A method and device for predicting capacity

Abstract

The present invention discloses a capacity prediction method and device, the method comprising: obtaining formation characteristic parameters of a lithium battery to be tested; the formation characteristic parameters include any one or more of voltage, current, capacity, energy and temperature at each time point during the formation period; performing data preprocessing on the formation characteristic parameters to obtain the interval in which the formation characteristic parameters are located; according to the interval in which the formation characteristic parameters are located, calling the corresponding interval capacity prediction model from the capacity prediction model group; the capacity prediction model group includes a number of trained capacity prediction models of different intervals; inputting the formation characteristic parameters into the corresponding interval capacity prediction model for prediction; and outputting the battery capacity of the lithium battery to be tested. The present invention can predict the final capacity of the battery without going through the capacity division process, thereby saving the time of the processes such as static and capacity division, saving equipment costs and plant area, and also improving production capacity.

Description

Technical Field

[0001] The present invention relates to the field of batteries, and in particular to a capacity prediction method and device. Background Art

[0002] In the lithium-ion battery production process, the capacity data in the battery grading process is an important indicator for evaluating the quality of single lithium-ion batteries before leaving the factory. Manufacturers will grade lithium-ion batteries according to different capacity levels, and batteries of the same level will be combined into battery packs to obtain better stability and life.

[0003] The commonly used method to obtain lithium-ion battery capacity data is to first fully charge the lithium-ion battery through charging and discharging equipment, and then perform multiple sets of step-by-step current discharges, that is, to fully discharge the battery, and finally add up the total capacity of each discharge step as the factory capacity of the battery.

[0004] The disadvantage of this method is that the entire charging and discharging process is extremely time-consuming and energy-consuming. Taking the 18650 battery with a standard capacity of 3200mAH as an example, the entire charging and discharging process will last 4 to 5 hours, and each charging and discharging stage will consume a lot of electricity.

[0005] In addition, there is a necessary formation process before the volume separation process.

[0006] After the lithium-ion battery is filled with electrolyte and sealed, its physical structure is basically assembled. In order to fully infiltrate the electrodes with electrolyte and fully activate them, it needs to go through an activation process with multiple constant voltage charge-discharge cycles as the main steps, and finally make the battery have discharge capacity. This electrochemical process is the formation process.

[0007] After formation is completed, the battery will generally be left at room temperature or high temperature for at least one day to allow its chemical properties to stabilize before undergoing a capacity separation process to test its capacity.

[0008] In summary, the formation process is one of the indispensable processes in the production of lithium-ion batteries, but the subsequent standing and capacity separation processes greatly extend the production time. At the same time, the capacity separation process consumes a lot of electricity, increasing the production cost of the battery.

[0009] In addition to the conventional capacity prediction process, some people have tried to combine artificial intelligence in the existing technology. For example, Xi'an Jiaotong University has developed a capacity prediction method based on deep neural network (Deep Neutral Network, DNN) ("Method, system, equipment and storage medium for online capacity prediction of lithium-ion batteries", CN113743661A). This method embeds the battery capacity and charge and discharge voltage curves, and then passes them to the fully connected layer after weighting by the attention mechanism, and finally outputs the capacity value. This method has the following problems:

[0010] 1) The data of the capacity division process (discharge voltage curve) is still used as training data. In essence, all batteries need to go through the discharge step, and capacity prediction no longer has the significance of reducing equipment and energy consumption;

[0011] 2) The Attention mechanism is a bright spot, but combined with the description of truncating the voltage curve into a sequence, the significance of introducing the Embedding layer should be to encode the voltage curve into a time series for use. However, the lengths of the last two vectors are both 300. Even the Transformer cannot be competent for weight memory of such a long sequence. If downsampling is used to shorten the sequence length (not described), then the introduction of the Attention mechanism is meaningless. This method seems to lack consideration in the architecture of the neural network. Summary of the invention

[0012] In order to solve the above technical problems, the present invention provides a capacity prediction method and device. Specifically, the technical solution of the present invention is as follows:

[0013] On the one hand, the present application discloses a capacity prediction method, comprising: obtaining formation characteristic parameters of a lithium battery to be tested; the formation characteristic parameters include any one or more of voltage, current, capacity, energy and temperature at each time point during the formation period; the capacity is the integral of the current over time; the energy is the integral of the voltage over time; performing data preprocessing on the formation characteristic parameters to obtain the interval in which the formation characteristic parameters are located; according to the interval in which the formation characteristic parameters are located, calling a corresponding interval capacity prediction model from a capacity prediction model group; the capacity prediction model group includes a number of trained capacity prediction models for different intervals; inputting the formation characteristic parameters into the corresponding interval capacity prediction model for prediction; and outputting the battery capacity of the lithium battery to be tested.

