CN110658462B - An online life prediction method of lithium battery based on data fusion and ARIMA model - Google Patents

Abstract

The invention discloses an online life prediction method for lithium batteries based on data fusion and ARIMA model. The steps are as follows: step 1: collecting data of equal voltage discharge, charging time interval and battery capacity of the lithium battery; step 2: calculating the weight of data layer fusion And fuse the data; Step 3: Train and estimate the parameters of the ARIMA model and test the ARIMA model; Step 4: Use the ARIMA model to predict the RUL and SOH of the next cycle through the data fused in Step 2; Step 5: Input the battery obtained online in real time For state observation data, repeat steps 2 to 4 to update the ARIMA prediction model to realize online prediction. The invention realizes the lithium battery online life prediction based on data fusion and ARIMA model, improves the life prediction accuracy of the ARIMA model, realizes the online life prediction function of the lithium battery, and completes the reliability analysis of the aerospace device lithium battery.

Description

An online life prediction method of lithium battery based on data fusion and ARIMA model

technical field

The design of the invention belongs to the field of life prediction and health management (Prognosis and Health Management, PHM), and specifically relates to a lithium battery online life prediction method based on data fusion and ARIMA model.

Background technique

Compared with other energy storage devices, lithium batteries have the advantages of high energy density, light weight, and stable discharge. Lithium batteries are used as energy storage devices for spacecraft, which can improve the storage efficiency and stability of spacecraft energy systems, increase payload and reduce launch costs. Due to the complexity, large-scale and intelligence of spacecraft equipment, and special working environment, post-event maintenance and regular maintenance can no longer meet the requirements. At present, the technical means to maintain the reliable operation of spacecraft mainly rely on the health status assessment technology and prediction technology.

The traditional life prediction method uses the life data of product samples, and fits the failure time of each sample through the life distribution model to obtain the life distribution estimate of the product. However, for spacecraft lithium batteries with degraded failure modes and long service life, generally only very little failure data can be obtained in life tests, which makes traditional life prediction methods no longer effective for spacecraft lithium battery health management. However, spacecraft components have a large amount of available condition monitoring data and sensor historical data, so data-driven life prediction methods and models have become a research hotspot in the field of spacecraft product health state management and prediction technology. Data-driven life prediction methods can be further divided into statistical regression analysis methods and artificial intelligence analysis methods.

The Autoregressive Integrated Moving Average model (ARIMA) is one of the commonly used statistical regression methods. It uses the time series historical moment and the observation value of noise to perform linear weighting, and then uses the statistical information criterion to estimate the parameters of the ARIMA model. get the best estimation model. The ARIMA model has a simple prediction structure and only uses the observations of its own time series, and no additional auxiliary variables are required for prediction. The traditional ARIMA model cannot capture nonlinear changes, and the prediction accuracy of nonlinear change data is not high. prediction error.

Information fusion technology is a comprehensive information processing technology that utilizes the redundancy and complementarity of multiple sensor information to obtain more reliable, more accurate and more efficient information processing effects than a single sensor through information fusion algorithms. Information fusion technology can appear at different information layers, including data layer fusion, feature layer fusion and decision layer fusion.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a lithium battery online life prediction method based on data fusion and ARIMA model, using the data layer fusion method to improve the accuracy of lithium battery life prediction using the ARIMA model alone, and by updating the ARIMA model online, improve The online life prediction accuracy of the ARIMA model can complete the prediction of the health status and remaining service life of the spacecraft lithium battery, realize the real-time tracking of the battery change trend, and enhance the reliability of the later lithium battery life prediction.

In order to achieve the above object, the present invention provides an online life prediction method of lithium battery based on data fusion and ARIMA model. The invention first collects the equal-voltage discharge time interval and equal-voltage charging time interval of the lithium battery as the indirect health factor input for online prediction, and then uses the correlation coefficient between the two data and the battery capacity as the weight to perform information fusion at the data layer. , and finally the fused health factor is predicted online through the ARIMA model, and the predicted value of the remaining service life and health state is obtained. The invention effectively improves the prediction accuracy of the ARIMA model and realizes the online life prediction of the spacecraft lithium battery.

