A method and device for predicting the health status of electric ship batteries

Abstract

The present application provides a method and device for predicting the health status of batteries in electric ships. The method comprises: collecting multidimensional data of electric ships through a sensor network, covering the power system, environment, and hull status; constructing a sample data set containing this data, and designing an encoder based on multi-head attention, cross convolution, and feature attention aggregation, as well as a decoder based on deconvolution and conditional self-attention mechanisms; using the encoder and decoder to construct an initial prediction model, training it through the sample data set and a preset training configuration to obtain a prediction model for the state of the ship's pure battery power system; deploying the model on the ship to be predicted, and predicting the battery health status based on the battery capacity value, ultimately obtaining a prediction result for the battery health status. The method and device can predict the future health status of the ship's pure battery in real time with high prediction accuracy, providing technical support for the safe and efficient operation of the ship's pure battery power system.

Description

Technical Field

[0001] The present application relates to the field of battery management technology, and more specifically, to a method and device for predicting the health status of batteries in electric ships. Background Art

[0002] Currently, the global demand for green shipping technologies is becoming increasingly urgent, mainly due to increasingly severe environmental problems and the continuous tightening of ship emission regulations. Pure battery power systems have excellent performance in reducing pollution, reducing noise, and improving energy efficiency, becoming an important direction for the transformation and upgrading of the shipping industry. Compared with traditional diesel generators, pure battery systems have a higher level of electrification, and the DC network design optimizes energy transmission and spatial layout. However, its high complexity also increases the probability of equipment failure, and the variety of failures seriously threatens the safety of system operation. In particular, the performance of lithium-ion batteries gradually declines with the charge and discharge cycle, placing higher demands on accurate assessment and real-time monitoring of battery status. Traditional regular maintenance and post-repair strategies lack dynamic assessment capabilities and are difficult to detect potential faults in advance. They can no longer meet the requirements of pure battery power systems for high efficiency, stability, and reliability. Summary of the Invention

[0003] The purpose of the embodiments of the present application is to provide a method and device for predicting the health status of batteries in electric ships, which can predict the future health status of the ship's pure batteries in real time with high prediction accuracy, and provide technical support for the safe and efficient operation of the ship's pure battery power system.

[0004] The first aspect of the present application provides a method for predicting the health status of an electric ship battery, comprising:

[0005] Acquiring multidimensional data of the electric ship collected based on a sensor network; wherein the multidimensional data of the electric ship includes power system data, environmental data and hull status;

[0006] Constructing a sample data set according to the multidimensional data of the electric ship;

[0007] Construct an encoder based on multi-head attention mechanism, cross convolution and feature attention aggregation mechanism, and a decoder based on deconvolution and conditional self-attention mechanism;

[0008] constructing an initial prediction model based on the encoder and the decoder;

[0009] Training the initial prediction model according to the sample data set and preset training configuration data to obtain a trained ship pure battery power system state prediction model;

[0010] Deploying the ship's pure battery power system state prediction model onto the hardware platform of the ship to be predicted;

[0011] Obtaining the battery capacity value of the ship to be predicted according to a preset output window size;

[0012] A battery health state prediction is performed based on the battery capacity value and the ship pure battery power system state prediction model to obtain a battery health state prediction result.

[0013] The beneficial effects of this application are: (1) through the combination of multi-head self-attention mechanism and cross convolutional network, the model can capture the dynamic changes of the battery system (including fluctuations of important parameters such as voltage, current, temperature, etc.) from both short-term and long-term perspectives, thereby ensuring that the model can fully reflect the operating status of the ship's pure battery power system and achieve accurate identification of important abnormal characteristic changes, thereby enabling earlier detection of potential faults in the ship's pure battery power system, reducing the risk of sudden shutdown, and enhancing the safety of the system.

[0014] (2) The feature attention aggregation mechanism enables the model to assign adaptive weights to multi-scale features, focusing on features with greater influence in the prediction process, thereby reducing noise interference and improving the model's sensitivity to state changes. At the same time, the FAA mechanism makes the model more robust under different working conditions and environmental changes, thereby being able to identify key state parameters (such as current peaks or temperature anomalies) in pure battery power systems, thereby providing timely warnings for ship operations and improving the accuracy of fault detection.

[0015] (3) The model’s multi-dimensional prediction of the key operating parameters of the battery (including voltage, current, temperature and other indicators) can not only show the current system health status, but also infer the future performance of the battery, thereby helping ship operators to manage the battery rationally throughout its life cycle (including planning charge and discharge cycles, predicting remaining mileage, judging maintenance timing, etc.), and thus optimizing the battery’s efficiency and life. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 A flow chart of a method for predicting the health status of an electric ship battery provided in an embodiment of the present application;

[0018] Figure 2 This is a schematic diagram of an example process framework of a method for predicting the health status of an electric ship battery provided in an embodiment of the present application;

[0019] Figure 3 A schematic diagram of a portion of the flow chart of another method for predicting the health status of batteries in electric ships provided in an embodiment of the present application;

[0020] Figure 4 A schematic diagram of another portion of the flow chart of another method for predicting the health status of batteries in electric ships provided in an embodiment of the present application;

[0021] Figure 5 A framework for constructing a sample data set provided in an embodiment of the present application;

[0022] Figure 6 This is a schematic diagram of an example flow chart of a method for predicting the health status of an electric ship battery provided in an embodiment of the present application;

[0023] Figure 7 A comparison chart of the battery health status prediction simulation of an electric ship provided in an embodiment of the present application;

[0024] Figure 8 A schematic diagram of the structure of a device for predicting the health status of batteries of electric ships provided in an embodiment of the present application;

[0025] Figure 9 A schematic structural diagram of another electric ship battery health status prediction device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0027] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0028] Example 1

[0029] Please see Figure 1 , Figure 1 This is a flow chart of a method for predicting the health status of batteries in electric ships provided in this embodiment. The method for predicting the health status of batteries in electric ships includes:

[0030] S101. Acquire multi-dimensional data of the electric ship collected based on a sensor network; wherein the multi-dimensional data of the electric ship includes power system data, environmental data and hull status.

[0031] In this embodiment, the power system data includes at least rated voltage, rated current, rated power and power system ambient temperature;

[0032] Environmental data include at least water flow rate, ambient temperature and relative humidity;

[0033] The hull status includes at least the bearing temperature and the ship speed;

[0034] The sensor network includes battery pack sensors and environmental sensors;

[0035] Battery pack sensors are used to collect power system data;

[0036] The acquisition frequency of the battery pack sensor is the first preset acquisition frequency;

[0037] Environmental sensors include humidity sensors, flow rate sensors, temperature sensors and GPS sensors. Environmental sensors are used to collect environmental data and hull status;

[0038] The collection frequency of the environmental sensor is the second preset collection frequency.

[0039] S102: Construct a sample data set based on the multi-dimensional data of electric ships.