[0014] Preferably, the capacity prediction method also includes: training the interval capacity prediction model; specifically including: collecting formation characteristic parameter samples of lithium batteries from different batches; treating all the formation characteristic parameter samples of lithium batteries as a data population for normal distribution, and obtaining the distribution position of the formation characteristic parameter sample of each lithium battery in the data population; based on the normal distribution, dividing the data population into several intervals, and the data in different intervals are distributed in different positions in the data population; according to the distribution position of the formation characteristic parameter sample of each lithium battery in the data population, retrieving the formation characteristic parameter sample of each lithium battery to the interval sample training set of the corresponding interval; training different interval capacity prediction models through different interval sample training sets; the interval capacity prediction model corresponds one-to-one to the interval sample training set.

[0015] Further preferably, the formation characteristic parameter is subjected to data preprocessing to obtain the interval in which the formation characteristic parameter is located; specifically comprising: assuming that the formation characteristic parameter conforms to a normal distribution, calculating the standard score z of the formation characteristic parameter by the following formula;

[0016]

[0017] Wherein, x is the value corresponding to the formation characteristic parameter; μ is the average value of the sample data population; σ is the standard deviation of the sample data population;

[0018] According to the standard score of the formation characteristic parameter, an interval corresponding to the formation characteristic parameter is determined.

[0019] Preferably, each interval capacity prediction model includes an upstream CNN network and a downstream ∞-former network; wherein the CNN network is used to extract characteristic data from the formation characteristic parameters; the ∞-former network, whose input end is connected to the output end of the CNN network, is used to reduce the density of the formation characteristic data, and associate the formation characteristic data after the density reduction with the capacity of the lithium battery, and output the capacity of the lithium battery to be tested.

[0020] Further preferably, the CNN network comprises: several groups of sequentially connected convolutional pooling combination layers, a fully connected layer and an output layer; wherein the convolutional pooling combination layer comprises a convolutional layer and a pooling layer.

[0021] Further preferably, the CNN network comprises 2 groups of sequentially connected convolutional pooling combination layers; and the convolution layer is a 2-dimensional convolution layer with a kernel area of ​​3*3; the pooling layer is a maximum pooling layer with an area of ​​2*2; the fully connected layer includes 3 superimposed Dense layers, and the number of neurons thereof are 128, 64 and 32 respectively; the output layer includes 1 Dense layer, and the number of neurons thereof is 16.

[0022] On the other hand, the present application discloses a capacity prediction device, comprising: a data acquisition module, used to obtain formation characteristic parameters of a lithium battery to be tested; the formation characteristic parameters include any one or more of voltage, current, capacity, energy and temperature at each time point during the formation period; the capacity is the integral of the current over time; the energy is the integral of the voltage over time; a preprocessing module, used to perform data preprocessing on the formation characteristic parameters to obtain the interval in which the formation characteristic parameters are located; a model selection module, used to retrieve the corresponding interval capacity prediction model from a capacity prediction model group according to the interval in which the formation characteristic parameters are located; the capacity prediction model group includes a number of trained capacity prediction models for different intervals; a capacity prediction module, used to input the formation characteristic parameters into the corresponding interval capacity prediction model for prediction, and output the battery capacity of the lithium battery to be tested.

[0023] Preferably, the capacity prediction device also includes: a model training module, which is used to train each interval capacity prediction model; the model training module specifically includes: a sample collection submodule, which is used to collect formation characteristic parameter samples of lithium batteries from different batches; a sample processing submodule, which is used to treat all the formation characteristic parameter samples of lithium batteries as a data population for normal distribution, and obtain the distribution position of the formation characteristic parameter sample of each lithium battery in the data population; an interval division submodule, which is used to divide the data population into several intervals based on the normal distribution, and the data in different intervals are distributed in different positions in the data population; a sample retrieval submodule, which is used to retrieve the formation characteristic parameter sample of each lithium battery to the interval sample training set of the corresponding interval according to the distribution position of the formation characteristic parameter sample of each lithium battery in the data population; a training submodule, which is used to train different interval capacity prediction models through different interval sample training sets; the interval capacity prediction model corresponds one-to-one to the interval sample training set.

[0024] Further preferably, the preprocessing module comprises: a processing calculation submodule, for assuming that the formation characteristic parameter conforms to a normal distribution, and calculating the standard score z of the formation characteristic parameter by the following formula;

[0025]

[0026] Wherein, x is the value corresponding to the formation characteristic parameter; μ is the average value of the sample data population; σ is the standard deviation of the sample data population;

[0027] The interval determination submodule is used to determine the interval corresponding to the formation characteristic parameter according to the standard score of the formation characteristic parameter.

