KR20220043630A - A method for predicting failure of aluminum electrolytic capacitors using machine learning - Google Patents

Abstract

The present invention relates to a method for predicting a failure of an aluminum electrolytic capacitor using machine learning, which comprises the steps of: measuring charging voltage and discharging voltage for a predetermined period at a limit temperature or more of a capacitor, and converting the same into capacitance through equations 1 and 2; constructing a data set by performing a preprocessing task of averaging data using a centered moving average to reduce an error in the acquired data of the capacitor; and performing a long-short term memory (LSTM) using 70% of the data in the data set as training data and the remaining 30% thereof as test data. Therefore, a failure can be predicted through machine learning.

Description

A method for predicting failure of aluminum electrolytic capacitors using machine learning

The present invention relates to a failure prevention method, and more particularly, to a failure prevention method for an aluminum electrolytic capacitor using machine learning.

With the 4th industrial revolution, artificial intelligence using big data has been applied to various fields such as finance, stocks, medical care, agriculture and fishery industry, machinery, electronics and energy, and R&D is being actively carried out. Artificial intelligence is a technology that realizes human perception, learning, reasoning, thinking, and understanding of natural language with computer programs. Active research is being conducted in the field of improving equipment operation rate and reducing costs through failure prediction technology that predicts equipment failure in advance.

Aluminum electrolytic capacitors, which are used in power supplies, frequency control devices, and charging devices of most industrial equipment, have a large charge capacity compared to their size by using a thin oxide film dielectric. It happens often. In addition, since various electronic components as well as capacitors are mounted in the electronic circuit, it is difficult to recognize even if a failure occurs in the capacitor, so that the replacement of the electronic circuit board is required. In past research on capacitors, the lifespan of capacitors is judged by photographing the swelling phenomenon that swells according to the degree of deterioration with a camera, and the life expectancy is predicted through tests. not confirmed.

The present invention has been proposed to solve the above technical problems, and failure prediction was performed by applying LSTM (Long Short-Term Memory) and Vanila LSTM among machine learning techniques for aluminum electrolytic capacitors, and for each technique, To evaluate the loss according to the epoch and the accuracy of the model, the performance is compared using the scales RMSE (Root Mean Square Error) and MAE (Mean Absolute Error).

According to an embodiment of the present invention to solve the above problem, the capacitance is detected by measuring the charging voltage and the discharging voltage for a predetermined period above the limit temperature of the capacitor, and based on the detected data, a Long Short-Term (LSTM) Memory) or Vanila LSTM is applied to determine the failure prediction; an aluminum electrolytic capacitor failure prediction method using machine learning is provided, including.

In addition, according to another embodiment of the present invention, the steps of measuring the charging voltage and discharging voltage for a predetermined period above the limit temperature of the capacitor, and converting it into capacitance through Equations 1 and 2, and obtaining data of the capacitor In order to reduce the error in the target, the data set is constructed by performing preprocessing work to average the data using a centered moving average, and 70% of the data in the data set is the training data, and the remaining 30% of the data is the test data. A method for predicting failure of an aluminum electrolytic capacitor using machine learning is provided, including the step of using LSTM (Long-Short Term Memory).

<Equation 1>

<Equation 2>

Through the proposed aluminum electrolytic capacitor failure prediction method using machine learning, it is possible to predict failure by performing big data and machine learning on aluminum electrolytic capacitors used in most electronic circuits. The results obtained are summarized as follows.

1) By measuring the capacitance of aluminum electrolytic capacitors at high temperatures, failures can be predicted through machine learning.

2) LSTM and Vanila LSTM decrease loss as the epoch increases, but Vanila LSTM minimizes loss faster than LSTM.

3) Vanila LSTM decreased RMSE by 8.6% and MAE by 1.7% compared to LSTM.

1 is a view showing the internal structure of an aluminum electrolytic capacitor;
2 is a diagram showing a circuit for an accelerated life test;
3 is a conceptual diagram of a failure prediction big data collection system
4 is a diagram showing the load voltage and current waveforms during the discharge period;
Fig. 5 is a diagram of capacitor voltage measurement for accelerated life test at 140 °C;
6 is a structural diagram of an LSTM.
7 is a capacitance graph of moving average data.
8 is a graph of correlation between collected information (Ground Truth) and actual data;
9 is a capacitance graph of actual data and predicted data.
10 is a graph showing the loss of LSTM and Vanilla LSTM;
11 is a result of calculating RMSE and MAE

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings in order to describe in detail enough that a person of ordinary skill in the art to which the present invention pertains can easily implement the technical idea of the present invention.

