A method for predicting battery capacity through multimodal feature fusion

Abstract

This invention relates to a multimodal feature fusion-based battery capacity prediction method. Specifically, it relates to a multimodal feature fusion-based battery capacity prediction method. The purpose of this invention is to address the problems of low accuracy and poor robustness in existing methods using deep learning to predict battery capacity. The process is as follows: a sample set is constructed based on historical battery production data features; the normalized sample set is divided into a training set and a test set; a hybrid neural network model of convolutional neural network and long short-term memory network is constructed, with the training set as the input and the predicted battery capacity value as the output; the hybrid neural network model is trained to obtain a trained model; the test set is input to the trained hybrid neural network model, and the trained model outputs the prediction result; the prediction result is inversely calculated to obtain the inverse value as the predicted battery capacity value.

Description

Technical Field

[0001] This invention relates to the field of battery capacity prediction, and more specifically to a battery capacity prediction method based on multimodal feature fusion. Background Technology

[0002] With the growing demand for electric vehicles and renewable energy storage, the manufacturing of high-efficiency and high-performance batteries has become a focal point of the industry. Battery capacity, as a key indicator of battery performance, is crucial for accurate prediction to optimize production processes and improve the quality of the final product. Traditionally, battery capacity is measured by performing a complete discharge cycle on newly manufactured batteries. Specifically, after battery production, the battery is charged and activated, at which point it contains 50%–60% charge. To measure battery capacity, the battery must first be fully charged, then completely discharged, with the amount of discharge measured during this process representing the battery capacity. Finally, the battery is charged to approximately 50% for long-term storage. This battery capacity measurement process is lengthy and consumes a significant amount of energy. Utilizing deep learning methods based on measurement data from the battery activation process to predict battery capacity can significantly save time and reduce energy waste. Summary of the Invention

[0003] The purpose of this invention is to address the problems of low accuracy and poor robustness in existing methods that use deep learning to predict battery capacity, and to propose a battery capacity prediction method based on multimodal feature fusion.

[0004] The specific process of a multimodal feature fusion method for predicting battery capacity is as follows:

[0005] Step 1: Obtain the characteristics of historical battery production data, and construct a sample set based on the characteristics of historical battery production data;

[0006] Step 2: Normalize each feature data in the sample set obtained in Step 1 to obtain a normalized sample set. Divide the normalized sample set into a training set and a test set.

[0007] Step 3: Construct a hybrid neural network model of convolutional neural network and long short-term memory network. The training set is used as the input of the hybrid neural network model of convolutional neural network and long short-term memory network, and the battery capacity prediction value is used as the output of the hybrid neural network model of convolutional neural network and long short-term memory network. Train the hybrid neural network model of convolutional neural network and long short-term memory network until the loss function converges to obtain the trained hybrid neural network model of convolutional neural network and long short-term memory network.

[0008] Step 4: Input the test set into the trained hybrid neural network model of convolutional neural network and long short-term memory network, and the trained hybrid neural network model of convolutional neural network and long short-term memory network outputs the prediction results;

[0009] The predicted results are then reverse-calculated to obtain the reverse-calculated value, which is used as the predicted battery capacity value.

[0010] The beneficial effects of this invention are as follows:

[0011] This invention proposes a hybrid neural network model combining convolutional neural networks and long short-term memory networks. This model effectively learns and extracts features from multiple measurement variables and predicts the capacity of newly produced batteries. The method has high accuracy and robustness.

[0012] This invention can be applied to production lines in battery manufacturing plants. During battery activation, various production parameters are collected in real time and fed into the deep learning model proposed in this invention. The model can predict battery capacity values ​​more accurately, reducing production time, reducing power consumption in traditional measurement processes, and saving energy. Attached Figure Description

[0013] Figure 1 This is a flowchart of the process of the present invention;

[0014] Figure 2 This is a comparison chart of the predicted capacity and the actual capacity of the model in this invention;

[0015] Figure 3 This is a graph showing the absolute error between the predicted capacity and the actual capacity of the model in this invention.

[0016] Figure 4 This is a bar chart showing the distribution of the absolute error between the predicted capacity and the actual capacity of the model in this invention, categorized by interval. Detailed Implementation

[0017] Specific Implementation Method 1: The specific process of the battery capacity prediction method based on multimodal feature fusion in this implementation method is as follows:

[0018] Step 1: Obtain the characteristics of historical battery production data, and construct a sample set based on the characteristics of historical battery production data;

[0019] Step 2: Normalize each feature data in the sample set obtained in Step 1 to obtain a normalized sample set. Divide the normalized sample set into a training set and a test set.

[0020] Step 3: Construct a hybrid neural network model of convolutional neural network and long short-term memory network. The training set is used as the input of the hybrid neural network model of convolutional neural network and long short-term memory network, and the battery capacity prediction value is used as the output of the hybrid neural network model of convolutional neural network and long short-term memory network. Train the hybrid neural network model of convolutional neural network and long short-term memory network until the loss function converges (the cross-entropy loss function is calculated based on the difference between the battery capacity prediction value and the actual battery capacity). The trained hybrid neural network model of convolutional neural network and long short-term memory network is obtained.

[0021] Step 4: Input the test set into the trained hybrid neural network model of convolutional neural network and long short-term memory network, and the trained hybrid neural network model of convolutional neural network and long short-term memory network outputs the prediction results;

[0022] The prediction result is a normalized value. It is necessary to back-calculate the prediction result to obtain the back-calculated value as the predicted battery capacity value.

[0023] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: in step 1, the characteristics of historical battery production data are obtained, and a sample set is constructed based on these characteristics; the specific process is as follows:

[0024] The battery production history data features include 7 time-series features, 14 non-time-series features, as well as the battery's actual capacity, charging step, and overall battery status.