[0028] Preferably, each interval capacity prediction model includes an upstream CNN network and a downstream ∞-former network; wherein the CNN network is used to extract characteristic data from the formation characteristic parameters; the ∞-former network, whose input end is connected to the output end of the CNN network, is used to reduce the density of the formation characteristic data, and associate the formation characteristic data after the density reduction with the capacity of the lithium battery, and output the capacity of the lithium battery to be tested.

[0029] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0030] 1. This application can predict the final capacity of the battery without going through the capacity division process, thus saving the time of the standing and capacity division processes after formation, and at the same time saving the electric energy consumed by charging and discharging in the capacity division process. This method not only meets the engineering demand for lithium-ion battery capacity grouping, but also compresses the production time, greatly reduces production energy consumption, indirectly brings about an increase in production capacity, and also saves equipment costs and plant area.

[0031] 2. The capacity prediction model group of the present application includes several interval capacity prediction models. Based on the collected formation characteristic parameters of the lithium battery to be tested, the interval to which it belongs is determined, and then the corresponding interval capacity prediction model is called for prediction, so that the capacity prediction can be performed more accurately and it can also adapt well to possible batch effects.

[0032] 3. The interval capacity prediction model architecture of the present application is composed of an upstream CNN network and a downstream ∞-former. Generally, CNN networks are used for image recognition processing, but for data, it is often not thought of using CNN networks for feature extraction. The present application breaks through this conventional cognition, converts the collected formation feature parameters into a two-dimensional matrix of feature * time, thereby simulating the color value of each pixel of a single channel of a grayscale image, and then uses the CNN to extract feature information from the two-dimensional matrix data, which can efficiently utilize the two-dimensional structure of the input data. In addition, the downstream ∞-former can handle contexts of arbitrary length, and its attention complexity can be independent of the context length, so that predictions can be made based on the global perspective of the entire formation process, further improving accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The preferred implementation modes will be described below in a clear and understandable manner with reference to the accompanying drawings to further illustrate the above-mentioned characteristics, technical features, advantages and implementation methods of the present invention.

[0034] Figure 1 is a flow chart of an embodiment of a capacity prediction method of the present application;

[0035] Figure 2is a flow chart of another embodiment of a capacity prediction method of the present application;

[0036] Figure 3 is a schematic diagram of the model architecture of the interval capacity prediction model in this application;

[0037] Figure 4 is a structural block diagram of an embodiment of a capacity prediction device of the present application;

[0038] Figure 5 It is a structural block diagram of another embodiment of a capacity prediction device of the present application. DETAILED DESCRIPTION

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings and other implementation methods can be obtained based on these drawings without creative work.

[0040] In order to simplify the drawings, only the parts related to the invention are schematically shown in each figure, and they do not represent the actual structure of the product. In addition, in order to simplify the drawings and facilitate understanding, in some figures, only one of the parts with the same structure or function is schematically drawn or marked. In this article, "one" not only means "only one", but also means "more than one".

[0041] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0042] In this document, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0043] In addition, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0044] In one embodiment, the reference specification Figure 1 , a capacity prediction method provided by the present invention is as follows Figure 1As shown, including:

[0045] S100, obtaining formation characteristic parameters of the lithium battery to be tested; the formation characteristic parameters include any one or more of voltage, current, capacity, energy and temperature at each time point during the formation period; the capacity is the integral of the current over time; the energy is the integral of the voltage over time;

[0046] Specifically, the formation process is one of the indispensable processes in the production of lithium batteries. In this embodiment, sampling is mainly performed during the formation process of the lithium battery to be tested, and the formation characteristic parameters during the formation are collected. This formation characteristic parameter mainly refers to any one or more combinations of voltage, current, capacity, energy and temperature at each time point during the formation. Among them, capacity is the integral of current over time, the unit is mA·h, and energy is the integral of voltage over time, the unit is mW·h.

[0047] It is worth noting that the same type of feature parameters are obtained during the prediction process, as are the feature parameters used during the training process. For example, when we train the model, we use the voltage and current data of the training sample at each time point during the formation period as input. When the model is actually applied after the training is completed, the voltage and current time of the lithium battery to be tested at each time point during the formation period are also collected.

[0048] S200, performing data preprocessing on the formation characteristic parameter to obtain the interval in which the formation characteristic parameter is located;

[0049] Specifically, in the present embodiment, it is not a general user neural network model to predict, but a model group is designed, which contains several models, each model is responsible for the extraction and association of the formation characteristic data of different intervals, and then capacity prediction is carried out, so that the batch effect that occurs when the actual use can be well adapted. During model training, according to the position of sample distribution in the data population (the sample data population conforms to normal distribution), it is divided into several sample sets, wherein the first sample set is located at the center of the population, the second is second, and so on. Then the sample in each sample set is used to train the corresponding model, and these different interval capacity prediction models trained constitute the capacity prediction model group. When predicting capacity, it is only necessary to carry out normal distribution to the formation characteristic parameters of the lithium battery to be tested obtained, determine the interval where the formation characteristic parameters of the lithium battery to be tested are located in the sample population, so as to facilitate subsequent call corresponding interval capacity prediction model to predict.

