A method for calculating the life of an electronic device based on kernel least square filter

Abstract

The application discloses a kind of electronic equipment life calculation methods based on kernel least square filter, it is related to adaptive signal processing and machine learning field.The application estimates the life of electronic equipment using adaptive filter, when the weight of adaptive filter is updated in the calculation process, the traditional integer order derivation method is abandoned, fractional order derivation is selected, the algorithm accuracy and algorithm robustness are improved.Compared with the traditional KLRS algorithm, the application has better performance when processing data containing a large number of outliers and mixed noise, and the estimation error is also improved to a certain extent.At the same time, compared with the traditional kernel adaptive filtering algorithm, the application method performs better when calculating the life of electronic equipment in the non-Gaussian noise condition, and the convergence speed and accuracy of the algorithm are improved.The life of electronic device is more accurate, and with more life calculation, the adaptive filter will be updated, so that subsequent calculation is more accurate.

Description

Technical Field

[0001] This invention relates to the fields of adaptive signal processing and machine learning. Background Technology

[0002] To date, kernel methods have been successfully applied in time series forecasting, nonlinear regression, and machine learning. Kernel adaptive filtering (KAF) employs an adaptive filtering (AF) algorithm to solve nonlinear problems in reproducing kernel Hilbert spaces (RKHS). In recent years, adaptive filtering algorithms have mainly focused on integer-order optimization methods, but fractional-order adaptive algorithms offer better performance in terms of convergence speed and estimation error compared to integer-order adaptive algorithms.

[0003] Kernel Least Squares (KRLS) exhibits excellent performance in Gaussian noise environments and is currently widely used for updating filter weights in signal processing. However, in many practical applications, noise, due to its significant peak characteristics or large outliers, does not follow a Gaussian distribution. Under such impulse noise, KRLS suffers from performance degradation or even failure. Therefore, to overcome this problem, it is essential to develop an innovative kernel adaptive filtering algorithm that performs fractional derivatives on the weight update formula to combat non-Gaussian noise and further improve the performance of KRLS. Summary of the Invention

[0004] This invention performs fractional derivatives on the weight update equation to address the limitations of existing technologies in handling data containing a large number of outliers and mixed noise.

[0005] The technical solution of this invention is: a method for calculating the lifetime of electronic devices based on kernel least squares filtering, the method comprising:

[0006] Step 1: Collect a sample database; Select multiple electronic devices and continuously collect voltage data between their input terminals and ground terminals in real time from the first operation of these electronic devices until the electronic devices reach the failure condition; Extract a fixed-length segment of data from the entire voltage data sequence and combine it with the remaining lifespan of the device corresponding to the end point of the data segment to form a training data; Use the method in Step 2 above to obtain the sample database;

[0007] Step 2: Construct an adaptive filter and train it using the sample database obtained in Step 1. Use the adaptive filter to predict the lifespan of the electronic device. The input of the filter is the voltage timing data between the input terminal and the ground terminal of the electronic device when it is working, and the output is the lifespan of the electronic device.

[0008] The cost function of the adaptive filter is:

[0009]

[0010] Where d(j) is the life expectancy prediction result, The input to the adaptive filter is represented by the voltage timing data between the input terminal and the ground terminal when the electronic device is operating; w represents the adaptive weight. To predict the output, use the integral operator. As a regularization term, a represents any lower limit of integration, α represents the α-order integral, and λ is the regularization parameter;

[0011] Step 3: Minimize the cost function. The parameters w that need to be updated each time during the training of the adaptive filter are:

[0012]

[0013] Where i is the iteration number, and e(j) represents the predicted lifetime output. The error between the expected lifetime output d(j);

[0014] make get:

[0015]

[0016] Error equation Substituting into the above equation, we get:

[0017]

[0018] get

[0019] In the formula:

[0020] V(i)=(v(1),v(2),…v(j)…,v(i))

[0021]

[0022]

[0023]

[0024] Step 4: Let That is, w i =V(i)Ω(i), and update Ω(i) using the following formula;

[0025]

[0026] in, Z represents the estimation error of the lifetime prediction results. i =P i-1 G(i-1)h i K i=m i P i-1 r i =(k i g(i)+λ-g(i)m i Z i ) -1 P i =(λI+G i H i ) -1 ,

[0027]

[0028]

[0029] <> represents the inner product operation;

[0030] Step 5: Repeat steps 3 and 4 until the adaptive filter training is complete, and use the trained adaptive filter to calculate the lifetime of the newly obtained electronic device.