[0040] S103. Construct an encoder based on multi-head attention mechanism, cross convolution and feature attention aggregation mechanism, and construct a decoder based on deconvolution and conditional self-attention mechanism.

[0041] In this embodiment, the core idea of the multi-head attention mechanism is to use multiple attention heads working in parallel to capture the relationships between key time steps in the data from different perspectives. In a pure battery-powered marine system, battery status and system operation are often influenced by multiple factors and have complex temporal dependencies. The multi-head attention mechanism helps the model identify which time steps and features are most important for predicting system status.

[0042] In this example, a cross-convolutional neural network (CCNN) extracts local features of input data using multi-scale convolution kernels and captures the spatial and temporal dependencies of the data through feature fusion. For state prediction of a ship's pure battery power system, the CCNN can help extract multi-scale features of the battery system, improving the model's ability to perceive complex dynamic data.

[0043] S104: Build an initial prediction model based on the encoder and decoder.

[0044] S105 . Train the initial prediction model according to the sample data set and the preset training configuration data to obtain a trained ship pure battery power system state prediction model.

[0045] S106. Deploy the ship's pure battery power system state prediction model on the hardware platform of the ship to be predicted.

[0046] S107. Obtain the battery capacity value of the ship to be predicted according to a preset output window size.

[0047] S108 , predicting the battery health state according to the battery capacity value and the ship pure battery power system state prediction model to obtain a battery health state prediction result.

[0048] Please see Figure 2 , Figure 2 A schematic diagram of an example process framework for predicting the health status of batteries in electric ships is shown. The diagram indicates that the ship's pure battery power system health prediction method is trained using a cross neural network model with a multi-head attention mechanism and a deconvolution conditional self-attention mechanism model using a training dataset.

[0049] In this embodiment, the execution subject of the method may be a computing device such as a computer or a server, and this is not limited in this embodiment.

[0050] In this embodiment, the execution subject of the method may also be a smart device such as a smart phone, a tablet computer, etc., which is not limited in this embodiment.

[0051] It can be seen that the implementation of the electric ship battery health status prediction method described in this embodiment can predict the future health status of the ship's pure battery in real time with high prediction accuracy, providing technical support for the safe and efficient operation of the ship's pure battery power system.

[0052] Example 2

[0053] Please see Figure 3 and Figure 4 , Figure 3 This is a partial flow chart of a method for predicting the health status of batteries in electric ships provided in this embodiment. Figure 4 This is another partial flow chart of a method for predicting the health status of batteries in electric ships provided in this embodiment. The method for predicting the health status of batteries in electric ships includes:

[0054] S201. Acquire multi-dimensional data of the electric ship collected based on a sensor network; wherein the multi-dimensional data of the electric ship includes power system data, environmental data and hull status.

[0055] In this embodiment, the method collects multidimensional data through a sensor network, including real-time collection of multidimensional data of the ship's power system's rated voltage, rated current, rated power, and power system ambient temperature, totaling 4 data points, and integrates environmental data (water flow rate, ambient temperature, relative humidity) and external factors of the hull state (bearing temperature, ship speed), totaling 5 data points, and thereby constructs a comprehensive data set containing internal states and external influencing factors.

[0056] In this embodiment, the sensor network is composed of battery pack sensors and environmental sensors.

[0057] In this embodiment, the battery pack sensor is used to collect the rated voltage V of the ship power system. b (t): 660V, rated current I b (t), system ambient temperature T b (t): 0~50 degrees, rated power P b (t): Information about the change in 1000kW. The acquisition frequency is once per second (1Hz). A total of 345,600 data points are acquired.

[0058] In this embodiment, the environmental sensor is composed of a humidity sensor, a flow rate sensor, a temperature sensor and a GPS sensor, which is used to collect the relative humidity H(t) and the water flow rate S in the environment. water (t), ambient temperature T env (t), bearing temperature T z (t) and ship speed v b (t) Data. The sampling frequency is once every five seconds, 0.2 Hz. A total of 5 data points are collected, resulting in approximately 51,840 data points.

[0059] S202: Perform time synchronization processing on the multi-dimensional data using a time interpolation algorithm to obtain synchronized data.

[0060] In this embodiment, since multi-source data needs to ensure that each data source is sampled on the same time axis, and the sampling frequencies of each data source in the ship's pure battery power system are different, this method uses time interpolation to achieve time alignment. The time synchronization formula is as follows:

[0061] Given sensor S i The sampling time is t (i，1) , t (i，2) ,…,t (i，n) , the corresponding data value is x (i，1) , x (i，2) ,…,x (i，n) The sampling points are aligned to the unified time axis T = {t1, t2, ..., t m}, the time interpolation formula is:

[0062]

[0063] in, Indicates sensor S i At the same time t j After all sensors are interpolated, a unified time series dataset is formed. i represents the sensor number, and j represents the value at time t.

[0064] S203: Using a weighted average algorithm to perform data fusion processing on the synchronized data to obtain a preliminary feature matrix.

[0065] In this embodiment, there are multiple sensors for the physical quantity of the same data tag. This method fuses the data of each sensor through a weighted average method to reduce the influence of the error of a single sensor.

[0066] Among them, the dataset obtained after time synchronization At time t j In the following example, taking temperature as an example, there are N sensors providing data x1(t j ), x2(t j ),…,x N (t j ). Through weighted averaging, the fused value is:

[0067]

[0068] Among them, w i is the weight, The obtained fusion data vector X(t j ) contains the physical quantities at time t j The representative values of are used to form a preliminary feature matrix.

[0069] S204: Perform vectorization processing on the preliminary feature matrix to obtain a multi-dimensional data vector.

[0070] In this embodiment, the fused data is j It is represented as a multidimensional data vector X(t j ), including various physical quantities:

[0071] X(t j )=[V b (t), I b (t), T b (t), P b (t), H(t), S water (t), T env (t), T z (t), v b (t)] T

[0072] Multidimensional vector X(t j ) indicates that at time t j A collection of systems. o As the final representation of the data acquisition part.

[0073] This implementation enables data fusion of internal and external factor datasets acquired by the sensor network, integrating multidimensional data (including battery parameters, motor status, and external environmental factors) into a global dataset, thereby facilitating the capture of overall system status information. The data fusion process includes time synchronization, weighted average-based data fusion, and multidimensional data vectorization.

[0074] S205. Calculate the mean and standard deviation of each data point in the multidimensional data vector.

[0075] In this embodiment, in the ship operation environment, the sensor data may be interfered by the external environment or sensor failure, resulting in abnormal values. Based on this, the method uses the three sigma method to detect abnormal values, setting the original data set X o =[x1, x2, ..., x n ], where each data point x i It is a status value collected by the system.

[0076] In this embodiment, the calculation formulas for the mean value μ and the standard deviation σ are as follows:

[0077]

[0078] S206 , performing outlier determination on each data point in the multidimensional data vector according to the mean value, standard deviation, and preset outlier determination conditions to obtain an outlier determination result.

[0079] In this embodiment, when the data point satisfies |x i When -μ|>3σ, the point is considered an outlier and is removed.