[0025] Charging steps, such as constant current charging (CC) and constant voltage charging (CV), are typical charging steps;

[0026] The overall status of the battery, such as charging, discharging, fully charged, or faulty;

[0027] The sample set consists of 7 temporal features and 14 non-temporal features;

[0028] The seven timing characteristics include battery internal pressure, temperature, measured voltage, current, and differential voltage. Differential capacitance value Theoretical capacity value Each input feature is ;

[0029] Among them are battery internal pressure, temperature, voltage, current, and differential voltage. Differential capacitance value Theoretical capacity value It contains 506 units of measurement data with a time resolution of 10 seconds (sampled every 10 seconds for battery internal pressure, temperature, voltage, current, and differential voltage). Differential capacitance value Theoretical capacity value A total of 506 measurements were taken of the battery's internal pressure, temperature, voltage, current, and differential voltage. Differential capacitance value Theoretical capacity value );

[0030] The 14 non-time-series characteristics include 1 discharge efficiency parameter, 10 positive and negative electrode material characteristic parameters, and 3 electrolyte characteristic parameters.

[0031] One discharge efficiency parameter is the battery discharge efficiency value;

[0032] The 10 positive and negative electrode material characteristic parameters are aluminum content, graphite purity, solid content of positive electrode slurry, solid content of negative electrode slurry, width of insulating tape on the front side of positive electrode head, width of insulating tape on the back side of positive electrode head, width of insulating tape on the front side of positive electrode tail, width of insulating tape on the back side of positive electrode tail, cobaltate content, and initial coulombic efficiency.

[0033] The three electrolyte characteristic parameters are electrolyte water content, electrolyte injection volume, and electrolyte temperature.

[0034] The other steps and parameters are the same as in Specific Implementation Method 1.

[0035] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that the theoretical capacity value... The calculation process is as follows:

[0036] The sampling time interval was calculated using the Simpson numerical integration method. The corresponding theoretical capacity value is given by the following formula:

[0037] in, For continuous sampling time points and , for The current value at that moment. for The current value at that moment. for The current value at that moment. Indicates the sampling time interval The corresponding theoretical capacity value;

[0038] The differential capacitance value The calculation process is as follows:

[0039]

[0040] The differential voltage value The calculation process is as follows:

[0041]

[0042] In the formula: express The differential capacitance value at time t. express The differential voltage value at time t. Indicates the sampling time interval The corresponding theoretical capacity value; This indicates the measured voltage value; express The measured voltage value at any given time; express The measured voltage value at any given time; express The theoretical capacity value at any given time; express The theoretical capacity value at any given time.

[0043] Other steps and parameters are the same as in specific implementation method one or two.

[0044] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that: in step 2, each feature data in the sample set obtained in step 1 is normalized to obtain a normalized sample set, which is then divided into a training set and a test set; the specific process is as follows:

[0045] Normalization specifically involves: normalizing each feature in the sample set. Normalized to a standard state distribution with a mean of 0 and a variance of 1, the formula is as follows:

[0046]

[0047] In the formula, For the sample set Each feature data, For the sample set The mean of each feature data, For the sample set The standard deviation of each feature data, The normalized version One feature data;

[0048] The normalized sample set is divided into a training set and a test set.

[0049] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0050] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that: in step 3, a hybrid neural network model combining convolutional neural networks and long short-term memory networks is constructed; the specific process is as follows:

[0051] Hybrid neural network models combining convolutional neural networks and long short-term memory networks include:

[0052] Non-time-series deep convolutional network models, time-series CNN+LSTM network models, time-series CNN network models, and time-series and non-time-series connection processing network models;

[0053] The non-time-series deep convolutional network model includes:

[0054] Input layer, first convolutional layer (Conv), first max pooling layer (MaxPool), first batch normalization (BN) layer, first LeakyReLU activation function layer, second convolutional layer (Conv), second max pooling layer (MaxPool), third convolutional layer (Conv), first add layer (adding elements one by one), second batch normalization (BN) layer, second LeakyReLU activation function layer, fourth convolutional layer (Conv), third LeakyReLU activation function layer, fifth convolutional layer (Conv), fourth LeakyReLU activation function layer, sixth convolutional layer (Conv), output layer;

[0055] The time-series CNN+LSTM network model includes:

[0056] First Transpose (similar to transpose, actually a dimensionality transformation, transforming the input of 7 time-series data (sequence length 506) into a 506x7 dimension, batch size 8), seventh convolutional layer (Conv), fifth LeakyReLU activation function layer, eighth convolutional layer (Conv), sixth LeakyReLU activation function layer, ninth convolutional layer (Conv), seventh LeakyReLU activation function layer, tenth convolutional layer (Conv), eleventh convolutional layer (Conv), second Add layer, third BN layer, eighth LeakyReLU activation function layer, second Transpose, first reshape, first Gather layer (indexing operation, collecting features along the specified dimension according to the index, an internal LSTM operator, automatically processed), first channel alignment Unsqueeze, first concatenated Concat layer (channel concatenation within the Concat layer), first Expand layer (flattening layer), second reshape, second Gather layer, second channel alignment Unsqueeze, second level The sequence is as follows: Concat layer, Expand layer, LSTM layer, Squeeze layer (channel compression, removing dimensions with a scale of 1), Reshape layer, Gather layer, Unsqueeze layer, Concat layer, Expand layer, Reshape layer, Gather layer, Unsqueeze layer, Concat layer, Expand layer, LSTM layer, Squeeze layer, Reshape layer, Gather layer, Unsqueeze layer, Concat layer, Expand layer, LSTM layer, Squeeze layer, Transpose layer;

[0057] The time-series CNN network model includes:

