Spacecraft lithium battery state of health estimation method based on physical information neural network

Abstract

The application provides a spacecraft lithium battery health state estimation method based on a physical information neural network, and has the characteristics that the following steps are included: a second-order hybrid equivalent circuit model of a battery monomer is established, and the relationship between the open circuit voltage of the equivalent circuit model and the state of charge of the battery is represented by a multilayer perceptron; the multilayer perceptron is trained by using the open circuit voltage and the state of charge of the battery in the collected sample data; the multilayer perceptron is used as a data driving layer of a physical neural network model, and the physical information neural network model is designed; the training of the physical neural network model is carried out by using the discharge curve in the collected sample data, and the estimation of the battery health state is realized by using the trained neural network model. The application does not need additional tests, and on the basis of using limited data to establish a physical information neural network model, the health state online estimation is accurately realized, and the stable operation of the spacecraft is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery state of health estimation, and particularly relates to a spacecraft lithium battery state of health estimation method based on a physical information neural network. BACKGROUND

[0002] Lithium-ion batteries have been widely used in various fields, including electric vehicles, mobile communications, power grids, spacecraft, and military industries, in recent years due to their high energy density, long cycle life, high safety, and low environmental pollution. The state of health (SOH) of a battery is defined as the ratio of the remaining available capacity to the rated capacity. When a lithium battery is in operation, irreversible chemical reactions occur inside the battery, causing the capacity to continuously decrease. When the battery capacity decreases to 80%, the lithium battery is considered to have reached the end of its life and should be replaced in a timely manner, otherwise it will affect the normal operation of the system and even cause accidents. In space applications, lithium-ion battery packs are the only source of energy for spacecraft in the Earth's shadow region, and monitoring their state of health is particularly important. However, the capacity of the battery is difficult to measure directly online, therefore, accurate estimation of SOH is an important guarantee for the reliable operation of spacecraft.

[0003] Existing SOH estimation methods are mainly divided into model-based methods and data-driven methods. The models mainly include equivalent circuit models and electrochemical models. Electrochemical models describe the dynamic changes inside the battery in detail, but their structure and calculations are complex and not suitable for online applications. The parameters in the equivalent circuit model change with cycling, and in space applications, parameters cannot be identified through additional testing, so they can only be estimated online through methods such as Kalman filtering and least squares. These methods require the state of the model to be fully excited, but the charging and discharging current and time interval of the spacecraft battery are usually constant, resulting in poor estimation results. Data-driven methods do not require accurate modeling, but they rely on a large amount of training data and are difficult to extrapolate to unobserved situations. The discharge depth of spacecraft lithium batteries is often shallow, resulting in limited available data, therefore, how to model and estimate the state of health of lithium batteries with limited telemetry data is a problem to be solved. SUMMARY

[0004] To solve the above problems, the present application provides a spacecraft lithium battery state of health estimation method based on a physical information neural network, which does not require additional testing and accurately estimates the state of health online based on the establishment of a physical information neural network model using limited data, ensuring the stable operation of spacecraft.

[0005] The technical solutions adopted by the present application are as follows:

[0006] A spacecraft lithium battery state of health estimation method based on a physical information neural network, comprising the following steps:

[0007] A second-order hybrid equivalent circuit model of the battery cell is established, and an open circuit voltage U oc The relationship between the open circuit voltage and the state of charge SOC of the battery is represented by a multilayer perceptron;

[0008] The multilayer perceptron is trained by using the open circuit voltage and the state of charge of the battery in the collected sample data;

[0009] The multilayer perceptron is used as a data-driven layer of a physical neural network model, and the physical information neural network model is designed;

[0010] The physical neural network model is trained by using the discharge curve in the collected sample data, and the trained neural network model is used to estimate the state of health of the battery.

[0011] Further, the open circuit voltage and the state of charge of the battery in the sample data are obtained by ground pulse testing, and the discharge curve in the sample data is a B0005 battery data set of the NASA Prognostics Center of Excellence.

[0012] Further, the second-order hybrid equivalent circuit model includes an ohmic resistance R0 and a second-order parallel RC network.

[0013] Further, the physical neural network model is divided into a physical layer, a data-driven layer and a sum layer; one output end of the physical layer is connected to an input end of the sum layer, and the other output end is connected to an input end of the data-driven layer; and an output end of the data-driven layer is connected to an input end of the sum layer.

[0014] Further, the input of the physical layer is a current x k , and the update formula is:

[0015] h k =g(W hh h k-1 +W hx x k +b h )

[0016]

[0017] wherein, and respectively represent the voltages of two parallel parts of the RC network, g(·) is an activation function, SOC k represents the state of charge of the battery, W hh and W hx represent weight coefficients, and b h is a set bias.