[0080] S207 , performing outlier elimination processing on the multidimensional data vector according to the outlier determination result to obtain eliminated data.

[0081] S208 : Decompose the first processed data to obtain approximation coefficients and detail coefficients of each layer.

[0082] In this embodiment, the method uses wavelet transform noise reduction to effectively separate signals of different frequencies, thereby removing noise in the operation of the ship power system. Specifically, the method performs noise reduction on the original signal X o Perform discrete wavelet transform to decompose detail coefficients and approximation coefficients of different frequency bands.

[0083] Specifically, X o Decompose to J layers and get the approximation coefficient a J and the detail coefficient d of each layer j (j=1, 2, ..., J).

[0084]

[0085] S209 , performing soft threshold filtering on the detail coefficients to obtain filtered target detail coefficients.

[0086] In this embodiment, the method performs the detail coefficient d j The soft threshold filtering method is used. The detail coefficients and approximation coefficients after noise reduction are reconstructed into the cleaned data X by inverse wavelet transform. c .

[0087] S210 , performing inverse wavelet transform reconstruction processing according to the target detail coefficient and the target detail coefficient to obtain reconstructed noise reduction data.

[0088] S211. Perform EMD decomposition on the denoised data to obtain EMD decomposition data.

[0089] In this embodiment, for parameters such as battery temperature and voltage, there may be a trend drift with a relatively slow change. This method uses empirical mode decomposition to separate the trend term and short-term fluctuations, eliminating the impact of system drift.

[0090] Specifically, this method first performs EMD decomposition and transforms the data signal X c It is decomposed into several intrinsic mode functions (IMFs) and expressed as:

[0091]

[0092] Among them, the IMF k is the kth mode function, and r is the residual trend term.

[0093] S212: Perform trend item reconstruction processing on the EMD decomposition data to obtain trend item reconstructed data.

[0094] S213 : Smoothing the trend item reconstructed data to obtain first processed data.

[0095] In this embodiment, the method retains the lower frequency IMF k The trend term is further smoothed using the moving average method, using low-frequency IMF and residual r to represent the trend part of the system.

[0096]

[0097] S214 , performing residual denoising and reconstructing processing on the first processed data to obtain second processed data.

[0098] In this embodiment, the method can denoise and reconstruct the residuals, and convert X c The smooth trend term X t Remove X from the original signal f =X c -X t, the drift-removed data signal X is obtained f .

[0099] S215 : Perform normalization processing on the second processed data to obtain multidimensional time series data.

[0100] In this embodiment, normalization is performed to standardize the data to a distribution with a mean of 0 and a standard deviation of 1 to ensure that the relative weight of each variable to the model input remains consistent.

[0101]

[0102] S216. Construct a sample data set based on the multidimensional time series data.

[0103] In this embodiment, the normalized data X obtained after data cleaning is n , removes the interference of outliers, noise, and trend drift, and can be directly used for model state prediction.

[0104] Please see Figure 5 , Figure 5 The figure shows a framework for constructing a sample data set, which is actually a data set obtained by collecting and processing data related to the ship's power system.

[0105] By implementing this embodiment, the method can ensure the accuracy and robustness of the prediction model through data cleaning (which includes outlier detection, noise reduction and smoothing), thereby avoiding the influence of noise, environmental interference and sensor error during the acquisition process (these influences will make the original data X o including outliers, noise, and drift), thereby improving data usability.

[0106] S217. Construct an encoder based on multi-head attention mechanism, cross convolution and feature attention aggregation mechanism, and construct a decoder based on deconvolution and conditional self-attention mechanism.

[0107] In this embodiment, in the ship pure battery power system state prediction scenario, in order to accurately predict the future system state and capture the feature dependency and multi-dimensional feature interaction relationship in the time series data, this method designs a multi-head attention feature aggregation cross convolutional neural network encoder to achieve deep feature extraction of input data. The encoder receives the multi-dimensional time series data X of the ship pure battery system. n ,After encoding processing, an embedding representation of global dependencies and multi-scale features is generated as the input of the prediction model.

[0108] S218. Build an initial prediction model based on the encoder and decoder.

[0109] S219. Train the initial prediction model according to the sample data set and the preset training configuration data to obtain a trained ship pure battery power system state prediction model.

[0110] In this embodiment, the training configuration data includes at least the model training algorithm, loss function, gradient update algorithm, number of iterations, and training termination conditions;

[0111] In this embodiment, the model training algorithm is a supervised learning algorithm, the loss function is a mean square error loss function, and the gradient update algorithm is an adaptive learning rate optimization algorithm. Based on an encoder and decoder, this model constructs a novel encoder-decoder state prediction model for a ship's pure battery power system. By inputting a labeled training set into the model, the network is trained using supervised learning to obtain the pure battery power system state.

[0112] In this embodiment, the loss function is first defined: the mean square error loss function is used:

[0113]

[0114] Among them, n represents the number of samples, y i is the true label, Y i is the predicted value of the model.

[0115] Then, the gradient is updated: Adam is a commonly used adaptive learning rate optimization algorithm that combines the ideas of momentum gradient descent and RMSProp algorithm. The following is the formula of Adam optimization algorithm:

[0116] Let t be the current iteration number, η is the learning rate, β1 and β2 are the decay rate parameters, and ò is a small constant. First, calculate the first-order moment estimate m and the second-order moment estimate v of the gradient:

[0117] m t =β1·m t-1 +(1-β1)·g t ;

[0118]

[0119] Among them, g t is the gradient of the current iteration, and then the first-order moment estimate and the second-order moment estimate are bias corrected:

[0120]

[0121] Finally, update the model parameters:

[0122]

[0123] Among them, θt is the updated parameter.

[0124] In this embodiment, when the model reaches the set number of iterations or reaches the set loss function value in advance, the training stops.

[0125] In this embodiment, the final model parameters are saved at the end of the training as a deployment model for prediction of the ship's pure battery power system.

[0126] S220: Deploy the ship's pure battery power system state prediction model onto the hardware platform of the ship to be predicted.

[0127] S221. Obtain the battery capacity value of the ship to be predicted according to a preset output window size.

[0128] In this embodiment, the collected multi-dimensional sensor data set Xn of the ship pure battery power system is used, which includes the rated voltage V b (t), rated current I b (t), system ambient temperature T b (t), rated power P b (t), relative humidity H(t), water flow velocity Swater(t), ambient temperature Tenv(t), bearing temperature Tz(t) and ship speed v b (t).

[0129] In this embodiment, the training process may be performed based on a scale of 80% training set and 20% test set.

[0130] S222. Map the battery capacity value to a high-dimensional feature space through an embedding layer to obtain an embedded representation feature.

[0131] In this embodiment, the input data X n The dimensions are T × d, where d = 9, representing the dataset's nine dimensions. The encoding layer consists of three layers, employing a multi-head self-attention mechanism, cross-convolution kernels, and feature attention aggregation. The number of convolution channels per layer is set to 64, 128, and 256, respectively, gradually improving the representational power of feature extraction.