[0050] S300, according to the interval in which the formation characteristic parameter is located, calling a corresponding interval capacity prediction model from a capacity prediction model group; the capacity prediction model group includes a number of trained capacity prediction models of different intervals;

[0051] Specifically, the capacity prediction of this embodiment is achieved through a model group, wherein the capacity prediction model group includes several interval capacity prediction models, and the model architecture of each interval capacity prediction model is completely consistent, except that the sample data set passed in during training is different. The samples trained by each interval capacity prediction model during training are all samples of the corresponding interval, which improves the robustness of the model. After the training is completed, when performing capacity prediction, it is only necessary to pre-process the formation characteristic parameters of the lithium battery to be tested, determine the interval it is in, and then call the corresponding interval capacity prediction model for prediction, thereby achieving accurate prediction and better adapting to the batch effect that occurs in the lithium battery production process.

[0052] S400, inputting the formation characteristic parameters into a corresponding interval capacity prediction model for prediction;

[0053] Specifically, according to the formation characteristic parameters, a matrix of characteristic data*sampling time is constructed, and the constructed matrix data is input into the interval capacity prediction model.

[0054] S500: outputting the battery capacity of the lithium battery to be tested.

[0055] This embodiment collects a certain amount of curve data (voltage, current, capacity and temperature at each time point) during the formation process, and obtains a neural network model group through deep learning method training. According to the data collected during the formation process, the corresponding interval capacity prediction model in the model group is retrieved, and the capacity of the battery can be directly calculated, eliminating the process of standing and dividing the capacity after formation. This method not only meets the engineering demand for lithium-ion battery capacity grouping, but also compresses production time, greatly reduces production energy consumption, and brings about an increase in production capacity in disguise.

[0056] Preferably, in step S400, the formation characteristic parameters are input into the corresponding interval capacity prediction model for prediction, which specifically includes: constructing a two-dimensional matrix of characteristic data*sampling time according to the formation characteristic parameters, and then inputting the corresponding interval capacity prediction model. The formation characteristic parameters refer to any one or more of the voltage data, current data, temperature data, energy data, etc. at each time point during the formation period, so based on the formation characteristic parameters, a two-dimensional matrix of specific characteristics*time can be constructed, such as a two-dimensional matrix of current*time, a two-dimensional matrix of voltage*time, etc., to simulate the color value of each pixel point of a single channel of the grayscale image.

[0057] Another embodiment of the present application, based on the previous embodiment, further includes training the interval capacity prediction model; the specific training process is as follows Figure 2 As shown, specifically including:

[0058] S010, collecting formation characteristic parameter samples of different batches of lithium batteries;

[0059] S020, taking all the formation characteristic parameter samples of lithium batteries as a data population for normal distribution, and obtaining the distribution position of each formation characteristic parameter sample of lithium batteries in the data population;

[0060] S030, based on the normal distribution, dividing the data population into a plurality of intervals, where data in different intervals are distributed in different positions in the data population;

[0061] S040, according to the distribution position of the formation characteristic parameter sample of each lithium battery in the data population, calling the formation characteristic parameter sample of each lithium battery to the interval sample training set of the corresponding interval;

[0062] Specifically, for example, if the interval to which the formation characteristic parameters of lithium battery A belong is the second interval, the formation characteristic parameters of lithium battery A are placed in the second interval and are labeled (actual battery capacity) for subsequent training.

[0063] S050, training different interval capacity prediction models through different interval sample training sets; the interval capacity prediction models correspond one to one with the interval sample training sets.

[0064] How to make a trained model well adaptable to the subsequent batch effect is a relatively cutting-edge domain generalization problem in the field of artificial intelligence. In order to improve the robustness of the model, we divide the samples into several data sets according to their distribution in the data population. The first data set is located in the center of the population and is the first interval sample training set; the second is the second interval sample training set, and so on. In the subsequent model group training, we use these interval sample training sets in a gradient manner.

[0065] If the training formation characteristic parameter samples contain more than one type of characteristic data, such as four types of characteristic data including voltage, current, energy and temperature, the method for determining whether the sample belongs to the nth interval sample training set is as follows:

[0066] For any given sample, arrange its features in a fixed order in the x-axis direction, and use its value as y, then a curve can be obtained in the XoY plane. The feature values ​​of all samples are presented in this way, and then the coordinate axis is rotated 90° counterclockwise with the origin as the center, then the corresponding points on each curve conform to the normal distribution, and the samples can be divided into several interval sample training sets in turn.