The present invention is a lithium battery online life prediction method based on data fusion and ARIMA model, and its implementation steps are as follows:

Step 1: Collect the lithium battery's iso-voltage discharge, charging time interval and battery capacity data;

Step 2: Calculate the weight of data layer fusion and fuse the data;

Step 3: Train and estimate ARIMA model parameters and test the ARIMA model;

Step 4: Use the ARIMA model to predict the RUL and the SOH of the next cycle through the data fused in Step 2;

Step 5: Input the battery state observation data obtained online in real time, repeat steps 2 to 4, update the ARIMA prediction model, and realize online prediction.

Among them, in step 1, the method of "collecting the data of equal voltage discharge, charging time interval and battery capacity of lithium battery" is as follows:

与

i＝1,2,…,K，计算出等电压放电时间间隔序列，记为

采集锂电池前K个充电周期中电压从V₃充到V₄的时刻值，记为

与

i＝1,2,…,K，计算出等电压充电时间间隔序列，记为

采集锂电池前K个充放电周期的电池容量序列，记为{C_i|i＝ 1,2,...,K}，并采集剩余充放电周期的电池容量序列作为验证集，记为{C_i|i＝K+ 1,K+2,...,N}，N是所有充放电周期试验总数。The time value of the voltage from V ₁ to V ₂ in the K discharge cycles before the collection of lithium batteries is recorded as

and

i=1,2,...,K, calculate the time interval sequence of equal voltage discharge, denoted as

The value of the moment when the voltage is charged from V ₃ to V ₄ in the first K charging cycles of the lithium battery is recorded as

and

i=1,2,...,K, calculate the time interval sequence of equal voltage charging, denoted as

Collect the battery capacity sequence of the first K charge and discharge cycles of the lithium battery, denoted as {C _i |i = 1,2,...,K}, and collect the battery capacity sequence of the remaining charge and discharge cycles as the verification set, denoted as { C _i |i=K+1,K+2,...,N}, N is the total number of all charge-discharge cycle tests.

Among them, the method of "calculating the weight of data layer fusion and merging the data" described in step 2 is as follows:

S21. Calculate the Pearson Correlation Coefficient (Pearson Correlation Coefficient) of the equal-voltage charging and discharging time interval obtained in step 1 as the online indirect health factor and the battery capacity as the offline direct health factor;

S22, normalize the charging and discharging time interval data of the same voltage;

S23. Use the correlation coefficient calculated in step S21 to perform a weighted average on the equal-voltage charging and discharging time interval data normalized in step S22 to obtain input training data of the ARIMA model.

Among them, the "training and estimating ARIMA model parameters and testing the ARIMA model" described in step 3 is as follows:

S31. Perform stationarity judgment on the input training data after the fusion of the data layer in step 2. If the sequence is not stationary, it is necessary to perform d-order difference on the training data to confirm the parameter d of the ARIMA model;

S32. Determine the optimization range of the other two parameters p and q of the ARIMA model, including the order p of the autoregression and the order q of the moving average, and train the ARIMA model under each parameter;

S33, using the Bayesian Information Criterion (Bayesian Information Criterion, BIC) to estimate the optimal ARIMA model parameters;

S34. Check whether the parameters of the ARIMA model are significant by calculating the T-Statistic of the parameters of the ARIMA model.

Among them, the method of "predicting the RUL and the SOH of the next cycle through the ARIMA model of the fused data" described in step 4 is as follows:

S41. Obtain the optimal ARIMA model trained through the model parameters estimated in step 3, and perform autoregressive moving average operation on the fusion data in step 2 to predict the remaining service life of the battery and the state of health of the battery in the next test cycle .

S42. Normalize the untrained battery capacity sequence {C _i |i=K+1, K+2, . Validation dataset.

S43. Calculate the absolute error (Absolute Error) between the estimated value of the remaining service life prediction each time and the verification data value in S42, so as to measure the accuracy and effect of the online real-time prediction.

Among them, in step 5, "input the battery state observation data of the new test cycle, repeat steps 2 to 4, update the ARIMA prediction model, and realize online prediction", the method is as follows:

S51. Input the i-th (i=K+1, K+2, . The observations of i-1 trial periods constitute the new training dataset.