[0132] In this embodiment, the input data X n By mapping it to a high-dimensional feature space through the embedding layer, that is, the embedding representation Z, time series data can obtain a richer representation in multiple dimensions:

[0133] Z=E(X n );

[0134] Among them, the dimension of Z is T×d E , d EIt represents the embedding dimension, which is used to improve the feature expression of the input data. The initial number of convolution channels is set to 64. E is the embedding mapping function.

[0135] S223. Map the embedded representation features into query matrix, key matrix and value matrix through a multi-head self-attention mechanism layer.

[0136] In this embodiment, the multi-head self-attention mechanism is used to capture the global dependencies of time series data.

[0137] In this embodiment, the method maps the embedding representation Z into a query Q matrix, a key K matrix, and a value V matrix, that is, each attention head independently generates a set of Q, K, and V: Q = ZW Q , K=ZW K , V=ZW V , where W Q , W K , W V is the weight matrix of the attention head, and the dimension of each matrix is d E ×d q , d q is the dimension of the attention head, this method sets d q =8.

[0138] S224. Calculate the attention weight of each attention head based on the query matrix and the key matrix.

[0139] In this embodiment, the method can calculate the dependency between each time step based on each attention head of the attention mechanism, and obtain the attention weight by scaling the dot product of the query matrix Q and the key matrix K:

[0140]

[0141] S225. Calculate the output feature matrix of each attention head based on the attention weight and value matrix.

[0142] In this embodiment, the method can use the attention weight matrix A to perform weighted summation on the value matrix V to generate the output feature H of each head: H=AV.

[0143] S226. Merge the output feature matrix to obtain multi-head features.

[0144] S227. Perform linear transformation on the multi-head features to obtain the final output matrix of the multi-head self-attention mechanism layer.

[0145] In this embodiment, the method can merge the output features of all attention heads into multi-head features, and obtain the final output representation Z of the multi-head attention layer through linear transformation. md :

[0146] Zmd =Con(H1, H2, ..., H h )W O

[0147] Among them, H i =A i V, W O is the output weight matrix of the multi-head attention layer, Con is the merging function, and h represents the number of attention heads.

[0148] S228. Stabilize the embedded representation features through the multi-head self-attention mechanism layer to obtain the multi-head attention mechanism output data.

[0149] In this embodiment, in order to stabilize the training process and retain the original feature information, the output Z of the multi-head attention md With the input embedding Z, residual connections and layer normalization are used to improve stability:

[0150] Z n =LN(Z md +Z)

[0151] The obtained Z n The final multi-head attention mechanism output has enhanced global dependency information and important time step weighted information.

[0152] S229. Perform convolution processing on the output data of the multi-head attention mechanism through the cross convolution layer to obtain the output features of multiple different convolution kernels.

[0153] In this embodiment, in an application for predicting the state of a ship's pure battery power system, a cross-convolution layer uses multiple convolution kernels of different sizes to simultaneously extract multi-scale features, thereby capturing the state changes of the battery system at different time scales. The cross-convolution layer used in this method uses convolution kernels of multiple scales to capture feature information at different scales.

[0154] In this embodiment, the method can output Z according to the final multi-head attention mechanism n , use different convolution kernels to perform convolution operations and extract features at different time scales.

[0155] This method uses multi-scale convolution kernels (1*1, 3*3, 5*5), among which the 1*1 convolution kernel is used to capture the individual features of each time step and retain short-term information; the 3*3 convolution kernel extracts local features to capture short-term trend changes, such as current fluctuations in a short period of time; the 5*5 convolution kernel is used to extract features in a larger range to capture medium- and long-term trends and pattern changes, such as the long-term trend of voltage.

[0156] For a convolution kernel size of k*k, the corresponding extracted features are recorded as F k. Perform convolution operation:

[0157] F k =conv k*k (Z n ).

[0158] In this embodiment, the encoder includes an embedding layer, a multi-head self-attention mechanism layer, and a cross-convolution layer; the cross-convolution layer includes multiple multi-scale convolution kernels.

[0159] In this embodiment, the method fuses the features extracted by different convolution kernels to achieve the integration of multi-scale information and enhance the model's overall understanding of the battery status. Specifically, the output features F1, F3, and F5 of each convolution kernel are spliced through the channel dimension to form a fusion feature matrix F c Storing features of different scales in parallel allows subsequent layers to use these features for deeper learning, i.e.

[0160] F c =Con(F1, F3, F5).

[0161] S230. Obtain the convolution kernel weight corresponding to each convolution kernel.

[0162] In this example, to enable the model to automatically focus on features that are more critical in state prediction, a feature attention aggregation (FAA) mechanism is introduced in the cross-convolution layer. This mechanism assigns adaptive weights to each convolution feature and performs weighted fusion based on the feature importance of different convolution kernels.

[0163] Specifically, for each scale of convolution feature F k Assign a learned weight W_f^((k)), the size of the weight is automatically adjusted by the model according to the importance of the feature during training. The FAA process is as follows:

[0164]

[0165] Among them, F FAA is the final weighted fusion feature matrix. The FAA mechanism enables the model to focus more on features that have a strong explanatory power for battery state changes and suppress noise and irrelevant information.

[0166] S231. Perform channel-dimensional splicing processing on the output features of multiple different convolution kernels according to the convolution kernel weights to obtain a coding output matrix.

[0167] In this embodiment, the feature F after FAA processing FAA The final output of the cross convolution layer contains multi-scale fusion features and adaptively weighted feature information. The output features will serve as input to subsequent parts of the model for further decoding and prediction.

[0168] S232. Perform feature inverse expansion processing on the encoding output matrix through the deconvolution feature reconstruction layer to obtain decoding features.

[0169] In this embodiment, the decoder is responsible for converting the multi-scale and adaptive weighted features F output by the encoder into FAA The decoder uses deconvolution operations and conditional self-attention mechanisms to ensure that the model can accurately reconstruct the future state based on the features extracted by the encoder, while adaptively focusing on key features during the decoding process.

[0170] In this embodiment, similar to the encoder part, the data set is the time series data of the ship's pure battery system. During the training process, the ratio of the training set to the test set is 80% and 20%.

[0171] In this embodiment, the decoder has three deconvolution layers, each with a different kernel size and number of channels, gradually expanding the temporal dimension of features to reconstruct future system states. The number of deconvolution channels in each layer is set to 256, 128, and 64, ensuring a gradual reduction in the complexity of the data.

[0172] In this embodiment, the deconvolution operation is used to convert the encoding feature matrix F FAA Gradually expand back to the original time and feature dimensions to decode the future state characteristics of the system.

[0173] Specifically, according to the multi-scale weighted feature matrix F output by the encoder FAA , whose dimension is T'×d E Each layer uses different convolution kernels and strides to expand features in reverse.