[0067] Preferably, the batch size during model training is 131,072, RAdam is used as the optimizer, the starting learning rate is 1e-2, beta_1 is 0.458, beta_2 is 0.347, and weight_decay of 0.01 is applied.

[0068] In order to ensure that the capacity prediction model for each interval has stable and excellent performance in the production environment, a self-evaluation step is added after the model training step:

[0069] S060, obtaining a predicted value and a true value of a lithium battery sample predicted by an interval capacity prediction model;

[0070] S070, based on the battery barcode, predicted value and true value of the lithium battery sample, the number of samples with different error levels is displayed through a histogram or the distribution difference between the true capacity and the predicted capacity is displayed through a probability density curve.

[0071] In one embodiment, the step S200 performs data preprocessing on the formation characteristic parameters to obtain the interval in which the formation characteristic parameters are located; specifically includes:

[0072] S210, assuming that the formation characteristic parameter conforms to a normal distribution, calculating a standard score z of the formation characteristic parameter by the following formula;

[0073]

[0074] Wherein, x is the value corresponding to the formation characteristic parameter; μ is the average value of the sample data population; σ is the standard deviation of the sample data population;

[0075] S210, determining an interval corresponding to the formation characteristic parameter according to a standard score of the formation characteristic parameter.

[0076] Specifically, in general, for a single type of feature (formation characteristic parameter), the population formed by all the sampled data of the type of formation characteristic parameter of the lithium battery to be tested conforms to the normal distribution. Based on this premise, in the method of this embodiment, we use the value of the standard score z to describe the distribution position of the sample in the population. The larger the absolute value of z, the farther we think it is from the center of the sample population.

[0077] For multi-class features (chemical characteristic parameters), the method for determining whether the input chemical characteristic parameters belong to the nth data set (corresponding to the nth interval) is:

[0078] For any given feature parameter, arrange it in a fixed order in the x-axis direction, and use its value as y, then a curve can be obtained in the XoY plane. The feature values ​​of all samples are presented in this way, and then the coordinate axis is rotated 90° counterclockwise with the origin as the center, then the corresponding points on each curve conform to the normal distribution, and the samples can be divided into several data sets in turn.

[0079] In order to improve the robustness of the model, we divide the samples into several data sets according to their distribution in the data population, for example, into 8 subsets, where the first subset is located in the center of the population, the second is second, and so on. In the subsequent model group training, we use these subsets (interval sample training sets) in a gradient manner, that is, one interval capacity prediction model corresponds to one interval sample training set, and the corresponding interval capacity prediction model is trained through the interval sample training set.

[0080] In one embodiment, based on any of the above embodiments, the capacity prediction model group is mainly composed of several interval capacity prediction models. The model architecture of each interval capacity prediction model is the same, but the training set used during training is different. Each interval capacity prediction model includes an upstream CNN network and a downstream ∞-former network; wherein,

[0081] The CNN network is used to extract characteristic data from the formation characteristic parameters;

[0082] CNN neurons can extract information from pixels within a certain coverage area of ​​an image, and perform well in large image processing. This network structure generally consists of one or more convolutional layers and a fully connected layer (FC layer) at the top, as well as associated weights and pooling layers. This enables CNN to efficiently utilize the two-dimensional structure of the input data.

[0083] Inspired by the characteristics of CNN, in this application proposal, feature×time is laid out as a two-dimensional matrix to simulate the color value of each pixel in a single channel of a grayscale image.

[0084] The ∞-former network, whose input end is connected to the output end of the CNN network, is used to reduce the density of the formation characteristic data, associate the formation characteristic data after the density reduction with the capacity of the lithium battery, and output the capacity of the lithium battery to be tested.

[0085] The ∞-former is a Transformer model with unlimited long-time memory (LTM) that can handle contexts of arbitrary length. By utilizing a continuous spatial attention mechanism to handle long-term memory, the attention complexity of the ∞-former can be independent of the context length. Therefore, it is able to model contexts of arbitrary length and maintain "sticky memories" with a fixed computational cost. With this capability, our model can ultimately make predictions based on a global view of the entire formation process, further improving accuracy and robustness.

[0086] This application does not use the traditional transform model structure, but uses ∞-former, because the memory of the traditional transform is improved compared to the RNN network, but it is still not good enough. In this application, if the collected transformation feature parameters are relatively large, or the sampling frequency is relatively high, the density of the input transformation feature parameters will also be relatively large, or the data will be relatively long. If the transform architecture is used, the previous weights may not be remembered due to poor memory. With the ∞-former architecture, even if the input data is very long, after learning, although the weights will decay a little, the previous weights will still be remembered based on the global perspective.

