A multi-scale fusion prediction method for residual service life of lithium ion battery

Abstract

The application relates to the technical field of battery capacity detection, and discloses a multi-scale fusion prediction method for the residual service life of a lithium ion battery, which comprises the following steps: collecting the charging current, voltage, temperature, resistance and historical capacity data of each battery before failure; collecting the charging current, voltage, temperature, resistance and historical capacity data of each battery before the starting point of life prediction, respectively; arranging the historical resistance and capacity data into a one-dimensional matrix vector; adopting a three-layer dilated convolution network to perform two-dimensional convolution on the original V, I and T matrices; inputting the obtained feature mapping into a global average pooling layer to reduce the dimension of the feature mapping; capturing the change law of the V, I and T compact features and the original resistance data along with the number of cycles by using an LSTM (Long Short Term Memory), modeling the long-term dependence relationship between the features, and quantifying the aging behavior of the battery; and inputting the capacity attenuation trend after the starting point of battery life prediction into a multilayer perceptron to predict the capacity attenuation trend.

Description

Technical Field

[0001] This invention relates to the field of battery capacity detection technology, specifically a multi-scale fusion prediction method for the remaining service life of lithium-ion batteries. Background Technology

[0002] Lithium-ion batteries are widely used due to their low pollution, high energy density, high power density, and long lifespan. During battery use, their electrochemical performance changes with the remaining lifespan. When performance degradation exceeds a certain threshold, lithium-ion batteries may experience a sharp decline in performance, leading to potential safety issues or even catastrophic failures. Therefore, accurate prediction of battery lifespan is crucial for promoting the safe and green development of the battery industry and for more scientific battery management systems. Furthermore, the battery development process inevitably requires comprehensive and repeated evaluation of cycle life, but existing lifespan testing consumes significant human, material, and testing resources. Lifespan feedback for long-life batteries typically takes months or even years of testing, resulting in high time costs. The rapid development of machine learning allows this invention to find potential correlations from large amounts of historical test data, providing accurate predictions of battery lifespan.

[0003] Most current research on lithium battery life prediction does not adequately consider health indicators for measuring battery capacity aging, modeling only single or a few measurement variables without comprehensively considering multiple health indicators. Insufficient features and limited datasets restrict the predictive and generalization performance of data-driven models. Furthermore, most current research is based on an "input-model-output" modeling baseline. For complex models, the input-to-output transition involves multiple neural network layers, potentially leading to the forgetting of important features while non-critical variable information is extracted, interfering with the model's predictions.

[0004] In summary, current lithium-ion battery remaining life prediction models face several key challenges, including insufficient comprehensive consideration of multiple health indicators, limited selection of features and datasets, excessively long paths between raw data input and model output, and difficulty in inheriting and retaining key information. These challenges result in low accuracy, weak robustness, and poor generalization performance in predicting battery remaining life capacity trends. To address these issues, this invention proposes a multi-scale fusion prediction method for lithium-ion battery remaining life. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a multi-scale fusion prediction method for the remaining lifespan of lithium-ion batteries. It comprehensively selects voltage (V), current (I), temperature (T), resistance, and historical capacity data during the battery charging process as HI. The V, I, and T data are easy to collect and monitor online, thereby solving the aforementioned technical problems.

[0007] (II) Technical Solution

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] A multi-scale fusion prediction method for the remaining service life of lithium-ion batteries includes the following steps:

[0010] Step 1: Collect charging current, voltage, temperature, resistance, and historical capacity data for each battery before failure from multiple battery capacity aging datasets;

[0011] Step 2: Collect charging current, voltage, temperature, resistance and historical capacity data before the start of each battery life prediction as HI to measure battery capacity aging. Stack the current, voltage and temperature into a three-dimensional matrix to meet the calculation requirements of the dilated convolution layer. Arrange the historical resistance and capacity data into a one-dimensional matrix vector.

[0012] Step 3: Use a 3-layer dilated convolutional network to perform two-dimensional convolution on the original V, I, and T matrices respectively;

[0013] Step 4: Input the obtained feature map into the global average pooling layer to reduce the dimensionality of the feature map;

[0014] Step 5: Stack the feature sequence obtained in the previous step together with the battery resistance data through matrix concatenation, and use it as the input of LSTM. LSTM captures the changes of V, I, T compact features and original resistance data with the number of cycles, models the long-term dependency between them, and quantifies the aging behavior of the battery.

[0015] Step 6: Input the feature map obtained in the previous step along with the historical capacity data into the multilayer perceptron to predict the capacity decay trend after the battery life prediction start point.