S52. Steps 2 to 4 will be repeated every time the data of one test cycle is input, and the parameters of the ARIMA model will be continuously updated online until the data processing of all test cycles is completed, and finally the battery health of N-K test cycles will be obtained. State predicted value.

S53, calculating the root mean square error (Root Mean Square Error, RMSE) between the prediction result and the verification data, as an index for judging the quality of the prediction effect.

Through the above steps, the present invention realizes the lithium battery online life prediction based on data fusion and ARIMA model, improves the life prediction accuracy of the ARIMA model, realizes the online life prediction function of the lithium battery, and completes the reliability analysis of the aerospace device lithium battery .

According to the design of the present invention, the present invention realizes the lithium battery online life prediction based on data fusion and ARIMA model, the algorithm parameter identification method is simple, the complexity is low, and the continuous update of the ARIMA model is easily realized.

According to the design of the present invention, the present invention comprehensively utilizes the observation data in the battery degradation model through the data layer fusion method, effectively improving the accuracy of the ARIMA model using a single information source to predict the remaining service life, and improving the accuracy of the reliability analysis of the spacecraft lithium battery. and robustness (see Figure 1 for the prediction results).

According to the design of the present invention, the present invention extracts the indirect health factor that can characterize the state of health of the battery in the lithium battery degradation model online, predicts the state of health value of the next cycle each time, and continuously updates the ARIMA prediction model to realize the real-time change of the state of health of the battery. According to the observation, the prediction accuracy of the online life state of the lithium battery is improved (see Figure 2 for the prediction results).

Description of drawings

Figure 1 is a comparison chart of the RUL prediction curves of the ARIMA model using different data information in Cycle 60.

Figure 2 is a graph comparing the curves of the traditional ARIMA model and the improved ARIMA model that both use the fusion information of the data layer to predict the battery state of health at Cycle60.

FIG. 3 is a flow chart of the lithium battery online life prediction method based on data layer fusion and ARIMA model proposed by the present invention.

Figure 4 is a curve of the discharge voltage of B0005 battery Cycle 1 with time.

Figure 5 is the B0005 battery equal-voltage discharge time interval curve.

Figure 6 is the time curve of the charging voltage of the B0005 battery Cycle 1.

Fig. 7 is the time interval curve of constant voltage charging of B0005 battery.

Figure 8 is the battery capacity time curve of the B0005 battery.

Detailed ways

In order to have a further understanding and understanding of the features, purposes and functions of the present invention, the present invention will now be described in more detail with reference to specific embodiments and accompanying drawings.

As shown in FIG. 3 , the present invention provides an online life prediction method for lithium batteries based on data fusion and ARIMA model, and the specific implementation steps are as follows:

Step 1: Collect the lithium battery's iso-voltage discharge, charging time interval and battery capacity data.

The data used in the present invention comes from the lithium battery life test data set published by the NASA Failure Prediction Research Center (Prognostics Center of Excellence, PCoE), and the life prediction is performed by using the lithium battery data numbered B0005.

Draw the curve of voltage versus time during the first discharge process (Cycle 1) of the B0005 lithium battery, as shown in Figure 4. In Figure 4, the horizontal axis represents the lithium battery discharge time, in seconds (s), and the vertical axis represents the corresponding discharge voltage, in volts (V). The battery has experienced the discharge stage and the static stage successively. In the discharge stage, the voltage drops from 4.2V to 2.6V due to the continuous discharge of the battery, and the speed of the voltage value decline is first slow and then accelerated. During the resting phase, the voltage gradually rises back up to 3.3V due to the self-charging of the battery. Because the voltage recovery range in the stationary phase is affected by many factors and is difficult to determine, the present invention uses the time interval of the lithium battery discharge phase as one of the training data for online prediction.

待锂电池放电到V₂＝3.6V时再记录当下的时刻值

则第一个周期的等电压放电时间间隔

以此类推，计算出第i个周期的等电压放电时间间隔

其中，i＝1,2,…,N。如图5所示，得到锂电池所有试验周期的等电压放电时间间隔。During the first discharge cycle of the B0005 lithium battery, first record the time value when the voltage value V ₁ =3.8V

When the lithium battery is discharged to V ₂ =3.6V, record the current time value

Then the equal voltage discharge time interval of the first cycle

By analogy, the equal-voltage discharge time interval of the i-th cycle is calculated

where i=1,2,...,N. As shown in Figure 5, the equal-voltage discharge time intervals of all test cycles of the lithium battery are obtained.