[0087] In another embodiment of the present application, based on the above embodiment, the above-mentioned CNN network includes: several groups of sequentially connected convolutional pooling combination layers, fully connected layers and output layers; wherein the convolutional pooling combination layers include convolutional layers and pooling layers. The kernel area product of the convolutional layer is m*m; m is an odd number less than or equal to 9. Because if an even number * an even number is used, the final convolution may leave a column empty, and thus a complete convolution cannot be formed, an odd number * an odd number is generally used as the kernel area of ​​the convolutional layer. The pooling layer generally uses a maximum value pooling layer with an area of ​​n*n; wherein n is an even number less than 8. Generally, the CNN network uses 2 or 3 sequentially connected convolutional pooling combination layers, and then connected to subsequent fully connected layers and output layers.

[0088] Preferably, the convolutional pooling combination layer also includes a BN layer (Batch Normalization. After adding the BN layer, the distribution of the input value of any neuron in each layer of the neural network can be forcibly pulled back to a standard normal distribution with a mean of 0 and a variance of 1. In fact, it is to force the increasingly biased distribution back to a more standard distribution, so that the activation input value falls in an area where the nonlinear function is more sensitive to the input, so that a small change in the input will lead to a larger change in the loss function, which means that the gradient becomes larger and the gradient vanishing problem is avoided. Moreover, a larger gradient means a faster learning convergence speed, which can greatly speed up the training speed.

[0089] In another embodiment of the present application, the obtained formation characteristic parameters are arranged as a two-dimensional matrix of feature×time as input to simulate the color value of each pixel of a single channel of a grayscale image. The structure of the neural network is as follows Figure 3 As shown:

[0090] (1) 2D convolutional layer with a kernel area of ​​3×3;

[0091] (2) A maximum pooling layer with an area of ​​2×2;

[0092] (3) Batch Normalization;

[0093] (4) 2D convolutional layer with a kernel area of ​​3×3;

[0094] (5) Maximum pooling layer with an area of ​​2×2;

[0095] (6) Fully connected layer (3 Dense layers stacked, with 128, 64, and 32 neurons respectively);

[0096] (7) Output layer (Dense layer, 16 neurons).

[0097] The ∞-former takes the output of CNN as input, and the loss function of the final model is the mean absolute percentage error.

[0098] The loss function loss calculation formula is as follows:

[0099]

[0100] Among them, y_true is the true value and y_pred is the predicted value.

[0101] Based on the above CNN+∞-former model architecture, several models are trained using several different interval sample training sets mentioned above to form a model group. For example, 8 interval sample training sets are used to train 8 interval capacity prediction models. When predicting, first determine which interval the new formation data is in based on the existing data set, and then call the specific model for prediction.

[0102] Based on the same technical concept, another embodiment of the present application discloses a capacity prediction device, such as Figure 4 As shown, including:

[0103] The data acquisition module 10 is used to acquire the formation characteristic parameters of the lithium battery to be tested; the formation characteristic parameters include any one or more of the voltage, current, capacity, energy and temperature at each time point during the formation period; the capacity is the integral of the current over time; the energy is the integral of the voltage over time;

[0104] A preprocessing module 20, used for performing data preprocessing on the formation characteristic parameters to obtain the interval in which the formation characteristic parameters are located;

[0105] A model selection module 30 is used to retrieve a corresponding interval capacity prediction model from a capacity prediction model group according to the interval in which the formation characteristic parameter is located; the capacity prediction model group includes a number of trained capacity prediction models of different intervals;

[0106] The capacity prediction module 40 is used to input the formation characteristic parameters into the corresponding interval capacity prediction model for prediction, and output the battery capacity of the lithium battery to be tested.

[0107] By using the capacity prediction device of this embodiment, the final capacity of the battery can be directly predicted through the trained capacity prediction model group only by using the data collected in the formation process during the battery production process. The final capacity of the battery can be predicted without going through the capacity division process, thereby saving the time of the standing and capacity division processes, and at the same time saving the electrical energy consumed by charging and discharging in the capacity division process.

[0108] In another embodiment of the device of the present application, as shown in the figure, the capacity prediction device further includes:

[0109] The model training module 50 is used to train the capacity prediction model of each interval; the model training module specifically includes:

[0110] The sample collection submodule 51 is used to collect formation characteristic parameter samples of different batches of lithium batteries;

[0111] The sample processing submodule 52 is used to treat all the formation characteristic parameter samples of lithium batteries as a data population for normal distribution, and obtain the distribution position of each formation characteristic parameter sample of lithium batteries in the data population;

[0112] An interval division submodule 53 is used to divide the data population into a plurality of intervals based on the normal distribution, where data in different intervals are distributed in different positions in the data population;

[0113] The sample retrieval submodule 54 is used to retrieve the formation characteristic parameter sample of each lithium battery to the interval sample training set of the corresponding interval according to the distribution position of the formation characteristic parameter sample of each lithium battery in the data population;

[0114] The training submodule 55 is used to train different interval capacity prediction models through different interval sample training sets; the interval capacity prediction models correspond one to one with the interval sample training sets.