[0174] The deconvolution output of the lth layer is recorded as D_l. For the lth layer deconvolution operation, its formula is expressed as:

[0175] D l =Deconv(D l-1 ;k l , s l , c l )

[0176] Among them, k l is the convolution kernel size, s l represents the stride, c l is the number of channels. The deconvolution layer is set as follows:

[0177] First layer of deconvolution: convolution kernel size k l =5, stride s l =2, number of channels c l =256;

[0178] Second layer of deconvolution: convolution kernel size k l =3, stride s l =2, number of channels c l =128;

[0179] The third layer of deconvolution: convolution kernel size k l =1, stride s l =1, number of channels c l =64.

[0180] S233. Generate a conditional feature matrix of the encoding output matrix through the conditional self-attention mechanism layer.

[0181] In this embodiment, a conditional self-attention mechanism is introduced in the decoder part, which enables the decoder to adaptively focus on key states according to specific conditional features during the process of reconstructing future states, thereby enhancing prediction accuracy.

[0182] In this embodiment, the method generates a conditional feature matrix C by encoding the feature F_FAA, where C is used to guide the weighted selection of important features in the decoding process, where:

[0183] C=F FAA W C

[0184] Among them, W C is the weight matrix used to map the encoder features into the conditional space.

[0185] S234. Calculate the attention weight matrix based on the conditional feature matrix and the decoding features.

[0186] In this embodiment, the conditional feature C and the decoding feature D are combined l Calculate the attention weight matrix A l , giving higher weights to features in the decoded features that have stronger indications of the future state, which is expressed as follows:

[0187]

[0188] S235. Conditionally weight the decoded features through the conditional self-attention mechanism layer and the attention weight matrix to obtain a weighted feature matrix.

[0189] In this embodiment, the conditional weighted feature generation operation is performed, and the attention weight matrix A l Conditionally weight the decoding features to obtain the weighted feature matrix D F :

[0190] D F =A l D l .

[0191] S236. Construct cross-layer short connection features through residual connection and normalization layer, weighted feature matrix and decoding features.

[0192] In this embodiment, in order to enhance the stability and information transmission efficiency of the decoder, residual connection and layer normalization are used to process the output of each layer of the decoder to ensure that information is not lost during the decoding process.

[0193] In this embodiment, the method uses the weighted decoding feature D F Directly add the features of the previous layer to build a short connection across layers. The residual connection operation is expressed as:

[0194] D R =D F +D l+1 .

[0195] S237. Perform layer normalization on the cross-layer short connection features through residual connection and normalization layer to obtain a decoding feature matrix.

[0196] In this embodiment, layer normalization is used to standardize the features after residual connection to enhance the numerical stability of the features:

[0197] D N =LayerNorm(D R ).

[0198] S238: Input the decoded feature matrix to the output prediction layer for prediction processing to obtain a battery health status prediction result.

[0199] In this embodiment, the decoding feature matrix processed by conditional self-attention and residual normalization is input to the output layer D N , generate the future system state prediction result Y:

[0200] Y=FullyConnected(D N );

[0201] Among them, the FullyConnected layer maps the decoded high-dimensional features to the time step and feature dimension of the target state and outputs the predicted sequence Y.

[0202] In this embodiment, the prediction sequence Y represents the key operating parameters of the battery system's future state, reflecting the battery system's health, performance, and potential abnormal changes. Specifically, the prediction sequence Y includes the following main contents:

[0203] 1) Battery voltage: the battery voltage change trend in several future time steps;

[0204] 2) Battery current: the change in the battery’s charge and discharge current in future time steps;

[0205] 3) Battery temperature: the change of battery temperature in future time steps;

[0206] 4) Remaining power: the percentage of remaining power in the battery in the future;

[0207] 5) Power output: The power output of the battery system at the future time step.

[0208] In this embodiment, the battery health status prediction result includes key operating parameters of the future state of the battery system, and the key operating parameters include at least battery voltage, battery current, battery temperature, remaining power, and power output.

[0209] In this embodiment, the decoder includes a deconvolution feature reconstruction layer, a conditional self-attention mechanism layer, a residual connection and normalization layer, and an output prediction layer.

[0210] Please see Figure 6 , Figure 6 A flow chart showing an example of a method for predicting the health status of an electric ship battery is shown.

[0211] In this embodiment, the method deploys the trained ship pure battery power system prediction model to the ship hardware platform. By setting the input window size of the model, when the model reads the capacity value of the previous n cycles, the model can directly calculate the health status of the ship battery, that is: State = NN (V b (t), I b (t), T b (t), P b (t),H(t),Swater(t),Tenv(t),Tz(t),v b (t))

[0212] Among them, NN represents the trained network.

[0213] In this example, the model designed to describe the method has good performance. Figure 7 A comparison chart of battery health status prediction simulations for electric ships, specifically for a pure battery power system, is shown. The curves in the figure compare the true voltage with the voltage predicted by different algorithms. The figure shows that the proposed algorithm's prediction accuracy fluctuates around the true value, with a low mean square error (MSE). The LSTM and SVR algorithms, in contrast, are significantly less accurate than the proposed method, further demonstrating the effectiveness of the proposed algorithm.

[0214] In this embodiment, the execution subject of the method may be a computing device such as a computer or a server, and this is not limited in this embodiment.

[0215] In this embodiment, the execution subject of the method may also be a smart device such as a smart phone, a tablet computer, etc., which is not limited in this embodiment.

[0216] As can be seen, the electric ship battery health status prediction method described in this embodiment can use the 3σ criterion to dynamically eliminate outliers. Furthermore, a multi-layer noise reduction and decomposition method combining wavelet transform and empirical mode decomposition effectively separates noise and trend terms in the signal, making it suitable for the complex interference patterns of battery and environmental data. Furthermore, the use of Z-score-based normalization can maintain the consistency of model inputs. It can be seen that the combination of multiple cleaning steps and the above steps can ensure that the predicted data has high accuracy and low interference characteristics.

[0217] At the same time, the multi-head self-attention mechanism captures the long-term dependencies of time series data, enabling the model to automatically focus on important time steps and key features from different angles. The cross-convolutional network can extract multi-scale local features through parallel operations of different convolution kernel sizes. The introduction of the feature attention aggregation mechanism in the cross-convolution layer can adaptively weight the features extracted by different convolution kernels to ensure that the model focuses on the more critical parts of the multi-scale features for state prediction. The FAA mechanism can suppress the interference of noise features by dynamically allocating weights, allowing the model to prioritize features with predictive value (such as current peaks and temperature anomalies in battery systems), so that accurate predictions can still be made under complex working conditions.

[0218] In addition, the decoder uses multi-layer deconvolution to gradually restore the temporal dimension of the encoded features, and introduces residual connections and layer normalization operations at each layer to ensure the complete transfer of information during the decoding process. Multi-layer deconvolution gradually expands the feature dimension, ensuring that the model can gradually reconstruct future states, while residual connections and layer normalization enhance the stability of decoding, making the model more stable and accurate in long-time-step prediction tasks.