[0058] Fourth Transpose, Twelfth Convolutional Layer (Conv), Thirteenth Convolutional Layer (Conv), Fourteenth Convolutional Layer (Conv), Third Max Pooling Layer (MaxPool), Fourth BN Layer, Fourth Max Pooling Layer (MaxPool), Ninth ReLU Activation Layer, Fifteenth Convolutional Layer (Conv), Tenth Leaky ReLU Activation Layer, Sixteenth Convolutional Layer (Conv), Eleventh Leaky ReLU Activation Layer, Third Add Layer, Twelfth ReLU Activation Layer, Seventeenth Convolutional Layer (Conv), Thirteenth Leaky ReLU Activation Layer, Eighteenth Convolutional layer Conv, fourteenth Leaky ReLU activation layer, fourth Add layer, fifteenth ReLU activation layer, nineteenth Convolutional layer Conv, sixteenth Leaky ReLU activation layer, twentieth Convolutional layer Conv, seventeenth Leaky ReLU activation layer, twenty-first Convolutional layer Conv, fifth Add layer, eighteenth ReLU activation layer, twenty-second Convolutional layer Conv, nineteenth Leaky ReLU activation layer, twenty-third Convolutional layer Conv, twentieth Leaky ReLU activation layer, sixth Add layer. Layer 21 (ReLU activation function), Layer 24 (Conv), Layer 22 (LeakyReLU activation function), Layer 25 (Conv), Layer 23 (LeakyReLU activation function), Layer 26 (Conv), Layer 7 (Add), Layer 24 (ReLU activation function), Layer 27 (Conv), Layer 25 (LeakyReLU activation function), Layer 28 (Conv), Layer 26 (LeakyReLU activation function), Layer 8 (Add), Layer 27 (ReLU activation function), Layer 29 (Conv) 28th LeakyReLU activation layer, 30th convolutional layer Conv, 29th LeakyReLU activation layer, 31st convolutional layer Conv, 9th Add layer, 30th ReLU activation layer, 32nd convolutional layer Conv, 31st LeakyReLU activation layer, 33rd convolutional layer Conv, 32nd LeakyReLU activation layer, 10th Add layer, 33rd ReLU activation layer, 34th convolutional layer Conv, 34th ReLU activation layer, 35th convolutional layer Conv;

[0059] The time-series and non-time-series connection processing network model includes:

[0060] Input layer A, input layer B, and input layer C;

[0061] Input layer A receives the output features from a time-series CNN+LSTM network model;

[0062] Input layer B receives the output features of the time-series CNN network model;

[0063] The input layer C receives the output features from a non-time-series deep convolutional network model;

[0064] First Cast layer (data type converted to float32), Second Cast layer (data type converted to float32), Eleventh Add layer, Fifth BN layer, Seventh Concatenated Concat layer, Thirty-sixth Convolutional layer (Conv), Thirty-fifth ReLU activation function layer, Thirty-seventh Convolutional layer (Conv), Twelfth Add layer, Thirty-sixth LeakyReLU activation function layer, Thirty-eighth Convolutional layer (Conv), Thirty-seventh ReLU activation function layer, Thirty-ninth Convolutional layer (Conv), Thirteenth Add layer, Thirty-eighth LeakyReLU activation function layer, Fortieth Convolutional layer (Conv), Thirty-ninth LeakyReLU activation function layer, Fortyth Convolutional layer (Conv), Third Cast layer (data type converted to float32), Seventh Reshape layer, Eighth Reshape layer, Eighth Concatenated Concat layer, First Gemm layer, Second Gemm layer, Third Gemm layer, Fourth Gemm layer, Output layer;

[0065] The first Gemm layer, the second Gemm layer, the third Gemm layer, and the fourth Gemm layer are all fully connected layers.

[0066] The other steps and parameters are the same as in any of the specific implementation methods one to four.

[0067] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the working process of the non-time-series deep convolutional network model is as follows:

[0068] Non-temporal features are sequentially input into the first convolutional layer (Conv), the first max pooling layer (MaxPool), the first batch normalization (BN) layer, the first LeakyReLU activation function layer, the second convolutional layer (Conv), and the second max pooling layer (MaxPool). The second max pooling layer (MaxPool) outputs feature D.

[0069] Non-temporal features are input into the third convolutional layer Conv via the input layer, and the third convolutional layer Conv outputs feature E;

[0070] Input features D and E into the first Add layer, and the first Add layer outputs feature F.

[0071] Feature F is sequentially input into the second BN layer, the second LeakyReLU activation function layer, the fourth convolutional layer Conv, the third LeakyReLU activation function layer, the fifth convolutional layer Conv, the fourth LeakyReLU activation function layer, and the sixth convolutional layer Conv. The sixth convolutional layer Conv outputs feature G, which is then output by the output layer as the output feature of the non-time series deep convolutional network model.

[0072] The other steps and parameters are the same as those in any of the specific implementation methods one to five.

[0073] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the working process of the time series CNN+LSTM network model is as follows:

[0074] Temporal features are input to the first Transpose, and the first Transpose outputs features H.

[0075] Feature H is sequentially input into the seventh convolutional layer Conv, the fifth LeakyReLU activation function layer, the eighth convolutional layer Conv, the sixth LeakyReLU activation function layer, the ninth convolutional layer Conv, and the seventh LeakyReLU activation function layer. The seventh LeakyReLU activation function layer outputs feature I.

[0076] Feature H is sequentially input into the 10th convolutional layer Conv and the 11th convolutional layer Conv, and the 11th convolutional layer Conv outputs feature J;

[0077] Feature I and feature J are input into the second Add layer, and the second Add layer outputs feature K;

[0078] Feature K is sequentially input into the third BN layer, the eighth LeakyReLU activation function layer, and the second Transpose layer. The second Transpose layer outputs feature L.

[0079] Feature L is sequentially input into the first reshape layer, the first Gather layer, the first channel alignment Unsqueeze layer, the first concatenated Concat layer, and the first Expand layer (flattening layer). The first Expand layer (flattening layer) outputs feature M.

[0080] Feature L is sequentially input into the second reshape layer, the second Gather layer, the second channel alignment Unsqueeze layer, the second concatenated Concat layer, and the second Expand layer (flattening layer). The second Expand layer (flattening layer) outputs feature N.

[0081] Features L, M, and N are input into the first LSTM layer, and the first LSTM layer outputs feature O.

[0082] Feature O is input to the first Squeeze, and the first Squeeze outputs feature P;

[0083] The feature P is sequentially input into the third reshape layer, the third Gather layer, the third channel alignment Unsqueeze layer, the third concatenated Concat layer, and the third Expand layer (flattening layer). The third Expand layer (flattening layer) outputs the feature Q.