[0115] Preferably, the pre-processing module 20 includes:

[0116] The processing and calculation submodule 21 is used to assume that the formation characteristic parameter conforms to a normal distribution and calculate the standard score z of the formation characteristic parameter by the following formula;

[0117]

[0118] Wherein, x is the value corresponding to the formation characteristic parameter; μ is the average value of the sample data population; σ is the standard deviation of the sample data population;

[0119] The interval determination submodule 22 is used to determine the interval corresponding to the formation characteristic parameter according to the standard score of the formation characteristic parameter.

[0120] In another embodiment of the device of the present application, based on any of the above device embodiments, the architecture of each interval capacity prediction model is completely consistent, and only the interval training sample set passed in during training is different. Taking a single interval capacity prediction model as an example, the model consists of two parts: a CNN (convolutional neural network) model for abstracting features and a downstream ∞-former.

[0121] The CNN network is used to extract characteristic data from the formation characteristic parameters;

[0122] The network structure of CNN generally consists of one or more convolutional layers, a fully connected layer (FC layer) at the top, as well as associated weights and a pooling layer, etc. This enables CNN to efficiently utilize the two-dimensional structure of the input data.

[0123] The ∞-former network, whose input end is connected to the output end of the CNN network, is used to reduce the density of the formation characteristic data, associate the formation characteristic data after the density reduction with the capacity of the lithium battery, and output the capacity of the lithium battery to be tested.

[0124] In this application proposal, feature×time is laid out as a two-dimensional matrix to simulate the color value of each pixel in a single channel of a grayscale image. The structure of the neural network is:

[0125] 1) 2D convolutional layer with a kernel area of ​​3×3

[0126] 2) Maximum pooling layer with an area of ​​2×2

[0127] 3) Batch Normalization

[0128] 4) 2D convolutional layer with a kernel area of ​​3×3

[0129] 5) Maximum pooling layer with an area of ​​2×2

[0130] 6) Fully connected layer (3 Dense layers stacked, with 128, 64 and 32 neurons respectively)

[0131] 7) Output layer (Dense layer, 16 neurons)

[0132] The ∞-former is a Transformer model with unlimited long-time memory (LTM) that can handle contexts of arbitrary length. By utilizing a continuous spatial attention mechanism to handle long-term memory, the attention complexity of the ∞-former can be independent of the context length. Therefore, it is able to model contexts of arbitrary length and maintain "sticky memories" with a fixed computational cost. With this capability, our model can ultimately make predictions based on a global view of the entire formation process, further improving accuracy and robustness.

[0133] In this application proposal, the ∞-former takes the output of CNN as input, and the loss function of the final model is the mean absolute percentage error.

[0134] Mean absolute percentage error:

[0135]

[0136] Among them, y_true is the true value and y_pred is the predicted value.

[0137] In this application proposal, based on the above CNN+∞-former model architecture, 8 models were trained using the 8 different data subsets mentioned above to form a model group. When predicting, first determine in which interval the new formation data is located based on the existing data set, and then call the specific model for prediction.

[0138] It is worth noting that the device embodiment of the present application corresponds to the method embodiment of the present application, and the technical details of the method embodiment of the present application are also applicable to the device embodiment of the present application. In order to reduce repetition, they are not repeated here.

[0139] It should be noted that the above embodiments can be freely combined as needed. The above are only preferred embodiments of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A capacity prediction method, characterized in that: include: Obtaining formation characteristic parameters of the lithium battery to be tested; The formation characteristic parameters include any one or more of the voltage, current, capacity, energy and temperature at each time point during the formation; the capacity is the integral of the current over time; the energy is the integral of the voltage over time; Performing data preprocessing on the formation characteristic parameters to obtain the interval in which the formation characteristic parameters are located; According to the interval in which the formation characteristic parameter is located, a corresponding interval capacity prediction model is retrieved from a capacity prediction model group; the capacity prediction model group includes a number of trained capacity prediction models of different intervals; Inputting the formation characteristic parameters into a corresponding interval capacity prediction model for prediction; Outputting the battery capacity of the lithium battery to be tested; Training the interval capacity prediction model; specifically including: Collect formation characteristic parameter samples of different batches of lithium batteries; All the formation characteristic parameter samples of lithium batteries are normally distributed as a data population, and the distribution position of each formation characteristic parameter sample of lithium batteries in the data population is obtained; Based on the normal distribution, the data population is divided into a plurality of intervals, and data in different intervals are distributed in different positions in the data population; According to the distribution position of the formation characteristic parameter sample of each lithium battery in the data population, the formation characteristic parameter sample of each lithium battery is retrieved to the interval sample training set of the corresponding interval; Different interval capacity prediction models are trained through different interval sample training sets; the interval capacity prediction models correspond one to one with the interval sample training sets.