Same as above, draw the curve of voltage versus time during the first charging process (Cycle 1) of the B0005 lithium battery, as shown in Figure 6. The physical meaning represented by the abscissa and ordinate axes of FIG. 6 is the same as that of the discharge process. During the charging process, the lithium battery undergoes 1.5A constant current charging and 4.2V constant voltage charging successively. The invention uses the time interval of the lithium battery constant current charging stage as one of the data for online analysis and prediction.

待锂电池充电到V₄＝4.0V时再记录当下的时刻值

则第一个周期的等电压充电时间

以此类推，计算出第i个周期的等电压充电时间

其中，i＝1,2,……,k。如图7所示，得到锂电池所有试验周期的等电压充电时间间隔。During the first charging cycle of the B0005 lithium battery, first record the moment value when the voltage value is V ₃ =3.8V

When the lithium battery is charged to V ₄ =4.0V, record the current time value

Then the equal voltage charging time of the first cycle

By analogy, calculate the equal-voltage charging time of the i-th cycle

Among them, i=1,2,...,k. As shown in Figure 7, the equal-voltage charging time intervals of all test cycles of the lithium battery are obtained.

As shown in FIG. 8, the present invention directly uses the battery capacity information provided in the data set.

Step 2: Calculate the weight of data layer fusion and fuse the data.

If the equal-voltage discharge time interval sequence is {td _i |i=1,2,...,N}, and the equal-voltage charging time interval sequence is {tc _i |i=1,2,...,N}, the battery The capacity sequence is {C _i |i=1,2,...,N}, then the Pearson correlation coefficients r ₁ and r ₂ of the battery capacity between the equal-voltage discharge time interval and the equal-voltage charge time interval are calculated as follows: shown:

为等电压放电时间间隔序列的平均值，

为等电压充电时间间隔序列的平均值，

为电池容量序列的平均值。in,

is the average value of the equal-voltage discharge time interval sequence,

is the average value of a sequence of equal-voltage charging time intervals,

is the average value of the battery capacity series.

The equal-voltage discharge time interval, equal-voltage charge time interval and battery capacity obtained in the first step are not within an order of magnitude, so data normalization is performed before prediction.

The formula for calculating the state of health of the battery, which is normalized by the equal-voltage discharge time interval, is as follows:

The formula for calculating the state of health of the battery, which is normalized by the equal-voltage charging time interval, is as follows:

The calculation formula of the battery state of health expressed by the normalized battery capacity is as follows:

Among them, td ₁ , tc ₁ and C ₁ respectively represent the observed state of the lithium battery at the initial moment of the test.

Using the calculated Pearson correlation coefficients r ₁ , r ₂ , the normalized equal-voltage charging and discharging time interval data are weighted and fused, and the fused battery health state information is expressed as:

Step 3: Train and estimate ARIMA model parameters and test the ARIMA model.

The equal-voltage discharge time interval and the equal-voltage charging time interval show a downward trend with the increase of the number of charge and discharge uses, so the fused data is a non-stationary time series. In the training process of the ARIMA model, it is necessary to perform d-order difference on the training data until the differenced sequence satisfies stationarity.

After determining the difference order d, clarify the training range of the autoregressive order p and moving average order q of the ARIMA(p,d,q) model, train the parameters of each ARIMA model, and then use the Bayesian information criterion BIC to estimate model parameters to obtain the optimal ARIMA prediction model.

是模型的残差方差的估计值，则BIC函数的计算如下:Assuming that the sample time series is {X _t }, t = 0, ± 1, ... N,

is an estimate of the residual variance of the model, then the BIC function is calculated as follows:

The smaller the calculated value of BIC, the better the fitting effect of the relevant model and the smaller the prediction error.

As shown in Table 1, after the ARIMA model is trained, the T-statistic (T-Statistic) of the ARIMA model parameters is calculated to test whether the ARIMA model parameters are significant. If the parameters are significant, the overall fitting effect of the model is acceptable.