[0084] Feature P is sequentially input into the fourth reshape layer, the fourth Gather layer, the fourth channel alignment Unsqueeze layer, the fourth cascaded Concat layer, and the fourth Expand layer (flattening layer). The fourth Expand layer (flattening layer) outputs feature R.

[0085] Features P, Q, and R are input into the second LSTM layer, and the second LSTM layer outputs feature S.

[0086] Feature S is input to the second Squeeze, and the second Squeeze outputs feature T;

[0087] Feature T is sequentially input into the fifth reshape layer, the fifth Gather layer, the fifth channel alignment Unsqueeze layer, the fifth cascaded Concat layer, and the fifth Expand layer (flattening layer). The fifth Expand layer (flattening layer) outputs feature U.

[0088] Feature T is sequentially input into the sixth reshape layer, the sixth Gather layer, the sixth channel alignment Unsqueeze layer, the sixth cascaded Concat layer, and the sixth Expand layer (flattening layer). The sixth Expand layer (flattening layer) outputs feature V.

[0089] Features T, U, and V are input into the third LSTM layer, and the third LSTM layer outputs feature W.

[0090] Feature W is sequentially input into the third Squeeze and the third Transpose. The third Transpose outputs feature X, which serves as the output feature of the time series CNN+LSTM network model.

[0091] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0092] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that the working process of the time series CNN network model is as follows:

[0093] Temporal features are sequentially input into the fourth Transpose layer, the twelfth convolutional layer (Conv), the thirteenth convolutional layer (Conv), the fourteenth convolutional layer (Conv), the third max pooling layer (MaxPool), the fourth batch normalization (BN) layer, the fourth max pooling layer (MaxPool), and the ninth ReLU activation function layer. The ninth ReLU activation function layer outputs the features. ;

[0094] Output features of the ninth ReLU activation function layer The input layers are sequentially the 15th convolutional layer (Conv), the 10th LeakyReLU activation function layer, the 16th convolutional layer (Conv), and the 11th LeakyReLU activation function layer. The 11th LeakyReLU activation function layer outputs the features. ;

[0095] feature and characteristics Input to the third Add layer, output features of the third Add layer ;

[0096] feature The input is the twelfth ReLU activation function layer, and the output features of the twelfth ReLU activation function layer are... ;

[0097] feature The input layers are sequentially: the 17th convolutional layer (Conv), the 13th LeakyReLU activation layer, the 18th convolutional layer (Conv), and the 14th LeakyReLU activation layer. The 14th LeakyReLU activation layer outputs the features. ;

[0098] feature and characteristics Input to the fourth Add layer, output features from the fourth Add layer. ;

[0099] feature The input is the fifteenth ReLU activation function layer, and the output features of the fifteenth ReLU activation function layer are... ;

[0100] feature The input layers are sequentially processed by the nineteenth convolutional layer (Conv), the sixteenth LeakyReLU activation function layer, the twentieth convolutional layer (Conv), and the seventeenth LeakyReLU activation function layer. The seventeenth LeakyReLU activation function layer outputs the features. ;

[0101] feature Input to the 21st convolutional layer Conv, output features from the 21st convolutional layer Conv ;

[0102] feature and characteristics Input to the fifth Add layer, output features of the fifth Add layer ;

[0103] feature The input is the 18th ReLU activation function layer, and the output features of the 18th ReLU activation function layer are... ;

[0104] feature The input layers are sequentially: the 22nd convolutional layer (Conv), the 19th LeakyReLU activation function layer, the 23rd convolutional layer (Conv), and the 20th LeakyReLU activation function layer. The 20th LeakyReLU activation function layer outputs features. ;

[0105] feature and characteristics Input to the sixth Add layer, output features of the sixth Add layer ;

[0106] feature Input to the 21st ReLU activation function layer, output features from the 21st ReLU activation function layer. ;

[0107] feature The input layers are sequentially the 24th convolutional layer (Conv), the 22nd LeakyReLU activation layer, the 25th convolutional layer (Conv), and the 23rd LeakyReLU activation layer. The 23rd LeakyReLU activation layer outputs the features. ;

[0108] feature Input to the 26th convolutional layer Conv, output features from the 26th convolutional layer Conv ;

[0109] feature and characteristics Input to the seventh Add layer, output features of the seventh Add layer ;

[0110] feature Input to the 24th ReLU activation function layer, output features of the 24th ReLU activation function layer ;

[0111] feature The input layers are sequentially the 27th convolutional layer (Conv), the 25th LeakyReLU activation layer, the 28th convolutional layer (Conv), and the 26th LeakyReLU activation layer. The 26th LeakyReLU activation layer outputs the features. ;

[0112] feature and characteristics Input to the eighth Add layer, the eighth Add layer outputs features ;

[0113] feature Input to the 27th ReLU activation function layer, output features of the 27th ReLU activation function layer ;

[0114] feature The input layers are sequentially: the 29th convolutional layer (Conv), the 28th LeakyReLU activation layer, the 30th convolutional layer (Conv), and the 29th LeakyReLU activation layer. The 29th LeakyReLU activation layer outputs the features. ;

[0115] feature Input to the 31st convolutional layer Conv, output features from the 31st convolutional layer Conv ;

[0116] feature and characteristics Input to the ninth Add layer, output features from the ninth Add layer ;

[0117] feature The input layer is the 30th ReLU activation function layer, and the output layer is the feature. ;

[0118] feature The input layers are sequentially: the 32nd convolutional layer (Conv), the 31st LeakyReLU activation layer, the 33rd convolutional layer (Conv), and the 32nd LeakyReLU activation layer. The 32nd LeakyReLU activation layer outputs the features. ;

[0119] feature and characteristics Input to the 10th Add layer, output features of the 10th Add layer ;

[0120] feature The input layers are sequentially the 33rd ReLU activation layer, the 34th convolutional layer (Conv), the 34th ReLU activation layer, and the 35th convolutional layer (Conv). The 35th convolutional layer (Conv) outputs the features. ;

[0121] feature As the output of a time-series CNN network model.