[0016] Preferably, the battery life prediction start point is when the battery capacity decays to 90% of the rated capacity.

[0017] Preferably, the second step includes interpolating the V, I, and T data for each cycle by filling in missing values ​​and using Akima interpolation.

[0018] Preferably, the third step is as follows:

[0019] Let the nth thThe output feature map of each dilated convolutional layer is O n , where n = 1, 2, ..., L, represents the number of dilated convolutional layers;

[0020] The shape of the VIT matrix of the input dilated convolutional layer is M. n ×N n The shape of the convolution kernel is P n ×Q n The weight matrix and bias of the convolution kernel are W, respectively. n and b k The expansion rate is (d h，n d w，n );

[0021] nth th The k-th layer of the dilated convolutional layer th Output feature mapping It can be represented as:

[0022]

[0023] represent The element at position (i, j) in the array; For the (n-1)th th The r-th receptive field R of the dilated convolutional layer th The element at position (i+a, j+b) in the input feature map; The element at position (p, q) corresponds to the nth element. th Layer k th The weights of each convolutional kernel;

[0024] Sense Wild (R) h，n+1 ×R w，n+1 The size is calculated as follows:

[0025]

[0026]

[0027] Preferably, the specific steps of the fourth step are as follows:

[0028] Let D be the feature map after the dilated convolutional layer. n All cells have a shape of (w×h), where n = 1, 2, ..., C, representing the cell cycle number. The global average pooling process is then expressed as:

[0029]

[0030] dn is the nth element in the output compact feature sequence. th Each element.

[0031] Preferably, the specific steps of the fifth step are as follows:

[0032] The resistance data from all loops before the prediction start point are integrated with the compact features obtained from global average pooling and used as input to the LSTM network, as shown below:

[0033] f t =σ(W f ⊙[h t-1 x t ]+b f );

[0034] i t =σ(W i ⊙[h t-1 x t ]+b i );

[0035] g t =tanh(W g ⊙[h t-1 x t ]+b g );

[0036] C t =f t ⊙C t-1 +i t ⊙g t ;

[0037] o t =σ(W Xo ⊙[h t-1 x t ]+b o );

[0038] h t =o t ⊙tanh(C t ).

[0039] Preferably, the specific steps of step six are as follows:

[0040] The LSTM output is integrated with historical capacity data and denoted as... As input to the multilayer sensor, the multilayer sensor generates a trend prediction result for the remaining usable capacity of the lithium-ion battery, as shown below:

[0041]

[0042] Where L(·) is the linear activation function, and f(·) represents the Sigmoid function, as follows:

[0043]

[0044] (III) Beneficial Effects

[0045] Compared with existing technologies, the multi-scale fusion prediction method for the remaining service life of lithium-ion batteries provided by this invention has the following beneficial effects:

[0046] 1. The multi-scale fusion prediction method for the remaining service life of lithium-ion batteries comprehensively selects voltage, current, temperature, resistance and historical capacity data as HI for quantifying battery aging. The rich feature data and a large number of public datasets improve the prediction accuracy and generalization performance of the model.

[0047] 2. The multi-scale fusion prediction method for the remaining lifespan of lithium-ion batteries uses dilated convolution to extract multiple features from V, I, and T data under the same cycle state. The receptive field, which increases with the number of network layers, can fuse the features of the three variables at multiple scales, making the obtained feature mapping more representative.

[0048] 3. The multi-scale fusion prediction method for the remaining lifespan of lithium-ion batteries, through LSTM and hysteresis multi-head input, on the one hand, determines the inheritance and rejection of original information on the long processing path through the "gate" structure of LSTM, and on the other hand, the hysteresis input directly shortens the processing path between key features and output, better preserves key feature information, overcomes the long-term dependence and gradient explosion problem of long sequence data prediction, and provides high-accuracy capacity aging trend prediction. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the multi-scale fusion prediction method for the remaining lifespan of lithium-ion batteries according to an embodiment of the present invention.

[0050] Figure 2 This is a capacity decay curve of 124 battery datasets used in an embodiment of the present invention.

[0051] Figure 3 The regularity differences of V, I, and T under different cycles in the embodiments of the present invention are shown.

[0052] Figure 4 This is a trend graph showing the changes in internal resistance and capacity of batteries with different battery lifespans in embodiments of the present invention.