[0219] Finally, key operating parameters such as battery voltage, current, temperature, and power output are used as the model's primary prediction targets, constructing a multi-dimensional prediction sequence. This allows for real-time monitoring of battery health, charge and discharge efficiency, and potential faults, enabling proactive planning of maintenance and operational strategies to ensure safe and efficient ship operation.

[0220] Example 3

[0221] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of a device for predicting the health status of an electric ship battery provided in this embodiment. Figure 8 As shown, the electric ship battery health status prediction device includes:

[0222] A first acquisition unit 310 is configured to acquire multi-dimensional data of the electric ship collected based on a sensor network; wherein the multi-dimensional data of the electric ship includes power system data, environmental data, and hull status;

[0223] A first constructing unit 320 is configured to construct a sample data set based on the multi-dimensional data of the electric ship;

[0224] A second construction unit 330 is used to construct an encoder based on a multi-head attention mechanism, a cross convolution and a feature attention aggregation mechanism, and to construct a decoder based on a deconvolution and a conditional self-attention mechanism;

[0225] A third construction unit 340 is configured to construct an initial prediction model based on the encoder and the decoder;

[0226] A training unit 350 is configured to train the initial prediction model based on the sample data set and preset training configuration data to obtain a trained ship battery-only power system state prediction model;

[0227] A deployment unit 360 is used to deploy the ship pure battery power system state prediction model to the hardware platform of the ship to be predicted;

[0228] A second acquiring unit 370 is configured to acquire a battery capacity value of the ship to be predicted according to a preset output window size;

[0229] The prediction unit 380 is used to predict the battery health state according to the battery capacity value and the ship pure battery power system state prediction model to obtain a battery health state prediction result.

[0230] In this embodiment, the explanation of the electric ship battery health status prediction device can refer to the description in Example 1 or Example 2, and will not be repeated in this embodiment.

[0231] It can be seen that the electric ship battery health status prediction device described in this embodiment can predict the future health status of the ship's pure battery in real time with high prediction accuracy, providing technical support for the safe and efficient operation of the ship's pure battery power system.

[0232] Example 4

[0233] Please see Figure 9 , Figure 9 This is a schematic diagram of the structure of a device for predicting the health status of an electric ship battery provided in this embodiment. Figure 9 As shown, the electric ship battery health status prediction device includes:

[0234] A first acquisition unit 310 is configured to acquire multi-dimensional data of the electric ship collected based on a sensor network; wherein the multi-dimensional data of the electric ship includes power system data, environmental data, and hull status;

[0235] A first constructing unit 320 is configured to construct a sample data set based on the multi-dimensional data of the electric ship;

[0236] A second construction unit 330 is used to construct an encoder based on a multi-head attention mechanism, a cross convolution and a feature attention aggregation mechanism, and to construct a decoder based on a deconvolution and a conditional self-attention mechanism;

[0237] A third construction unit 340 is configured to construct an initial prediction model based on the encoder and the decoder;

[0238] A training unit 350 is configured to train the initial prediction model based on the sample data set and preset training configuration data to obtain a trained ship battery-only power system state prediction model;

[0239] A deployment unit 360 is used to deploy the ship pure battery power system state prediction model to the hardware platform of the ship to be predicted;

[0240] A second acquiring unit 370 is configured to acquire a battery capacity value of the ship to be predicted according to a preset output window size;

[0241] The prediction unit 380 is used to predict the battery health state according to the battery capacity value and the ship pure battery power system state prediction model to obtain a battery health state prediction result.

[0242] In this embodiment, the power system data includes at least rated voltage, rated current, rated power and power system ambient temperature;

[0243] Environmental data include at least water flow rate, ambient temperature and relative humidity;

[0244] The hull status includes at least the bearing temperature and the ship speed;

[0245] The sensor network includes battery pack sensors and environmental sensors;

[0246] Battery pack sensors are used to collect power system data;

[0247] The acquisition frequency of the battery pack sensor is the first preset acquisition frequency;

[0248] Environmental sensors include humidity sensors, flow rate sensors, temperature sensors and GPS sensors. Environmental sensors are used to collect environmental data and hull status;

[0249] The collection frequency of the environmental sensor is the second preset collection frequency.

[0250] As an optional implementation, the first building unit 320 includes:

[0251] The fusion subunit 321 is used to perform data fusion processing on the multi-dimensional data of the electric ship to obtain a multi-dimensional data vector;

[0252] a cleaning subunit 322, configured to perform data cleaning processing on the multidimensional data vector to obtain first processed data;

[0253] a denoising subunit 323, configured to perform residual denoising and reconstructing processing on the first processed data to obtain second processed data;

[0254] a normalization subunit 324 for performing normalization processing on the second processed data to obtain multidimensional time series data;

[0255] The construction subunit 325 is used to construct a sample data set according to the multi-dimensional time series data.

[0256] As an optional implementation, the fusion subunit 321 includes:

[0257] A synchronization processing module is used to perform time synchronization processing on multi-dimensional data using a time interpolation algorithm to obtain synchronized data;

[0258] A fusion processing module is used to perform data fusion processing on the synchronized data using a weighted average algorithm to obtain a preliminary feature matrix;

[0259] The vectorization processing module is used to perform vectorization processing on the preliminary feature matrix to obtain a multi-dimensional data vector.

[0260] As an optional embodiment, the cleaning subunit 322 includes:

[0261] A first calculation module is used to calculate the mean and standard deviation of each data point in the multidimensional data vector;

[0262] A determination module is used to determine the outlier value of each data point in the multidimensional data vector according to the mean value, standard deviation and preset outlier determination conditions to obtain an outlier determination result;

[0263] The elimination processing module is used to perform an outlier elimination process on the multidimensional data vector according to the outlier determination result to obtain eliminated data;

[0264] a decomposition processing module, configured to decompose the first processed data to obtain an approximation coefficient and detail coefficients of each layer;

[0265] A filtering processing module is used to perform soft threshold filtering on the detail coefficients to obtain the filtered target detail coefficients;

[0266] The reconstruction processing module is used to perform inverse wavelet transform reconstruction processing according to the target detail coefficient and the target detail coefficient to obtain reconstructed noise reduction data;

[0267] The decomposition processing module is also used to perform EMD decomposition on the noise reduction data to obtain EMD decomposition data;

[0268] The reconstruction processing module is also used to perform trend item reconstruction processing on the EMD decomposition data to obtain trend item reconstructed data;

[0269] The smoothing processing module is used to perform smoothing processing on the trend item reconstruction data to obtain first processed data.

[0270] As an optional implementation, the prediction unit 380 includes:

[0271] The encoding subunit 381 is used to perform feature encoding processing on the battery capacity value through an encoder to obtain an encoding output matrix;

[0272] The decoding subunit 382 is used to decode the encoded output matrix through a decoder to obtain a battery health status prediction result; wherein the battery health status prediction result includes key operating parameters of the future state of the battery system, and the key operating parameters include at least battery voltage, battery current, battery temperature, remaining power, and power output.