[0122] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0123] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that the working process of the time series and non-time series connection processing network model is as follows:

[0124] Input layer A receives the output features from a time-series CNN+LSTM network model. These output features are then input to the first Cast layer, which outputs the features. ;

[0125] Input layer B receives the output features of a time-series CNN network model. These output features are then input to the second cast layer, which outputs the features. ;

[0126] feature and characteristics Input to the eleventh Add layer, output features of the eleventh Add layer ;

[0127] feature Input to the fifth BN layer, output features from the fifth BN layer ;

[0128] feature and the output features of the first Cast layer Input to the seventh cascaded Concat layer, output features from the seventh cascaded Concat layer. ;

[0129] feature Inputting the 36th convolutional layer (Conv), the 35th ReLU activation layer, and the 37th convolutional layer (Conv), the 37th convolutional layer (Conv) outputs the features. ;

[0130] feature and characteristics Input to the twelfth Add layer, output features of the twelfth Add layer ;

[0131] feature The input layer is the 36th LeakyReLU activation function layer, and the output layer is the feature. ;

[0132] feature The input layers are sequentially the 38th convolutional layer (Conv), the 37th ReLU activation layer, and the 39th convolutional layer (Conv). The 39th convolutional layer (Conv) outputs the features. ;

[0133] feature and characteristics Input to the thirteenth Add layer, output features of the thirteenth Add layer ;

[0134] feature The input layers are sequentially: the 38th LeakyReLU activation layer, the 40th convolutional layer (Conv), the 39th LeakyReLU activation layer, the 41st convolutional layer (Conv), and the 8th deformation (reshape). The 8th deformation (reshape) outputs the features. ;

[0135] The input layer C receives the output features of a non-time-series deep convolutional network model. These output features are then input to the third Cast layer, which outputs the features. ;

[0136] feature Input the seventh deformation reshape, output the features of the seventh deformation reshape ;

[0137] feature and characteristics Input to the eighth concatenated concat layer, output features from the eighth concatenated concat layer. ;

[0138] feature Input the first Gemm layer, the second Gemm layer, the third Gemm layer, and the fourth Gemm layer in sequence, and output the result of the fourth Gemm layer;

[0139] The output of the fourth Gemm layer serves as the output of the network model for processing both time-series and non-time-series connections.

[0140] The output is the predicted capacity value, which is a normalized result, and the output size is 1.

[0141] The evaluation of the prediction results specifically involves using the goodness of fit. The root mean square error (RMSE) and mean absolute error (MAE) are used to quantitatively evaluate the prediction results. The calculation formulas are as follows:

[0142]

[0143]

[0144]

[0145] In the formula, This is the actual battery capacity value. This is a predicted value for battery capacity. For the number of data points, For all The average value.

[0146]

[0147] From the table and Figure 2 , Figure 3 , Figure 4 As can be seen, the multi-channel convolutional neural network and temporal convolutional network models perform well in predicting battery capacity, with small errors and achieving the expected fitting results. Through multiple fittings, the results show little difference, indicating that the above models are robust to battery capacity prediction.

[0148] Overall, the battery capacity prediction model established by the battery capacity prediction method based on multi-channel convolutional neural networks and temporal convolutional networks proposed in this invention performs well in tests on actual battery data and achieves the expected purpose.

[0149] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0150] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Nine in that the inverse calculation formula in step 4 is as follows:

[0151]

[0152] In the formula, For the predicted battery capacity value, This represents the average of the actual battery capacity data from step 1. The standard deviation of the actual battery capacity data in step 1. This is the predicted battery capacity value calculated in reverse.

[0153] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0154] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A battery capacity prediction method based on multimodal feature fusion, characterized in that: The specific process of the method is as follows: Step 1: Obtain the characteristics of historical battery production data, and construct a sample set based on the characteristics of historical battery production data; Step 2: Normalize each feature data in the sample set obtained in Step 1 to obtain a normalized sample set. Divide the normalized sample set into a training set and a test set. Step 3: Construct a hybrid neural network model of convolutional neural network and long short-term memory network. The training set is used as the input of the hybrid neural network model of convolutional neural network and long short-term memory network, and the battery capacity prediction value is used as the output of the hybrid neural network model of convolutional neural network and long short-term memory network. Train the hybrid neural network model of convolutional neural network and long short-term memory network until the loss function converges to obtain the trained hybrid neural network model of convolutional neural network and long short-term memory network. Step 4: Input the test set into the trained hybrid neural network model of convolutional neural network and long short-term memory network, and the trained hybrid neural network model of convolutional neural network and long short-term memory network outputs the prediction results; The predicted results are then reverse-calculated to obtain the reverse-calculated value, which is used as the predicted battery capacity value.

2. The battery capacity prediction method based on multimodal feature fusion according to claim 1, characterized in that: In step 1, the characteristics of historical battery production data are obtained, and a sample set is constructed based on the characteristics of historical battery production data. The specific process is as follows: The battery production history data features include 7 time-series features, 14 non-time-series features, as well as the battery's actual capacity, charging step, and overall battery status. The sample set consists of 7 temporal features and 14 non-temporal features; The seven timing characteristics include battery internal pressure, temperature, measured voltage, current, and differential voltage. Differential capacitance value Theoretical capacity value Each input feature is ; Among them are battery internal pressure, temperature, voltage, current, and differential voltage. Differential capacitance value Theoretical capacity value It is a measurement data with a length of 506 and a time resolution of 10s; The 14 non-time-series characteristics include 1 discharge efficiency parameter, 10 positive and negative electrode material characteristic parameters, and 3 electrolyte characteristic parameters. One discharge efficiency parameter is the battery discharge efficiency value; The 10 positive and negative electrode material characteristic parameters are aluminum content, graphite purity, solid content of positive electrode slurry, solid content of negative electrode slurry, width of insulating tape on the front side of positive electrode head, width of insulating tape on the back side of positive electrode head, width of insulating tape on the front side of positive electrode tail, width of insulating tape on the back side of positive electrode tail, cobaltate content, and initial coulombic efficiency. The three electrolyte characteristic parameters are electrolyte water content, electrolyte injection volume, and electrolyte temperature.