[0053] Figure 5 This is a schematic diagram of the LSTM network structure according to an embodiment of the present invention;

[0054] Figure 6 This invention discloses the performance of the method for predicting capacity aging trends after the prediction starting point. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example

[0057] Please see Figure 1-6 The multi-scale fusion prediction method for the remaining service life of lithium-ion batteries provided in this embodiment of the invention includes the following steps:

[0058] Step 1: Collect charging current, voltage, temperature, resistance, and historical capacity data for each battery before failure from 124 battery capacity aging datasets.

[0059] Step 1 is implemented in the following steps:

[0060] Step 1.1: The experimental dataset consists of 124 commercial lithium iron phosphate / graphite batteries with a rated capacity of 1.1 Ah and a rated voltage of 3.3 V. These batteries were cycled to failure in a 48-channel Abin charge / discharge cabinet and a 30°C constant temperature chamber under 72 charging strategies and a fixed discharge rate. The dataset includes 96,700 cycle times, with the longest-lived battery having 2,338 cycles. Battery failure was defined as a capacity reduction to 80% of the rated capacity, with 90% of the rated capacity as the predicted starting point.

[0061] Figure 2 Capacity decay curves for 124 battery datasets are presented.

[0062] Step 2: Collect charging current, voltage, temperature, resistance, and historical capacity data for each battery before the predicted lifespan starting point as the HI (Highest Hierarchy Point) for measuring battery capacity aging. Stack the current, voltage, and temperature data into a three-dimensional matrix to meet the computational requirements of the dilated convolutional layer. Arrange the historical resistance and capacity data into a one-dimensional matrix vector for subsequent calculations.

[0063] Step 2 is implemented in the following steps:

[0064] Step 2.1: Extract the HI data before the start of battery life prediction for each individual cell. Standardize the HI data for each cycle to a length of 1024 by filling in missing values ​​and using -Akima interpolation. The dataset is randomly divided into training and testing sets in an 8:2 ratio. This means there will be 99 battery datasets with a total of 52 charging strategies and 25 datasets with 20 charging strategies, used for RUL estimation and model generalization performance validation, respectively.

[0065] In step 2.2, the battery's V, I, and T data showed varying degrees of shift as the cycle progressed. Furthermore, batteries with different lifespans already exhibited significant differences in their initial cycle states, and these regular differences were strongly correlated with battery lifespan.

[0066] Figure 3 The regular differences in V, I, and T under different cycles are presented. Figure 3 It can be seen that the voltage, current and temperature of batteries with different lifespans exhibit a regular distribution under the same cycle conditions. The shorter the remaining lifespan, the stronger the correlation between these HI values ​​and capacity decay.

[0067] Figure 4 The trend graphs of internal resistance and capacity changes of batteries with different battery lifespans (taking four of them as examples) are given. It can be seen that as the number of cycles increases, the aging trend of resistance and capacity shows a strong negative correlation.

[0068] Step 2.3 defines a special one-hot code as a marker for the start and end of data reception by the recognition model. This one-hot code will appear at the start and end points before the feature input, i.e., at the 0th loop and the prediction start point, and at the start and end points of the target data in the training set, i.e., at the capacity at the prediction start point and the capacity at the failure point. The model will automatically determine the length of the output sequence (i.e., the remaining lifetime) during prediction, and the one-hot code will also be output at the sequence output and the end point of the model.

[0069] Step 3: A 3-layer dilated convolutional network is used to perform two-dimensional convolution on the original V, I, and T matrices respectively, which increases the information receptive field of the convolution calculation process and greatly reduces the amount of computation while maintaining the shape of the original input feature map.

[0070] Step 3 is implemented in the following steps:

[0071] Step 3.1, for ease of representation, let the nth... th The output feature map of each dilated convolutional layer is O n Where n = 1, 2, ..., L, represents the number of dilated convolutional layers. The shape of the VIT matrix of the input dilated convolutional layer is M. n ×N n The shape of the convolution kernel is P n ×Q n The weight matrix and bias of the convolution kernel are W, respectively. n and b k The expansion rate is (d h，n d w，n At this time, the nth... th The k-th layer of the dilated convolutional layer th Output feature mapping It can be represented as:

[0072]

[0073] in, represent The element at position (i, j) in the array; For the (n-1)th th The r-th receptive field R of the dilated convolutional layer th The element at position (i+a, j+b) in the input feature map; The element at position (p, q) corresponds to the nth element. th Layer k th The weights of each convolutional kernel. The receptive field (R) of this layer. h，n+1 ×R w，n+1 The size is calculated as follows:

[0074]

[0075]

[0076] The receptive field of dilated convolutional layers increases exponentially with the number of layers, significantly reducing computational cost compared to traditional convolution. Furthermore, dilated convolutional layers overcome the shortcomings of traditional convolution in examining edge data; the increased receptive field enables multi-scale fusion of voltage, current, and temperature data under the same cyclic state, and the extracted compact features encompass the contributions of all three input variables.