[0273] In this embodiment, the encoder includes an embedding layer, a multi-head self-attention mechanism layer, and a cross convolution layer;

[0274] The cross convolution layer includes multiple multi-scale convolution kernels;

[0275] The decoder includes a deconvolution feature reconstruction layer, a conditional self-attention mechanism layer, a residual connection and normalization layer, and an output prediction layer;

[0276] The training configuration data includes at least the model training algorithm, loss function, gradient update algorithm, number of iterations, and training termination conditions;

[0277] The model training algorithm is a supervised learning algorithm, the loss function is the mean square error loss function, and the gradient update algorithm is an adaptive learning rate optimization algorithm.

[0278] As an optional implementation, the encoding subunit 381 includes:

[0279] A mapping module is used to map the battery capacity value to a high-dimensional feature space through an embedding layer to obtain an embedded representation feature;

[0280] The mapping module is also used to map the embedded representation features into query matrix, key matrix and value matrix through the multi-head self-attention mechanism layer;

[0281] The second calculation module is used to calculate the attention weight of each attention head according to the query matrix and the key matrix;

[0282] The second calculation module is also used to calculate the output feature matrix of each attention head based on the attention weight and value matrix;

[0283] The merging module is used to merge the output feature matrix to obtain multi-head features;

[0284] The transformation module is used to perform linear transformation on multi-head features to obtain the final output matrix of the multi-head self-attention mechanism layer;

[0285] The stability processing module is used to perform stability processing on the embedded representation features through the multi-head self-attention mechanism layer to obtain the multi-head attention mechanism output data;

[0286] The convolution processing module is used to perform convolution processing on the output data of the multi-head attention mechanism through a cross convolution layer to obtain the output features of multiple different convolution kernels;

[0287] The acquisition module is used to obtain the convolution kernel weight corresponding to each convolution kernel;

[0288] The splicing module is used to perform channel-dimensional splicing on the output features of multiple different convolution kernels according to the convolution kernel weights to obtain the encoded output matrix.

[0289] As an optional implementation, the decoding subunit 382 includes:

[0290] The expansion processing module is used to perform feature inverse expansion processing on the encoding output matrix through the deconvolution feature reconstruction layer to obtain decoding features;

[0291] A generation module for generating a conditional feature matrix of the encoding output matrix through a conditional self-attention mechanism layer;

[0292] The third calculation module is used to calculate the attention weight matrix based on the conditional feature matrix and the decoding features;

[0293] The weighted processing module is used to perform conditional weighted processing on the decoded features through the conditional self-attention mechanism layer and the attention weight matrix to obtain a weighted feature matrix;

[0294] A construction module for constructing cross-layer short connection features through residual connections, normalization layers, weighted feature matrices, and decoding features;

[0295] The normalization processing module is used to perform layer normalization processing on the cross-layer short connection features through residual connection and normalization layer to obtain the decoding feature matrix;

[0296] The prediction processing module is used to input the decoded feature matrix into the output prediction layer for prediction processing to obtain the battery health status prediction result.

[0297] In this embodiment, the explanation of the electric ship battery health status prediction device can refer to the description in Example 1 or Example 2, and will not be repeated in this embodiment.

[0298] It can be seen that the electric ship battery health status prediction device described in this embodiment can predict the future health status of the ship's pure battery in real time with high prediction accuracy, providing technical support for the safe and efficient operation of the ship's pure battery power system.

[0299] An embodiment of the present application provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the electric ship battery health status prediction method in embodiment 1 or embodiment 2 of the present application.

[0300] An embodiment of the present application provides a computer-readable storage medium storing computer program instructions. When the computer program instructions are read and executed by a processor, the method for predicting the health status of an electric ship battery in embodiment 1 or embodiment 2 of the present application is executed.

[0301] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0302] In addition, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0303] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0304] The foregoing is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included within the scope of protection of the present application. It should be noted that similar reference numerals and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined or explained in subsequent figures.

[0305] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0306] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

Claims

1. A method for predicting the health status of an electric ship battery, characterized in that: include: Acquiring multidimensional data of the electric ship collected based on a sensor network; wherein the multidimensional data of the electric ship includes power system data, environmental data and hull status; Constructing a sample data set according to the multidimensional data of the electric ship; Construct an encoder based on multi-head attention mechanism, cross convolution and feature attention aggregation mechanism, and a decoder based on deconvolution and conditional self-attention mechanism; constructing an initial prediction model based on the encoder and the decoder; Training the initial prediction model according to the sample data set and preset training configuration data to obtain a trained ship pure battery power system state prediction model; Deploying the ship's pure battery power system state prediction model onto the hardware platform of the ship to be predicted; Obtaining the battery capacity value of the ship to be predicted according to a preset output window size; Performing battery health state prediction based on the battery capacity value and the ship pure battery power system state prediction model to obtain a battery health state prediction result; The battery health status prediction is performed based on the battery capacity value and the ship pure battery power system state prediction model to obtain a battery health status prediction result, including: Performing feature encoding processing on the battery capacity value by the encoder to obtain an encoding output matrix; Decoding the encoded output matrix by the decoder to obtain a battery health status prediction result; wherein the battery health status prediction result includes key operating parameters of the future state of the battery system, and the key operating parameters include at least battery voltage, battery current, battery temperature, remaining charge, and power output; The encoder includes an embedding layer, a multi-head self-attention mechanism layer, and a cross convolution layer; The cross convolution layer includes multiple multi-scale convolution kernels; The decoder includes a deconvolution feature reconstruction layer, a conditional self-attention mechanism layer, a residual connection and normalization layer, and an output prediction layer; The training configuration data includes at least a model training algorithm, a loss function, a gradient update algorithm, a number of iterations, and a training termination condition; The model training algorithm is a supervised learning algorithm, the loss function is a mean square error loss function, and the gradient update algorithm is an adaptive learning rate optimization algorithm; The step of performing feature encoding processing on the battery capacity value by the encoder to obtain an encoding output matrix includes: Mapping the battery capacity value to a high-dimensional feature space through the embedding layer to obtain an embedded representation feature; Mapping the embedding representation features into a query matrix, a key matrix, and a value matrix through the multi-head self-attention mechanism layer; Calculate the attention weight of each attention head according to the query matrix and the key matrix; Calculating an output feature matrix of each of the attention heads according to the attention weights and the value matrix; Merging the output feature matrices to obtain multi-head features; Performing a linear transformation on the multi-head features to obtain the final output matrix of the multi-head self-attention mechanism layer; Performing stability processing on the embedded representation features through the multi-head self-attention mechanism layer to obtain multi-head attention mechanism output data; Performing convolution processing on the output data of the multi-head attention mechanism through the cross convolution layer to obtain output features of multiple different convolution kernels; Get the convolution kernel weight corresponding to each convolution kernel; The output features of the multiple different convolution kernels are spliced in the channel dimension according to the convolution kernel weights to obtain an encoding output matrix.