[0018] Further, the updating formula of the additive layer is:

[0019] y k = W yh h k + W yx x k + b y

[0020] The weight matrix is:

[0021] W yh = [-1 -1 f(SOC)], W yx = -R0

[0022] Wherein, f(SOC) is the output of the data driven layer, b y is the set deviation.

[0023] Further, the application normalizes the sample data before training; the multilayer perceptron is retrained during the training of the physical information neural network model.

[0024] Further, the specific process of estimating the state of health of the battery by using the trained neural network model is: using the parameter R0 in the trained neural network model to obtain the battery capacity C now , and estimating the state of health of the battery based on the battery capacity. Wherein C0 represents the rated capacity of the battery.

[0025] Further, in the equivalent circuit model of step two, the ohmic internal resistance is used to describe the transient response of the battery to direct current excitation, the second-order RC network is used to represent the polarization effect, and the open circuit voltage is obtained by using the state of charge through a multilayer perceptron.

[0026] Further, the cost function of the model is Wherein N is the number of samples, U i is the voltage value in the data set, is the output value of the physical information neural network model, and the training is stopped when the cost function is less than the set threshold or the training times reach the set upper limit.

[0027] Advantages

[0028] The application provides a spacecraft lithium battery state of health estimation method based on a physical information neural network.

[0029] Firstly, the application adopts a multilayer perceptron as a data-driven layer in a neural network, and through the open-circuit voltage in sample data and the sample data of the state of charge of the battery, the discharge curve in the sample data is used for training the neural network, so as to solve the identification difficulty of the pure physical model and the large data demand of the pure data-driven model.

[0030] Secondly, the application adopts the idea of transfer learning, first uses ground data to pre-train the hybrid model, and then uses limited running data to re-train the model in the network, so as to effectively solve the problem that the model cannot be trained and updated when the discharge depth is shallow and the data amount is small.

[0031] Thirdly, the second-order hybrid equivalent circuit model adopted by the application has the advantages of higher precision and more flexible adjustment than general equivalent circuit models, and can better capture the nonlinearity of the system.

[0032] Fourthly, the physical information neural network model established by the application has the advantages of data-driven models and physical models, the model conforms to physical knowledge and can conform to data distribution rules, and the dependence of the data-driven method on data is reduced. The model can not only describe and predict the change of the battery voltage, but also calculate the battery capacity according to the parameter identification result, so as to estimate the health state of the battery.

[0033] Fifthly, the identification result of the trained physical neural network model can be used for SOH estimation of the battery.

[0034] Sixthly, the application does not need additional tests, and on the basis of establishing a physical information neural network model using limited data, the health state is accurately estimated online, and the stable operation of the spacecraft is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The battery capacity curve diagram used for the specific embodiment is shown in the figure;

[0036] Figure 2 The second-order hybrid equivalent circuit model established in the method of the application is shown in the figure;

[0037] Figure 3 The overall operation principle diagram of the method provided by the application is shown in the figure;

[0038] Figure 4 The battery health state estimation result diagram is shown in the figure. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical scheme and advantages of the application more clear, the application will be further completely explained and described below in combination with the drawings. It should be noted that the embodiments described here are not all embodiments of the application.

[0040] As Figure 3 indicated, the application provides a spacecraft lithium battery health state estimation method based on a physical information neural network, which specifically comprises the following implementation steps:

[0041] Step one, obtain a data set, the obtained data contains two parts, one is 30 discharge curves before the end of battery life, and the other is the corresponding relationship data of open circuit voltage and state of charge.

[0042] The B0005 battery data set of the NASA Prognostics Center of Excellence is used in this example. The charge and discharge experiment steps of this data set are as follows: first, charge the battery in a constant current mode of 1.5 A until the battery voltage reaches 4.2 V; then continue to charge in a constant voltage mode until the battery current drops to 20 mA; then discharge the battery in a constant current mode of 2 A until the voltage drops to 2.7 V. Repeat the above charge and discharge cycle to accelerate battery aging.

[0043] Figure 1 is a curve graph of the capacity of battery No. 5 in the data set changing with cycles. The initial capacity of the battery is the rated capacity of 2 Ah, and it is considered to reach the end of life (EOL) when it decays to 1.4 Ah. In this example, 30 discharge curves before EOL are selected at equal time intervals for model training.

[0044] Since there is no pulse test result in the data set, the corresponding relationship between open circuit voltage and state of charge identified in existing research is used in this example, and this data is only used for pre-training of the multilayer perceptron in step two.

[0045] Step two, establish a second-order hybrid equivalent circuit model of the battery monomer.