[0077] Step 4: Input the obtained feature map into the global average pooling layer to reduce the dimensionality of the feature map. At this point, the feature map has a global receptive field.

[0078] Step 4 is implemented in the following steps:

[0079] Step 4.1, for ease of representation, let the feature map after dilated convolution layer processing be denoted as Dn, with a shape of (w×h), where n = 1, 2, ..., C, representing the cycle number of the battery. Then the global average pooling process can be represented as:

[0080]

[0081] Where, d n Then it is the nth feature in the output compact feature sequence. th Each element.

[0082] Step 5: Stack the feature sequence obtained in the previous step together with the battery resistance data via matrix concatenation, and use it as input to the LSTM. The LSTM captures the changes in V, I, and T features and the original resistance data with the number of cycles, models the long-term dependencies between them, and quantifies the battery's aging behavior.

[0083] Step 5 is implemented in the following steps:

[0084] Step 5.1: First, the resistance data from all iterations before the prediction start point are integrated with the compact features obtained after global average pooling, and used as input to the LSTM network. The LSTM network layers use Hadamard product operations, as shown below:

[0085] f t =σ(W f ⊙[h t-1 x t ]+b f (5)

[0086] i t =σ(W i ⊙[h t-1 x t ]+b i (6)

[0087] g t =tanh(W g ⊙[h t-1 x t ]+b g (6)

[0088] C t =f t ⊙C t-1 +i t ⊙g t (7)

[0089] o t =σ(W Xo ⊙[h t-1 x t ]+b o (8)

[0090] h t =o t ⊙tanh(C t (9)

[0091] Figure 5 A schematic diagram of the LSTM network structure is given.

[0092] Step 6: Input the feature map obtained in the previous step along with the historical capacity data into the multilayer perceptron to predict the capacity decay trend after the battery life prediction start point.

[0093] Step 6 is implemented in the following steps:

[0094] Step 6.1: Integrate the LSTM output with the historical capacity data and record it as... As input to the multilayer sensor, the multilayer sensor generates a trend prediction result for the remaining usable capacity of the lithium-ion battery, as shown below:

[0095]

[0096] Where L(·) is the linear activation function. f(·) represents the Sigmoid function, as follows:

[0097]

[0098] Figure 6 The performance of the method disclosed in this invention in predicting capacity aging trends after the prediction starting point is given (for example, four batteries).

[0099] The core of the method provided in the above embodiments of the present invention is to take the battery capacity decay to 90% of the rated capacity as the prediction starting point for the remaining battery life, and use the battery charging voltage (V), current (I), temperature (T), resistance, and historical capacity data before the prediction starting point as HI; firstly, the hidden features in the VIT matrix are extracted by dilated convolution, which has a larger information receptive field than ordinary convolution; the high-dimensional feature mapping is reduced in dimensionality by global average pooling, and then a compact one-dimensional feature vector is generated by a multilayer perceptron; the compact feature and resistance data are input into an LSTM network to capture their long-term temporal dependency; finally, the above feature mapping and historical capacity data are input into the multilayer perceptron.

[0100] This invention provides a multi-scale fusion prediction method for the remaining lifespan of lithium-ion batteries. It comprehensively selects voltage (V), current (I), temperature (T), resistance, and historical capacity data during battery charging as HI. V, I, and T data are easy to collect and monitor online, while resistance and historical capacity data are strongly correlated with the capacity decay trend after the prediction starting point. Dilated convolution is used to extract features from V, I, and T data under the same cyclic state. Its receptive field, which grows exponentially with the number of network layers, enables multi-scale fusion of the three features. Unlike the traditional "input-model-output" baseline, the proposed method uses multi-head data input, with resistance and historical capacity data input separately with lag, greatly shortening the processing path between key features and output, and ensuring the inheritance of key information. The feature map after global average pooling is integrated with the resistance data and input into an LSTM to further extract temporal features. The inheritance of key features and the discarding of useless information are then accomplished by the LSTM. The historical capacity data is integrated with the LSTM output, and a two-layer perceptron is used to predict the capacity aging trend after the prediction starting point. At this point, the historical capacity data is very close to the model output, enabling capacity aging prediction based on the existing capacity data. LSTM and lag-based multi-head input overcome the long-term dependence and gradient explosion problems in long-sequence data prediction, contributing accurate prediction results.

