; ò is a constant to prevent division by zero; Residual connection and normalization layer formula: c=Norm(a+SubLayer(a)) Among them, c is the output of residual connection and normalization; Norm is layer normalization; SubLayer is the sublayer operation applied to input a; a is the previous output.

9. The method for estimating the health status of an electric ship battery based on a multi-objective swarm intelligence algorithm and a deep learning neural network according to claim 8, characterized in that: The MOPSO velocity and position update formula in step S6 is as follows: in, is the velocity of particle i at the next moment t+1; w is the inertia weight; c1 and c2 are learning factors; r1 and r2 are random numbers in the range [0,1]; pbest i is the best individual position found so far by particle i; gbest is the global best position found in the current group; is the position of particle i at the current time t; is the new position of particle i at the next moment t+1.

10. The method for estimating the health status of electric ship batteries based on a multi-objective swarm intelligence algorithm and a deep learning neural network according to claim 9, characterized in that: In each iteration of step S6, the particle is calculated based on its own pbest i And the gbest of the group is used to update the position, and the Transformer network model is run to evaluate the RMSE and MAE values ​​of the new position. If particle i is at the new position x i The RMSE and MAE values ​​are better than its historical best position pbest i If the value of is good, update pbest i The new x for the current position i ; Among the new positions of all particles in the group, the position with the best RMSE and MAE values ​​is selected as the new global optimal position gbest. In the iterative process, the above steps are repeated continuously, and each particle is updated according to the pbest. i and gbest adjust their own positions, while continuously evaluating and adjusting these positions based on the objective function values ​​returned by the Transformer model.

Images