[0031] This invention employs an adaptive filter to predict the lifespan of electronic devices. Specifically, during the calculation process, when updating the weights of the adaptive filter, it abandons the traditional integer-order derivative method and instead uses fractional-order derivatives, improving the algorithm's accuracy and robustness. Compared to the traditional KLRS algorithm, this invention performs better when handling data containing a large number of outliers and mixed noise, and the estimation error is also improved. Furthermore, under non-Gaussian noise conditions, compared to the traditional kernel adaptive filtering algorithm, this invention's method performs better in calculating the lifespan of electronic devices, with improved convergence speed and accuracy. This makes the calculation of electronic device lifespan more accurate, and with each lifespan calculation, the adaptive filter is updated, further enhancing the accuracy of subsequent calculations. Attached Figure Description

[0032] Figure 1 A comparison of the MSE convergence curves of the relevant algorithms under MG data with impulse noise;

[0033] Figure 2 A comparison of the MSE convergence curves of the correlation algorithm on Lorenz time series data with impulse noise;

[0034] Figure 3 University of Maryland battery data;

[0035] Figure 4 NASA Prediction Center battery data;

[0036] Figure 5 Results of one-step capacity estimation for battery A3;

[0037] Figure 6 Results of one-step capacity estimation for battery B00005;

[0038] Figure 7 Results of 50-step capacity estimation for battery A3;

[0039] Figure 8 Results of 120-step capacity estimation for battery A5;

[0040] Figure 9 Results of 120-step capacity estimation for battery A8;

[0041] Figure 10 Results of 200-step capacity estimation for battery A12;

[0042] Figure 11 Results of 100-step capacity estimation for battery B0005;

[0043] Figure 12 Results of 120-step capacity estimation for battery B0006;

[0044] Figure 13 Results of 110-step capacity estimation for battery B0018. Detailed Implementation

[0045] To verify the effectiveness of the proposed FrKRLS method, this simulation experiment uses the Mackey-Glass time series prediction problem, the Lorenz time series prediction problem, and the battery RUL prediction problem to validate the performance of FrKRLS. This section compares the algorithm before and after the improvement. In the simulation experiments, different noise types were used to test the algorithm's performance for different time series prediction problems. For the simulation experiments of the first two problems, the training set size was 1000, the test set size was 100, and ten Monte Carlo experiments were used to reduce randomness. The mean squared error was used as the metric for prediction accuracy. In addition, battery RUL prediction was performed using the lithium battery experimental datasets from the University of Maryland and NASA to verify the algorithm's performance. To eliminate the possible impact of computer performance, all experiments were run on the same computer with the following configuration: Intel(R) Core(TM) i5-12600K@3.70GHz, 16GB RAM, Windows 10 environment, and MATLAB R2019a running platform.

[0046] like Figure 3 As shown, the improved FrKRLS algorithm is less sensitive to impulse noise, while the MSE of the traditional KRLS algorithm shows a significant upward trend, and the steady-state MSEs of the two are quite different. The performance of the KRLS algorithm by fractional derivative is improved.

[0047] like Figure 6 As shown, under impulse noise conditions, compared to Gaussian noise conditions, the accuracy of the FrKRLS algorithm is improved from -35dB to -50dB, while the performance of the KRLS algorithm remains basically unchanged. This proves that the algorithm improved by fractional derivative is more adaptable to complex noise conditions, and the performance of FrKRLS is better than that of KRLS. It can provide better prediction performance and achieve the minimum steady-state MSE.

[0048] To verify the performance of the algorithm, this paper uses battery datasets from NASA's Prediction Center (PCoE) and the University of Maryland. The battery data is as follows: Figure 7 and Figure 8 As shown.

[0049] from Figure 7 As can be seen, NASA's dataset consists of three identical batteries, numbered B0005, B0006, and B0018, operating at room temperature according to commercial 18650 battery specifications through three different operating configurations (charge, discharge, and impedance). In constant current mode, the batteries were charged at 1.5A until the voltage reached 4.2V, then charged in constant voltage mode until the charging current dropped to 20mA. They were then discharged at a constant current of 2A until the voltages of batteries B0005, B0006, and B0018 dropped to 2.7V, 2.5V, and 2.5V, respectively. This type of battery has a rated capacity of 2Ah, and reaches its end-of-life (EOL) standard when its capacity decays to 70% of its rated capacity (2Ah - 1.4Ah).