[0101] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-scale fusion prediction method for remaining useful life of a lithium-ion battery, characterized in that, The method comprises the following steps: Step 1: Collect the charging current I, voltage V, temperature T, resistance and historical capacity data of each battery before failure from a plurality of battery capacity aging data sets; Step 2: Collect the charging current I, voltage V, temperature T, resistance and historical capacity data of each battery before the start of battery life prediction as HI, which measures the battery capacity aging, stack V, I, T through a matrix to form a three-dimensional matrix to meet the calculation requirements of the dilated convolution layer, and arrange the historical resistance and capacity data into a one-dimensional matrix vector; Step 3: Use a 3-layer dilated convolution network to perform two-dimensional convolution on the original V, I and T matrix respectively; Step 4: Input the obtained feature mapping into a global average pooling layer to reduce the dimension of the feature mapping; Step 5: Stack the feature sequence obtained in the previous step and the resistance data of the battery through matrix connection together as the input of the LSTM, capture the change rule of the compact features of V, I and T and the original resistance data with the number of cycles through the LSTM, model the long-term dependence relationship between them, and quantify the aging behavior of the battery; Step 6: Input the feature mapping obtained in the previous step and the historical capacity data into a multilayer perceptron to predict the capacity attenuation trend after the start of battery life prediction.

2. The multi-scale fusion prediction method for the remaining useful life of a lithium-ion battery according to claim 1, characterized in that: The start of battery life prediction is when the battery capacity decays to 90% of the rated capacity.

3. The multi-scale fusion prediction method for the remaining useful life of a lithium-ion battery according to claim 1, characterized in that: Step 2 includes interpolating the V, I and T data under each cycle by filling in missing values and Akima interpolation method.

4. The multi-scale fusion prediction method for the remaining useful life of a lithium-ion battery according to claim 1, characterized in that: The specific content of Step 3 is as follows: Let the output feature map of the nth th dilated convolutional layer be O n n = 1, 2,..., L, representing the number of dilated convolutional layers; The shape of the VIT matrix input into the dilated convolution layer is M n ×N n , the shape of the convolution kernel is P n ×Q n , wherein the weight matrix and the bias of the convolution kernel are W n and b k , respectively, and the dilation rate is (d h,n ,d w,n ). n th the kth th output feature map of the dilated convolution layer may be represented as: representing an element at position (i,j) in the matrix M; is an element at position (i,j) in the matrix M; th an element at position (i+a,j+b) in the rth input feature map in the receptive field R of the dilated convolution layer; th is an element at position (p,q) in the matrix M, corresponding to the weight of the kth convolution kernel of the nth layer; th th the kth convolution kernel of the nth layer;​​ Receptive field (R h,n+1 ×R w,n+1 ) size is calculated as follows:

5. The multi-scale fusion prediction method for the remaining useful life of a lithium-ion battery according to claim 1, characterized in that: The specific steps of Step 4 are as follows: Let the feature map after the dilated convolution layer be denoted as D n , and the shape is (w x h) where n = 1, 2, …, C, representing the cycle number of the battery. The process of global average pooling is represented as: d n then is the n-th element in the output compact feature sequence. th ​ 6. The multi-scale fusion prediction method for the remaining useful life of a lithium-ion battery according to claim 1, characterized in that: The specific steps of Step 5 are as follows: Integrate the resistance data under all cycles before the prediction start point and the compact features after global average pooling together as the input of the LSTM network as follows: f t = σ(W f ⊙ [h t-1 , x t ] + b f ); i t = σ(W i ⊙ [h t-1 , x t ] + b i ); g t = tanh(W g ⊙ [h t-1 , x t ] + b g ); C t = f t ⊙ C t-1 + i t ⊙ g t ; o t = σ(W Xo ⊙ [h t-1 , x t ] + b o ); h t = o t ⊙ tanh(C t ).

7. The multi-scale fusion prediction method for the remaining useful life of a lithium-ion battery according to claim 1, characterized in that: The specific steps of Step 6 are as follows: The output of the LSTM is integrated with the historical capacity data and denoted as As input to the multilayer perceptron, the lithium-ion battery remaining-use capacity trend prediction result is generated by the multilayer perceptron as follows: Where L(·) is a linear activation function, and f(·) represents a Sigmoid function as follows:

Images