Table 1 is the static saliency properties of the ARIMA(1,1,1) model parameters.

Step 4: Use the fused data to predict the RUL and the SOH of the next cycle through the ARIMA model.

Through the model parameters calculated in the third step, the optimal ARIMA model is obtained. The fused data is used to predict the remaining service life RUL of the battery and the state of health SOH of the battery in the next test cycle by using the ARIMA model. Assuming that the stationary time series is {X _t }, the white noise series is {ε _t }, t=0, ±1, ..., the calculation formula of ARIMA (p, d, q) is as follows:

X _t =φ ₁ X _t-1 +φ ₂ X _t-2 +…+φ _p X _tp +ε _t -θ ₁ ε _t-1 -θ ₂ ε _t-2 -…-θ _q ε _tq

Among them, p is the order of the autoregressive model, q is the order of the moving average model, φ ₁ , φ ₂ , ...φ _p are the parameters of the autoregressive model, θ ₁ , θ ₂ , ..., θ _q are the moving average model parameters. The above parameters are obtained through the model training in the third step.

The untrained battery capacity {C _i |i=K+1, K+2,...,N} in the first step is normalized and represented as the battery's state of health SOH, which is used as the prediction effect of the ARIMA model. Validation dataset.

After each prediction, the absolute error (Absolute Error) between the estimated value of the remaining service life RUL prediction and the real value of the validation set is calculated to measure the accuracy and effect of online real-time prediction. The formula for calculating the absolute error is as follows:

AE _i = |RUL _i -RUL _{i_predict} |

Among them, AE _i is the absolute error of the ith prediction, RUL _{i_predict} is the estimated value of the remaining service life of the ith prediction, and RUL _i is the true value of the remaining service life of the ith test.

As shown in Table 2, the ARIMA model uses the equal voltage discharge time interval and the equal voltage charge time interval to predict the absolute error of the remaining service life, which is generally larger than the prediction error of the ARIMA model using the data layer fusion information. It can be seen that the present invention improves the accuracy and robustness of the ARIMA model for predicting the remaining service life through the method of data layer fusion.

Table 2 shows the absolute error AE of the remaining useful life predicted by the ARIMA model using different data sources.

Step 5: Input the battery state observation data obtained online in real time, repeat steps 2 to 4, update the ARIMA prediction model, and realize online prediction.

When the second online prediction is to be performed after the first life prediction is completed, the real battery health state value of the K+1th trial and the first K lithium battery state observation data are used as the training data set, and then the AIRMA The model performs a new round of training and prediction, and obtains the prediction value and prediction error of the K+2th trial period.

By analogy, every time the data of a new test cycle is input, steps 2 to 4 will be repeated, and the ARIMA model parameters will be continuously updated online until the data of all test cycles are processed, and the online prediction of lithium battery life is completed. Finally, through continuous model training, the predicted value of the battery state of health of N-K test cycles will be obtained, and the root mean square error (RMSE) of the predicted value and the verification data will be calculated, and the prediction effect of the model will be evaluated.

The formula for calculating the root mean square error of lithium battery health state prediction is as follows:

Among them, SOH _{i_predict} is the estimated value of the state of health of the battery in the i-th prediction, SOH _i is the real observed value of the state of health of the lithium battery in the i-th test, and n is the number of life predictions that have been made.

The traditional prediction method uses K known data to train the ARIMA model, and after solving the optimal ARIMA model, predicts the battery health state of the remaining N-K cycles. The traditional prediction method is not very accurate in predicting the state of health of the battery in the later stage, and the ARIMA training model is not dynamically updated, so it cannot be used for the battery model with accelerated degradation of the state of health in the later stage. The invention improves the ARIMA model prediction method, and predicts the state value of the next cycle each time, which is conducive to online real-time observation of the local changes of the battery health state. As shown in Table 3, the prediction accuracy of the improved ARIMA model of the present invention is far superior to the prediction accuracy of the traditional ARIMA model. When there are enough training samples in the later stage, it can be directly used to predict the state change of the remaining cycle and predict the remaining service life of the lithium battery.