2. The method for predicting the health status of an electric ship battery according to claim 1, characterized in that: The power system data includes at least rated voltage, rated current, rated power and power system ambient temperature; The environmental data at least includes water flow rate, ambient temperature and relative humidity; The hull status includes at least bearing temperature and ship speed; The sensor network includes a battery pack sensor and an environmental sensor; The battery pack sensor is used to collect the power system data; The acquisition frequency of the battery pack sensor is a first preset acquisition frequency; The environmental sensors include a humidity sensor, a flow rate sensor, a temperature sensor and a GPS sensor, and the environmental sensors are used to collect the environmental data and the hull status; The acquisition frequency of the environmental sensor is a second preset acquisition frequency.

3. The method for predicting the health status of an electric ship battery according to claim 1, characterized in that: The constructing of a sample data set according to the multi-dimensional data of the electric ship includes: Performing data fusion processing on the multi-dimensional data of the electric ship to obtain a multi-dimensional data vector; Performing data cleaning processing on the multidimensional data vector to obtain first processed data; performing residual denoising and reconstructing processing on the first processed data to obtain second processed data; performing normalization processing on the second processed data to obtain multidimensional time series data; A sample data set is constructed according to the multidimensional time series data.

4. The method for predicting the health status of an electric ship battery according to claim 3, characterized in that: The performing data fusion processing on the multi-dimensional data of the electric ship to obtain a multi-dimensional data vector includes: Performing time synchronization processing on the multidimensional data using a time interpolation algorithm to obtain synchronized data; A weighted average algorithm is used to perform data fusion processing on the synchronous data to obtain a preliminary feature matrix; The preliminary feature matrix is vectorized to obtain a multi-dimensional data vector.

5. The method for predicting the health status of an electric ship battery according to claim 3, characterized in that: The performing data cleaning processing on the multidimensional data vector to obtain first processed data includes: Calculating the mean and standard deviation of each data point in the multidimensional data vector; Performing outlier determination on each data point in the multidimensional data vector according to the mean value, the standard deviation, and a preset outlier determination condition to obtain an outlier determination result; Performing outlier elimination processing on the multidimensional data vector according to the outlier determination result to obtain eliminated data; Decomposing the first processed data to obtain an approximation coefficient and detail coefficients of each layer; Performing soft threshold filtering on the detail coefficients to obtain filtered target detail coefficients; Performing inverse wavelet transform reconstruction processing according to the target detail coefficient and the target detail coefficient to obtain reconstructed noise reduction data; Performing EMD decomposition on the noise reduction data to obtain EMD decomposition data; Performing trend item reconstruction processing on the EMD decomposition data to obtain trend item reconstructed data; The trend item reconstructed data is smoothed to obtain first processed data.

6. The method for predicting the health status of an electric ship battery according to claim 1, characterized in that: The decoding the encoded output matrix by the decoder to obtain a battery health status prediction result includes: Performing feature inverse expansion processing on the encoding output matrix through the deconvolution feature reconstruction layer to obtain decoding features; Generate a conditional feature matrix of the encoding output matrix through the conditional self-attention mechanism layer; Calculating an attention weight matrix according to the conditional feature matrix and the decoding features; Conditionally weighting the decoded features using the conditional self-attention mechanism layer and the attention weight matrix to obtain a weighted feature matrix; Constructing a cross-layer short connection feature through the residual connection and normalization layer, the weighted feature matrix and the decoding feature; Performing layer normalization processing on the cross-layer short connection features through the residual connection and normalization layer to obtain a decoding feature matrix; The decoded feature matrix is input into the output prediction layer for prediction processing to obtain a battery health status prediction result.

7. A device for predicting the health status of an electric ship battery, characterized in that: The electric ship battery health status prediction device includes: A first acquisition unit is configured to acquire multi-dimensional data of the electric ship collected based on a sensor network; wherein the multi-dimensional data of the electric ship includes power system data, environmental data, and hull status; A first constructing unit is configured to construct a sample data set based on the multi-dimensional data of the electric ship; The second building block is used to construct an encoder based on a multi-head attention mechanism, cross convolution, and feature attention aggregation mechanism, and a decoder based on a deconvolution and conditional self-attention mechanism; A third construction unit is configured to construct an initial prediction model based on the encoder and the decoder; A training unit, configured to train the initial prediction model according to the sample data set and preset training configuration data to obtain a trained ship pure battery power system state prediction model; A deployment unit, configured to deploy the ship pure battery power system state prediction model onto a hardware platform of the ship to be predicted; A second acquiring unit, configured to acquire the battery capacity value of the ship to be predicted according to a preset output window size; a prediction unit, configured to predict the battery health state based on the battery capacity value and the ship pure battery power system state prediction model to obtain a battery health state prediction result; The prediction unit includes: an encoding subunit, configured to perform feature encoding processing on the battery capacity value through the encoder to obtain an encoding output matrix; a decoding subunit, configured to decode the encoded output matrix through the decoder to obtain a battery health status prediction result; wherein the battery health status prediction result includes key operating parameters of the future state of the battery system, and the key operating parameters include at least battery voltage, battery current, battery temperature, remaining charge, and power output; The encoder includes an embedding layer, a multi-head self-attention mechanism layer, and a cross convolution layer; The cross convolution layer includes multiple multi-scale convolution kernels; The decoder includes a deconvolution feature reconstruction layer, a conditional self-attention mechanism layer, a residual connection and normalization layer, and an output prediction layer; The training configuration data includes at least a model training algorithm, a loss function, a gradient update algorithm, a number of iterations, and a training termination condition; The model training algorithm is a supervised learning algorithm, the loss function is a mean square error loss function, and the gradient update algorithm is an adaptive learning rate optimization algorithm; The encoding subunit includes: A mapping module, configured to map the battery capacity value to a high-dimensional feature space through the embedding layer to obtain an embedded representation feature; A mapping module is further configured to map the embedding representation features into a query matrix, a key matrix, and a value matrix through the multi-head self-attention mechanism layer; A second calculation module is used to calculate the attention weight of each attention head according to the query matrix and the key matrix; A second calculation module is further used to calculate the output feature matrix of each attention head according to the attention weight and the value matrix; A merging module, used to merge the output feature matrices to obtain multi-head features; A transformation module, configured to perform a linear transformation on the multi-head features to obtain a final output matrix of the multi-head self-attention mechanism layer; a stability processing module, configured to perform stability processing on the embedded representation features through the multi-head self-attention mechanism layer to obtain multi-head attention mechanism output data; A convolution processing module, configured to perform convolution processing on the output data of the multi-head attention mechanism through the cross convolution layer to obtain output features of multiple different convolution kernels; The acquisition module is used to obtain the convolution kernel weight corresponding to each convolution kernel; A splicing module is used to perform channel-dimensional splicing processing on the output features of the multiple different convolution kernels according to the convolution kernel weights to obtain an encoded output matrix.

Images