3. The battery capacity prediction method based on multimodal feature fusion according to claim 2, characterized in that: The theoretical capacity value The calculation process is as follows: The sampling time interval was calculated using the Simpson numerical integration method. The corresponding theoretical capacity value is given by the following formula: in, For continuous sampling time points and , for The current value at that moment. for The current value at that moment. for The current value at that moment. Indicates the sampling time interval The corresponding theoretical capacity value; The differential capacitance value The calculation process is as follows: The differential voltage value The calculation process is as follows: In the formula: express The differential capacitance value at time t. express The differential voltage value at time t. Indicates the sampling time interval The corresponding theoretical capacity value; This indicates the measured voltage value; express The measured voltage value at any given time; express The measured voltage value at any given time; express The theoretical capacity value at any given time; express The theoretical capacity value at any given time.

4. The battery capacity prediction method based on multimodal feature fusion according to claim 3, characterized in that: In step 2, each feature data in the sample set obtained in step 1 is normalized to obtain a normalized sample set. The normalized sample set is then divided into a training set and a test set. The specific process is as follows: Normalization specifically involves: normalizing each feature in the sample set. Normalized to a standard state distribution with a mean of 0 and a variance of 1, the formula is as follows: In the formula, For the sample set Each feature data, For the sample set The mean of each feature data, For the sample set The standard deviation of each feature data, The normalized version One feature data; The normalized sample set is divided into a training set and a test set.

5. The battery capacity prediction method based on multimodal feature fusion according to claim 4, characterized in that: Step 3 involves constructing a hybrid neural network model combining convolutional neural networks and long short-term memory networks; the specific process is as follows: Hybrid neural network models combining convolutional neural networks and long short-term memory networks include: Non-time-series deep convolutional network models, time-series CNN+LSTM network models, time-series CNN network models, and time-series and non-time-series connection processing network models; The non-time-series deep convolutional network model includes: Input layer, first convolutional layer (Conv), first max pooling layer (MaxPool), first batch normalization (BN) layer, first LeakyReLU activation function layer, second convolutional layer (Conv), second max pooling layer (MaxPool), third convolutional layer (Conv), first add layer, second batch normalization (BN) layer, second LeakyReLU activation function layer, fourth convolutional layer (Conv), third LeakyReLU activation function layer, fifth convolutional layer (Conv), fourth LeakyReLU activation function layer, sixth convolutional layer (Conv), output layer; The time-series CNN+LSTM network model includes: First Transpose, Seventh Convolutional Layer (Conv), Fifth Leaky ReLU Activation Layer, Eighth Convolutional Layer (Conv), Sixth Leaky ReLU Activation Layer, Ninth Convolutional Layer (Conv), Seventh Leaky ReLU Activation Layer, Tenth Convolutional Layer (Conv), Eleventh Convolutional Layer (Conv), Second Add Layer, Third BN Layer, Eighth Leaky ReLU Activation Layer, Second Transpose, First Reshape, First Gather Layer, First Channel Alignment Unsqueeze, First Concatenated Concat Layer, First Expand Layer, Second Reshape, Second Gather Layer, Second Channel Alignment Unsqueeze, Second Concatenated Concat Layer, Second Expand Layer, First LSTM Layer, First Squeeze The layers are: ze, third reshape, third Gather layer, third channel alignment Unsqueeze, third concatenated Concat layer, third Expand layer, fourth reshape, fourth Gather layer, fourth channel alignment Unsqueeze, fourth concatenated Concat layer, fourth Expand layer, second LSTM layer, second Squeeze, fifth reshape, fifth Gather layer, fifth channel alignment Unsqueeze, fifth concatenated Concat layer, fifth Expand layer, sixth reshape, sixth Gather layer, sixth channel alignment Unsqueeze, sixth concatenated Concat layer, sixth Expand layer, third LSTM layer, third Squeeze, third Transpose; The time-series CNN network model includes: Fourth Transpose, Twelfth Convolutional Layer (Conv), Thirteenth Convolutional Layer (Conv), Fourteenth Convolutional Layer (Conv), Third Max Pooling Layer (MaxPool), Fourth BN Layer, Fourth Max Pooling Layer (MaxPool), Ninth ReLU Activation Layer, Fifteenth Convolutional Layer (Conv), Tenth Leaky ReLU Activation Layer, Sixteenth Convolutional Layer (Conv), Eleventh Leaky ReLU Activation Layer, Third Add Layer, Twelfth ReLU Activation Layer, Seventeenth Convolutional Layer (Conv), Thirteenth Leaky ReLU Activation Layer, Eighteenth Convolutional layer Conv, fourteenth Leaky ReLU activation layer, fourth Add layer, fifteenth ReLU activation layer, nineteenth Convolutional layer Conv, sixteenth Leaky ReLU activation layer, twentieth Convolutional layer Conv, seventeenth Leaky ReLU activation layer, twenty-first Convolutional layer Conv, fifth Add layer, eighteenth ReLU activation layer, twenty-second Convolutional layer Conv, nineteenth Leaky ReLU activation layer, twenty-third Convolutional layer Conv, twentieth Leaky ReLU activation layer, sixth Add layer. Layer 21 (ReLU activation function), Layer 24 (Conv), Layer 22 (LeakyReLU activation function), Layer 25 (Conv), Layer 23 (LeakyReLU activation function), Layer 26 (Conv), Layer 7 (Add), Layer 24 (ReLU activation function), Layer 27 (Conv), Layer 25 (LeakyReLU activation function), Layer 28 (Conv), Layer 26 (LeakyReLU activation function), Layer 8 (Add), Layer 27 (ReLU activation function), Layer 29 (Conv) 28th LeakyReLU activation layer, 30th convolutional layer Conv, 29th LeakyReLU activation layer, 31st convolutional layer Conv, 9th Add layer, 30th ReLU activation layer, 32nd convolutional layer Conv, 31st LeakyReLU activation layer, 33rd convolutional layer Conv, 32nd LeakyReLU activation layer, 10th Add layer, 33rd ReLU activation layer, 34th convolutional layer Conv, 34th ReLU activation layer, 35th convolutional layer Conv; The time-series and non-time-series connection processing network model includes: Input layer A, input layer B, and input layer C; Input layer A receives the output features from a time-series CNN+LSTM network model; Input layer B receives features output from the time-series CNN network model; The input layer C receives the output features from a non-time-series deep convolutional network model; First Cast layer, Second Cast layer, Eleventh Add layer, Fifth BN layer, Seventh Concatenated Concat layer, Thirty-sixth Convolutional layer (Conv), Thirty-fifth ReLU activation function layer, Thirty-seventh Convolutional layer (Conv), Twelfth Add layer, Thirty-sixth LeakyReLU activation function layer, Thirty-eighth Convolutional layer (Conv), Thirty-seventh ReLU activation function layer, Thirty-ninth Convolutional layer (Conv), Thirteenth Add layer, Thirty-eighth LeakyReLU activation function layer, Fortieth Convolutional layer (Conv), Thirty-ninth LeakyReLU activation function layer, Fortyth Convolutional layer (Conv), Third Cast layer, Seventh Reshape layer, Eighth Reshape layer, Eighth Concatenated Concat layer, First Gemm layer, Second Gemm layer, Third Gemm layer, Fourth Gemm layer, Output layer; The first Gemm layer, the second Gemm layer, the third Gemm layer, and the fourth Gemm layer are all fully connected layers.