Prediction starting point Traditional ARIMA forecasting method Improved ARIMA prediction method 60 0.0474 0.0132 65 0.1025 0.0135 70 0.1430 0.0139 75 0.1942 0.0142 80 0.2122 0.0144 85 0.2218 0.0148 90 0.1000 0.0097 95 0.1628 0.0098 100 0.1736 0.0101

Table 3 shows the RMSE of the traditional ARIMA model and the improved ARIMA model that both use the fusion information of the data layer to predict the battery capacity.

Although the present invention has been disclosed by the above embodiments, it is only an example and is not intended to limit the present invention. Any person with ordinary knowledge in the technical field can make various equivalents without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the scope defined by the appended claims.

Claims (1)

1. A lithium battery online service life prediction method based on data fusion and an ARIMA model is characterized in that: the method comprises the following steps:

the method comprises the following steps: collecting data of the equal-voltage discharge and charge time interval and the battery capacity of the lithium battery;

step two: calculating the weight of data layer fusion and fusing data;

step three: training and estimating ARIMA model parameters and checking the ARIMA model;

step four: predicting RUL and SOH of the next period by the data fused in the step two through an ARIMA model;

step five: inputting battery state observation data acquired online in real time, repeating the second step to the fourth step, updating an ARIMA prediction model, and realizing online prediction;

the specific process of the step one is as follows:

collecting voltage from V in K discharge cycles before lithium battery₁Is put to V₂Time value of (D) is recorded as

And

calculate the sequence of equal voltage discharge time intervals, record

Collecting the voltage from V in K charging periods before the lithium battery₃Fill to V₄Time value of (D) is recorded as

And

calculating the charging time interval of the constant voltageSequence, is described as

Collecting battery capacity sequences of K charge-discharge cycles before the lithium battery, and recording the battery capacity sequences as { C_i1,2, and collecting a battery capacity sequence of the residual charge-discharge period as a verification set, and recording the battery capacity sequence as { C [ ]_iI ═ K +1, K +2,.., N }, where N is the total number of all charge-discharge cycle tests;

the second step comprises the following specific processes:

s21, respectively calculating Pearson correlation coefficients of the equal-voltage charging and discharging time intervals as the online indirect health factors and the battery capacity as the offline direct health factors, which are obtained in the step one;

s22, carrying out normalization processing on the data of the interval between the charging time and the discharging time of the equivalent voltage;

s23, carrying out weighted average on the equal-voltage charging and discharging time interval data subjected to normalization processing in the step S22 by using the correlation coefficient calculated in the step S21 to obtain input training data of the ARIMA model;

the third step comprises the following specific processes:

s31, performing stationarity judgment on the input training data after the data layer fusion in the step two, if the sequence is not stable, performing d-order difference on the training data, and confirming a parameter d of the ARIMA model;

s32, determining the optimizing ranges of the other two parameters p and q of the ARIMA model, including the autoregressive order p and the moving average order q, and training the ARIMA model under each parameter;

s33, estimating the optimal ARIMA model parameters by using a Bayesian information criterion;

s34, checking whether the ARIMA model parameters are significant by calculating the T statistic of the ARIMA model parameters;

the specific process of the step four is as follows:

s41, obtaining a trained optimal ARIMA model through the model parameters estimated in the third step, performing autoregressive moving average operation on the fusion data in the second step, and predicting the remaining service life of the battery and the battery health state in the next test period;

s42, the battery capacity sequence { C without training in the step one_iNormalizing the i | (K +1, K + 2.., N }, representing the i | (K + 2) | into the health state of the battery, and using the health state as a verification data set of the model prediction effect;

s43, calculating the absolute error between the estimated value of the residual service life prediction and the verification data value of S42 each time, and measuring the accuracy and effect of online real-time prediction;

the concrete process of the step five is as follows:

s51, sequentially inputting the ith lithium battery state observation data of a new test period, including equal voltage discharge, charging time intervals and battery capacity, and forming a new training data set with the input i-1 test period observation data, wherein i is K +1, K +2, … and N;

s52, repeating the second step to the fourth step every time data of one test period are input, continuously updating the ARIMA model parameters on line until the data processing of all the test periods is finished, and finally obtaining the predicted value of the battery health state of N-K test periods;

and S53, calculating the root mean square error of the prediction result and the verification data, and taking the root mean square error as an index for judging whether the prediction effect is good or bad.

Images