2. A capacity prediction method according to claim 1, characterized in that: Performing data preprocessing on the formation characteristic parameters to obtain the interval in which the formation characteristic parameters are located; comprising: Assuming that the formation characteristic parameter conforms to a normal distribution, the standard score z of the formation characteristic parameter is calculated by the following formula: Wherein, x is the value corresponding to the formation characteristic parameter; μ is the average value of the sample data population; σ is the standard deviation of the sample data population; According to the standard score of the formation characteristic parameter, an interval corresponding to the formation characteristic parameter is determined.

3. A capacity prediction method according to claim 1, characterized in that: Each interval capacity prediction model includes an upstream CNN network and a downstream ∞-former network; The CNN network is used to extract characteristic data from the formation characteristic parameters; The ∞-former network, whose input end is connected to the output end of the CNN network, is used to reduce the density of the formation characteristic parameters, associate the formation characteristic parameters after the density reduction with the capacity of the lithium battery, and output the capacity of the lithium battery to be tested.

4. A capacity prediction method according to claim 3, characterized in that: The CNN network includes: a plurality of groups of sequentially connected convolutional pooling combination layers, a fully connected layer and an output layer; wherein the convolutional pooling combination layer includes a convolutional layer and a pooling layer.

5. A capacity prediction method according to claim 4, characterized in that: The CNN network comprises two groups of sequentially connected convolutional pooling combination layers; the convolutional layer is a 2-dimensional convolutional layer with a kernel area of ​​3*3; the pooling layer is a maximum pooling layer with an area of ​​2*2; The fully connected layer includes three superimposed Dense layers, the number of neurons of which are 128, 64 and 32 respectively; The output layer includes 1 Dense layer, and the number of neurons is 16.

6. A capacity prediction device, characterized in that: include: A data acquisition module is used to obtain the formation characteristic parameters of the lithium battery to be tested; The formation characteristic parameters include any one or more of the voltage, current, capacity, energy and temperature at each time point during the formation; the capacity is the integral of the current over time; the energy is the integral of the voltage over time; A preprocessing module, used for performing data preprocessing on the formation characteristic parameters to obtain the interval in which the formation characteristic parameters are located; A model selection module, used to retrieve a corresponding interval capacity prediction model from a capacity prediction model group according to the interval in which the formation characteristic parameter is located; the capacity prediction model group includes a number of trained capacity prediction models of different intervals; A capacity prediction module, used for inputting the formation characteristic parameters into a corresponding interval capacity prediction model for prediction, and outputting the battery capacity of the lithium battery to be tested; A model training module is used to train the capacity prediction model of each interval; the model training module specifically includes: The sample collection submodule is used to collect formation characteristic parameter samples of different batches of lithium batteries; The sample processing submodule is used to treat all the formation characteristic parameter samples of lithium batteries as a data population for normal distribution, and obtain the distribution position of each formation characteristic parameter sample of lithium batteries in the data population; An interval division submodule, for dividing the data population into a plurality of intervals based on the normal distribution, wherein data in different intervals are distributed in different positions in the data population; A sample retrieval submodule is used to retrieve the formation characteristic parameter sample of each lithium battery to the interval sample training set of the corresponding interval according to the distribution position of the formation characteristic parameter sample of each lithium battery in the data population; The training submodule is used to train different interval capacity prediction models through different interval sample training sets; the interval capacity prediction models correspond one to one with the interval sample training sets.

7. A capacity prediction device according to claim 6, characterized in that: The preprocessing module comprises: A processing and calculation submodule, used for assuming that the formation characteristic parameter conforms to a normal distribution, and calculating a standard score z of the formation characteristic parameter by the following formula; Wherein, x is the value corresponding to the formation characteristic parameter; μ is the average value of the sample data population; σ is the standard deviation of the sample data population; The interval determination submodule is used to determine the interval corresponding to the formation characteristic parameter according to the standard score of the formation characteristic parameter.

8. A capacity prediction device according to any one of claims 6 to 7, characterized in that: Each interval capacity prediction model includes an upstream CNN network and a downstream ∞-former network; The CNN network is used to extract characteristic data from the formation characteristic parameters; The ∞-former network, whose input end is connected to the output end of the CNN network, is used to reduce the density of the formation characteristic parameters, and associate the formation characteristic parameters after the density reduction with the capacity of the lithium battery, and output the capacity of the lithium battery to be tested.

Images