6. The battery capacity prediction method based on multimodal feature fusion according to claim 5, characterized in that: The working process of the non-time-series deep convolutional network model is as follows: Non-temporal features are sequentially input into the first convolutional layer (Conv), the first max pooling layer (MaxPool), the first batch normalization (BN) layer, the first LeakyReLU activation function layer, the second convolutional layer (Conv), and the second max pooling layer (MaxPool). The second max pooling layer (MaxPool) outputs feature D. Non-temporal features are input into the third convolutional layer Conv via the input layer, and the third convolutional layer Conv outputs feature E; Input features D and E into the first Add layer, and the first Add layer outputs feature F. Feature F is sequentially input into the second BN layer, the second LeakyReLU activation function layer, the fourth convolutional layer Conv, the third LeakyReLU activation function layer, the fifth convolutional layer Conv, the fourth LeakyReLU activation function layer, and the sixth convolutional layer Conv. The sixth convolutional layer Conv outputs feature G, which is then output by the output layer as the output feature of the non-time series deep convolutional network model.

7. The battery capacity prediction method based on multimodal feature fusion according to claim 6, characterized in that: The working process of the time series CNN+LSTM network model is as follows: Temporal features are input to the first Transpose, and the first Transpose outputs features H. Feature H is sequentially input into the seventh convolutional layer Conv, the fifth LeakyReLU activation function layer, the eighth convolutional layer Conv, the sixth LeakyReLU activation function layer, the ninth convolutional layer Conv, and the seventh LeakyReLU activation function layer. The seventh LeakyReLU activation function layer outputs feature I. Feature H is sequentially input into the 10th convolutional layer Conv and the 11th convolutional layer Conv, and the 11th convolutional layer Conv outputs feature J; Feature I and feature J are input into the second Add layer, and the second Add layer outputs feature K; Feature K is sequentially input into the third BN layer, the eighth LeakyReLU activation function layer, and the second Transpose layer. The second Transpose layer outputs feature L. Feature L is sequentially input into the first reshape layer, the first Gather layer, the first channel alignment Unsqueeze layer, the first concatenated Concat layer, and the first Expand layer. The first Expand layer outputs feature M. Feature L is sequentially input into the second reshape layer, the second Gather layer, the second channel alignment Unsqueeze layer, the second concatenated Concat layer, and the second Expand layer. The second Expand layer outputs feature N. Features L, M, and N are input into the first LSTM layer, and the first LSTM layer outputs feature O. Feature O is input to the first Squeeze, and the first Squeeze outputs feature P; Feature P is sequentially input into the third reshape layer, the third Gather layer, the third channel alignment Unsqueeze layer, the third concatenated Concat layer, and the third Expand layer. The third Expand layer outputs feature Q. Feature P is sequentially input into the fourth reshape layer, the fourth Gather layer, the fourth channel alignment Unsqueeze layer, the fourth concatenation Concat layer, and the fourth Expand layer. The fourth Expand layer outputs feature R. Features P, Q, and R are input into the second LSTM layer, and the second LSTM layer outputs feature S. Feature S is input to the second Squeeze, and the second Squeeze outputs feature T; Feature T is sequentially input into the fifth reshape layer, the fifth Gather layer, the fifth channel alignment Unsqueeze layer, the fifth concatenated Concat layer, and the fifth Expand layer. The fifth Expand layer outputs feature U. Feature T is sequentially input into the sixth reshape layer, the sixth Gather layer, the sixth channel alignment Unsqueeze layer, the sixth concatenated Concat layer, and the sixth Expand layer. The sixth Expand layer outputs feature V. Features T, U, and V are input into the third LSTM layer, and the third LSTM layer outputs feature W. Feature W is sequentially input into the third Squeeze and the third Transpose. The third Transpose outputs feature X, which serves as the output feature of the time series CNN+LSTM network model.