[0050] same, Figure 8 The battery data was obtained in different laboratory settings at the University of Maryland and consists of four sets of data (numbered A1, A2, A3, and A4), with a failure threshold of 0.7A.

[0051] Furthermore, the root mean square error (RMSE) of the predicted capacity is used to evaluate the accuracy of the prediction. The calculation formula is:

[0052]

[0053] First, a one-step prediction was performed on the battery data, with a filter order of 7. The capacity estimation results for batteries B0005 and A3 are as follows: Figure 9 and Figure 10 As shown in the figure. Furthermore, the RMSE of the mean cross-validation results is shown in Table 1.

[0054] It can be seen that both algorithms can accurately track battery degradation data, with RMSE less than 0.04Ah. Furthermore, FrKRLS has higher accuracy than KRLS, indicating that the performance of the fractional-order algorithm has been improved, which can effectively reduce tracking errors and enhance algorithm performance.

[0055] The effectiveness of the fractional-order algorithm in long-term prediction was also verified by conducting a single-starting-point long-term capacity prediction experiment, and comparing the fractional-order algorithm with the original algorithm. In this experiment, we selected two different datasets and set different prediction lengths to improve reliability. The filter order was 7. The capacity estimation results are as follows: Figure 7-13 As shown in Table 2, the estimated RUL values ​​of the battery under different prediction lengths are as follows.

[0056] Table 1 shows the RMSE (mAh) of the algorithm based on one-step prediction.

[0057]

[0058] As can be observed from the figure, the RUL curves predicted by both algorithms fluctuate around the actual values, and the curve predicted by the FrKRLS algorithm is closer to the actual value. As shown in Table 2, both algorithms can accurately predict the battery RUL with a prediction error of less than 5, and the algorithm based on the fractional derivative improvement has a higher prediction accuracy than the traditional KRLS algorithm.

[0059] Table 2 Battery RUL Predicted in N Steps for Different Prediction Lengths

[0060]

Claims

1. A method for calculating the lifetime of electronic devices based on kernel least squares filtering, the method comprising: Step 1: Collect a sample database; Select multiple electronic devices and continuously collect voltage data between their input terminals and ground terminals in real time from the first operation of these electronic devices until the electronic devices reach the failure condition; extract a fixed-length segment of data from the entire voltage data sequence and combine it with the remaining lifespan of the device corresponding to the end point of the data segment to form a training data; use the method in step 2 above to obtain a sample database; Step 2: Construct an adaptive filter and train it using the sample database obtained in Step 1. Use the adaptive filter to predict the lifespan of the electronic device. The input of the filter is the voltage timing data between the input terminal and the ground terminal of the electronic device when it is working, and the output is the lifespan of the electronic device. The cost function of the adaptive filter is: Where d(j) is the life expectancy prediction result, The input to the adaptive filter is represented by the voltage timing data between the input terminal and the ground terminal when the electronic device is operating; w represents the adaptive weight. To predict the output, use the integral operator. As a regularization term, a represents any lower limit of integration, α represents the α-order integral, and λ is the regularization parameter; Step 3: Minimize the cost function. The parameters w that need to be updated each time during the training of the adaptive filter are: Where i is the iteration number, and e(j) represents the predicted lifetime output. The error between the expected lifetime output d(j); make get: Error equation Substituting into the above equation, we get: get In the formula: V(i)=(v(1),v(2),…v(j)…,v(i)) D(i)=(d(1),d(2),…d(j)…,d(i)) T , Step 4: Let That is, w i =V(i)Ω(i), and update Ω(i) using the following formula; in, Z represents the estimation error of the lifetime prediction results. i =P i-1 G(i-1)h i K i =m i P i-1 r i =(k i g(i)+λ-g(i)m i Z i ) -1 P i =(λI+G i H i ) -1 , <> represents the inner product operation; Step 5: Repeat steps 3 and 4 until the adaptive filter training is complete, and use the trained adaptive filter to calculate the lifetime of the newly obtained electronic device.

Images