[0046] The second-order hybrid equivalent circuit model established in this example is shown in Figure 2 The equivalent circuit model is composed of open circuit voltage U oc , ohmic resistance R0 and RC network. The ohmic resistance describes the transient response of the battery to direct current excitation. The RC network includes the polarization resistance and the polarization capacitance, which are used to describe the polarization effect of the battery. U oc is used to describe the static characteristics of the battery, and there is a nonlinear relationship between it and the state of charge (SOC) of the battery.

[0047] The discrete model can be represented by the following formula:

[0048]

[0049] Wherein, subscript k represents the sampling time, Δt represents the sampling interval, and respectively represent the voltage of the two parallel parts of the RC network, R1, R2 are polarization resistances, C1, C2 are polarization capacitances, τ1 = R1·C1, τ2 = R2·C2, I k is the current flowing through the battery monomer, η is the discharge efficiency of the battery, and C0 is the rated capacity of the battery. oc (SOC) represents the open-circuit voltage U oc and the state of charge (SOC) of the battery.

[0050] The present application describes U oc (SOC) using a six-layer multilayer perceptron, the input of the network is the state of charge, and the output is the open-circuit voltage. Compared with the relationship table and the polynomial, the multilayer perceptron can better capture the nonlinearity of the system, and the parameter fine-tuning is flexible, which is convenient for realizing the online update of the model. The multilayer perceptron is pre-trained using the corresponding relationship data between the open-circuit voltage of the battery and the charge state obtained in step one, as prior knowledge for subsequent training.

[0051] Step three, establishing a physical information neural network model of the battery.

[0052] According to the hybrid equivalent circuit model established in step two, the physical information neural network model proposed by the present application divides the recurrent neural network into a physical layer, a data-driven layer and a summation layer. The recurrent neural network is a kind of feedforward neural network suitable for processing time series data, in which all nodes are connected in a chain. The physical layer and the summation layer are updated according to the following mathematical formula:

[0053] h k =g(W hh h k-1 +W hx x k +b h ),

[0054] y k =W yh h k +W yx x k +b y ,

[0055] wherein, is the hidden layer state, g(·) is the activation function, x k = I k is the network input, b h and b y are biases, y k = U k+1 is the output, W hh , W hx and W yh , W yxare input and output weight matrix respectively. To build the physical information neural network model, the weight matrix and bias are given values according to the physical knowledge, so that the network is updated under the condition of meeting the physical law.

[0056] In this embodiment, let g(·) be the activation function is 1, b h and b y is the bias is 0, specifically, the input is the current x k =I k , RC terminal voltage and SOC as the state of the recurrent network and according to the following physical formula update value in the physical layer:

[0057] h k = W hh h k-1 + W hx x k ,

[0058] Among them, the designed input weight matrix is:

[0059]

[0060] The multilayer perceptron is used as the data driven layer, U oc =f(SOC), the data input is the SOC calculated after the physical layer k .

[0061] Finally, the sum layer uses the output of the physical layer and and the output of the data driven layer U oc to perform summation operation, the summation of the voltage is realized in the sum layer and output:

[0062] y k =U k = W yh h k + W yx x k ,

[0063] Among them, the designed output weight matrix is:

[0064] W yh =[-1 -1 f(SOC)], W yx =-R0

[0065] The parameters to be optimized in the network are R0, R1, R2, C1, C2, and the parameters in the multilayer perceptron f(SOC), which are corrected according to the error of the network output value and the true measured value.

[0066] The incorporation of the equivalent model imposes certain physical constraints on the training of the network, making the physical information neural network model not only conform to the data distribution law, but also have physical interpretability, while reducing the dependence of data-driven methods on data.

[0067] Step four, model training.

[0068] The current and voltage in the cyclic data set in step one are standardized and pretreated by the following formula, so that the results are mapped to the interval [-1, 1]:

[0069]

[0070] where x norm is the normalized data, x is the original data, x min and x max are the minimum and maximum values in the original data, respectively.

[0071] The standardized data set is input into the network. The network initialization parameters are set, the resistance and capacitance values use empirical values, and the multilayer perceptron uses the pre-training results in step two. The Adam optimizer is used to minimize the loss function where N is the number of samples, U i is the voltage value in the data set, and U is the output value of the physical information neural network model. The initial learning rate is set to 0.001, and the training period is 800. After training, the network weight parameters are saved, and the physical information neural network model of the battery is obtained.

[0072] Step five, estimate the state of health of the battery.