8. The battery capacity prediction method based on multimodal feature fusion according to claim 7, characterized in that: The working process of the time series CNN network model is as follows: Temporal features are sequentially input into the fourth Transpose layer, the twelfth convolutional layer (Conv), the thirteenth convolutional layer (Conv), the fourteenth convolutional layer (Conv), the third max pooling layer (MaxPool), the fourth batch normalization (BN) layer, the fourth max pooling layer (MaxPool), and the ninth ReLU activation function layer. The ninth ReLU activation function layer outputs the features. ; Output features of the ninth ReLU activation function layer The input layers are sequentially the 15th convolutional layer (Conv), the 10th LeakyReLU activation function layer, the 16th convolutional layer (Conv), and the 11th LeakyReLU activation function layer. The 11th LeakyReLU activation function layer outputs the features. ; feature and characteristics Input to the third Add layer, output features of the third Add layer ; feature The input is the twelfth ReLU activation function layer, and the output features of the twelfth ReLU activation function layer are... ; feature The input layers are sequentially: the 17th convolutional layer (Conv), the 13th LeakyReLU activation layer, the 18th convolutional layer (Conv), and the 14th LeakyReLU activation layer. The 14th LeakyReLU activation layer outputs the features. ; feature and characteristics Input to the fourth Add layer, output features from the fourth Add layer. ; feature The input is the fifteenth ReLU activation function layer, and the output features of the fifteenth ReLU activation function layer are... ; feature The input layers are sequentially processed by the nineteenth convolutional layer (Conv), the sixteenth LeakyReLU activation function layer, the twentieth convolutional layer (Conv), and the seventeenth LeakyReLU activation function layer. The seventeenth LeakyReLU activation function layer outputs the features. ; feature Input to the 21st convolutional layer Conv, output features from the 21st convolutional layer Conv ; feature and characteristics Input to the fifth Add layer, output features of the fifth Add layer ; feature The input is the 18th ReLU activation function layer, and the output features of the 18th ReLU activation function layer are... ; feature The input layers are sequentially: the 22nd convolutional layer (Conv), the 19th LeakyReLU activation function layer, the 23rd convolutional layer (Conv), and the 20th LeakyReLU activation function layer. The 20th LeakyReLU activation function layer outputs features. ; feature and characteristics Input to the sixth Add layer, output features of the sixth Add layer ; feature Input to the 21st ReLU activation function layer, output features from the 21st ReLU activation function layer. ; feature The input layers are sequentially the 24th convolutional layer (Conv), the 22nd LeakyReLU activation layer, the 25th convolutional layer (Conv), and the 23rd LeakyReLU activation layer. The 23rd LeakyReLU activation layer outputs the features. ; feature Input to the 26th convolutional layer Conv, output features from the 26th convolutional layer Conv ; feature and characteristics Input to the seventh Add layer, output features of the seventh Add layer ; feature Input to the 24th ReLU activation function layer, output features of the 24th ReLU activation function layer ; feature The input layers are sequentially the 27th convolutional layer (Conv), the 25th LeakyReLU activation layer, the 28th convolutional layer (Conv), and the 26th LeakyReLU activation layer. The 26th LeakyReLU activation layer outputs the features. ; feature and characteristics Input to the eighth Add layer, the eighth Add layer outputs features ; feature Input to the 27th ReLU activation function layer, output features of the 27th ReLU activation function layer ; feature The input layers are sequentially: the 29th convolutional layer (Conv), the 28th LeakyReLU activation layer, the 30th convolutional layer (Conv), and the 29th LeakyReLU activation layer. The 29th LeakyReLU activation layer outputs the features. ; feature Input to the 31st convolutional layer Conv, output features from the 31st convolutional layer Conv ; feature and characteristics Input to the ninth Add layer, output features from the ninth Add layer ; feature The input layer is the 30th ReLU activation function layer, and the output layer is the feature. ; feature The input layers are sequentially: the 32nd convolutional layer (Conv), the 31st LeakyReLU activation layer, the 33rd convolutional layer (Conv), and the 32nd LeakyReLU activation layer. The 32nd LeakyReLU activation layer outputs the features. ; feature and characteristics Input to the 10th Add layer, output features of the 10th Add layer ; feature The input layers are sequentially the 33rd ReLU activation layer, the 34th convolutional layer (Conv), the 34th ReLU activation layer, and the 35th convolutional layer (Conv). The 35th convolutional layer (Conv) outputs the features. ; feature As the output of a time-series CNN network model.

9. The battery capacity prediction method based on multimodal feature fusion according to claim 8, characterized in that: The working process of the time series and non-time series connection processing network model is as follows: Input layer A receives the output features from a time-series CNN+LSTM network model. These output features are then input to the first Cast layer, which outputs the features. ; Input layer B receives the output features of a time-series CNN network model. These output features are then input to the second cast layer, which outputs the features. ; feature and characteristics Input to the eleventh Add layer, output features of the eleventh Add layer ; feature Input to the fifth BN layer, output features from the fifth BN layer ; feature and the output features of the first Cast layer Input to the seventh cascaded Concat layer, output features from the seventh cascaded Concat layer. ; feature Inputting the 36th convolutional layer (Conv), the 35th ReLU activation layer, and the 37th convolutional layer (Conv), the 37th convolutional layer (Conv) outputs the features. ; feature and characteristics Input to the twelfth Add layer, output features of the twelfth Add layer ; feature The input layer is the 36th LeakyReLU activation function layer, and the output layer is the feature. ; feature The input layers are sequentially the 38th convolutional layer (Conv), the 37th ReLU activation layer, and the 39th convolutional layer (Conv). The 39th convolutional layer (Conv) outputs the features. ; feature and characteristics Input to the thirteenth Add layer, output features of the thirteenth Add layer ; feature The input layers are sequentially: the 38th LeakyReLU activation layer, the 40th convolutional layer (Conv), the 39th LeakyReLU activation layer, the 41st convolutional layer (Conv), and the 8th deformation (reshape). The 8th deformation (reshape) outputs the features. ; The input layer C receives the output features of a non-time-series deep convolutional network model. These output features are then input to the third Cast layer, which outputs the features. ; feature Input the seventh deformation reshape, output the features of the seventh deformation reshape. ; feature and characteristics Input to the eighth concatenated concat layer, output features from the eighth concatenated concat layer. ; feature Input the first Gemm layer, the second Gemm layer, the third Gemm layer, and the fourth Gemm layer in sequence, and output the result of the fourth Gemm layer; The output of the fourth Gemm layer serves as the output of the network model for processing both time-series and non-time-series connections. The output is the predicted capacity value.

10. The battery capacity prediction method based on multimodal feature fusion according to claim 9, characterized in that: The inverse calculation formula in step 4 is as follows: In the formula, For the predicted battery capacity value, This represents the average of the actual battery capacity data from step 1. The standard deviation of the actual battery capacity data in step 1. This is the predicted battery capacity value calculated in reverse.

Images