[0073] The parameter identification results of R0 in step four are labeled together with the true capacity of the battery in the figure, as shown in Figure 4 There is an obvious negative linear relationship between the two. Further, the Pearson correlation coefficient r is calculated by the following formula to reflect the linear relationship between the two from a statistical point of view:

[0074]

[0075] where x i and y i are the R0 and battery capacity values of each sampling point, and N = 30. The value of r is in the interval [-1, 1], r = 1 and r = -1 represent complete positive correlation and complete negative correlation between the two variables, respectively, and the closer r is to 0, the less obvious the linear relationship between the two variables. Generally, when |r| > 0.8, it is considered that the linear relationship between the two is relatively strong. After calculation, the Pearson correlation coefficient between the parameter identification results of R0 and the true capacity of the battery is -0.965, indicating a strong negative linear relationship.

[0076] Furthermore, delete a point that significantly deviates from the trend of other points (in... Figure 4 (In the box at the top right) the fitted linear function expression is y = -16.422x + 2.374. The fitted curve and 95% confidence interval are as follows: Figure 4 As shown. Therefore, the actual battery capacity can be linearly calculated based on the parameter identification results of R0, and based on... Estimate the battery's state of health, where C now This represents the current battery capacity.

[0077] To verify the effectiveness of this invention, the error of the estimation method proposed in this invention is calculated, such as... Figure 4 As shown in the bar chart, the mean absolute error (MAE) and root mean square error (RMSE) were calculated to be 0.03 Ah and 0.02 Ah, respectively, both less than 2%. This demonstrates that the method proposed in this invention is feasible and has high accuracy.

[0078] It should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, substitutions, variations, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for estimating the health status of spacecraft lithium batteries based on physical information neural networks, characterized in that, Includes the following steps: A second-order hybrid equivalent circuit model of a single battery cell is established, and the relationship between the open-circuit voltage of the equivalent circuit model and the state of charge of the battery is represented by a multilayer sensor. The multilayer sensor is trained using the open-circuit voltage and battery state of charge from the collected sample data. Using the multilayer perceptron as the data-driven layer of the physical neural network model, a physical information neural network model is designed. The discharge curves in the collected sample data are used to train a physical neural network model, and the trained neural network model is used to estimate the battery health status. The physical neural network model is divided into a physical layer, a data-driven layer, and an additive layer; wherein one output of the physical layer is connected to the input of the additive layer, the other output is connected to the input of the data-driven layer, and the output of the data-driven layer is connected to the input of the additive layer. The input to the physical layer is current x. k Its update formula is: h k =g(W hh h k-1 +W hx x k +b h ) in, and Let g(·) represent the voltages of the two parallel parts of the RC network, respectively, where g(·) is the activation function and SOC is the voltage. k W represents the state of charge of the battery. hh and W hx b represents the weighting coefficient. h To set the deviation; The update formula for the summation layer is: y k =W yh h k +W yx x k +b y The weight matrix is: W yh =[-1-1f(SOC)],W yx =-R0 Where f(SOC) is the output of the data-driven layer, b y To set the deviation.

2. The spacecraft lithium battery health status estimation method based on physical information neural network according to claim 1, characterized in that, The open-circuit voltage and state of charge of the battery in the sample data were obtained through ground pulse testing, and the discharge curves in the sample data are from the B0005 battery dataset of NASA's Center of Excellence for Prediction.

3. The spacecraft lithium battery health status estimation method based on physical information neural network according to claim 1, characterized in that, The second-order hybrid equivalent circuit model includes an ohmic resistor R0 and a second-order parallel RC network.

4. The spacecraft lithium battery health status estimation method based on physical information neural network according to claim 3, characterized in that, In the equivalent circuit model described in step two, the ohmic internal resistance is used to describe the transient response of the battery to DC excitation, the second-order RC network is used to characterize the polarization effect, and the open-circuit voltage is obtained by using the state of charge through a multilayer sensor.

5. The spacecraft lithium battery health status estimation method based on physical information neural network according to claim 1 or 2, characterized in that, Before training using the sample data, the sample data is normalized; during the training of the physical information neural network model, the multilayer perceptron is retrained.

6. The spacecraft lithium battery health status estimation method based on physical information neural network according to claim 1, characterized in that, The specific process of estimating battery health status using the trained neural network model is as follows: The battery capacity C is obtained using the parameters R0 in the trained neural network model. now Based on the battery capacity, estimate the battery's health status. C0 represents the rated capacity of the battery.

7. The spacecraft lithium battery health status estimation method based on physical information neural network according to claim 1, characterized in that, The cost function used during the training of the physical neural network model is: Where N is the sample size, U i These are the voltage values ​​in the dataset. This is the output value of the physical information neural network model. Training stops when the cost function is less than a set threshold or the number of training iterations reaches a set upper limit.

Images