- Failure mode

Table 1 is a table showing the FMEA of the aluminum electrolytic capacitor.

FMEA (Failure mode and effects analysis) results of aluminum electrolytic capacitors are as shown in Table 1, short circuit due to excessively applied voltage, open circuit due to adhesive/coating material and stress, current and electrolyte due to deterioration and overheating of sealant material. It can be broadly divided into four categories: evaporation and reduction in capacitance due to excessively applied voltage and electrolyte reduction. As can be seen from the table, the decrease in current and electrolyte with the highest RPN (Risk Priority Number) is a factor affecting the decrease in capacitance. Accelerated life test for failure prediction was performed.

<Table 1>

1 is a view showing an internal structure of an aluminum electrolytic capacitor.

That is, the internal structure of the aluminum electrolytic capacitor is the same as in FIG. 1 , and it is composed of an electrode plate, a lead wire, and an insulator, and the DC current flowing into the lead wire is blocked by the insulator, and energy is stored on the principle that electric charges are stored at the boundary. Capacitance is the ability of a capacitor to store charge, a decrease in capacitance means a decrease in the capacitance of the capacitor, which causes capacitor failure, and according to standard approaches, the capacitance is reduced by 15~ A 20% decrease is considered a failure.

The biggest reason for the decrease in capacitance is due to heat generation. As the power consumed in the aluminum electrolytic capacitor increases, the temperature inside the capacitor rises, and the temperature of the electrolyte increases, causing evaporation of the electrolyte. The capacitance of the capacitor decreases, resulting in capacitor failure.

- Test apparatus and method

2 is a diagram illustrating a circuit for an accelerated life test, and FIG. 3 is a conceptual diagram of a failure prediction big data collection system.

That is, in order to collect data from the aluminum electrolytic capacitor, a circuit was designed as shown in FIG. 2, and a circuit using a capacitor, a chamber, and a data acquisition (DAQ) test apparatus were configured as shown in FIG. Through this, voltage data by repeated charging and discharging of the capacitor was collected and stored using Labview. At this time, the sampling rate was set to 1000sps and the number of samplings was set to 1, and 1000 data were received per second.

Table 2 is a table showing the specifications of the aluminum electrolytic capacitor.

The specifications of the aluminum electrolytic capacitor used in the test are shown in Table 2, and as a result of the high temperature test at 140°C, it was confirmed that the failure occurred after about 72 to 480 hours.

- Data collection and extraction

4 is a diagram illustrating load voltage and current waveforms during a discharge period.

From the failure mode analysis result of FMEA, data was collected and analyzed by selecting temperature as the main parameter. there is.

The capacitance (C) of the capacitor can be expressed as Equation (1) from the relationship between energy (E) and voltage (V).

<Equation 1>

In addition, the energy (E) can be expressed as Equation (2) from the relationship between the voltage (V), the current (I), the measurement period (T), and the number of samples (n).

<Equation 2>

<Table 2>

The voltage data of the aluminum electrolytic capacitor was organized by applying the above formula, and only the necessary data was extracted using Python and converted into capacitance through the capacitance estimation technique. The accuracy of the capacitance can be confirmed through Table 3, and the error rate tends to decrease as the number of samples (n) is high and the measurement period (T) is small, so n = 30 and T = 30ms were tested. proceeded.

5 is a diagram of capacitor voltage measurement for an accelerated life test at 140°C.

When the capacitor temperature is 140° C., the charging and discharging voltage cycles are displayed one by one at the same time intervals of 4 days, such as the 1st, 5th, and 9th days, as shown in FIG. 5(a). Comparing the results measured on the 1st and 21st days, when the aluminum electrolytic capacitor was fully charged, the voltage increased from about 3.8V to about 4.2V, and after being discharged for 30ms, the voltage decreased from about 3V to about 0.1V.

<Table 3>

When the voltage data measured for 21 days is converted into capacitance using equations (1) and (2), it can be displayed as shown in Fig. 5(b), and the capacitance slightly increases from the initial 11.07㎌ However, it shows a decreasing trend. Although the capacitance increases as the initial discharge voltage increases, it can be seen that after about 7 days, the capacitance decreases as the initial discharge voltage increases and the voltage decreases after discharging for 30 ms.

- LSTM and Vanila LSTM

6 is a structural diagram of an LSTM.

In RNN (Recurrent Neural Networks), the most basic model for data processing in the form of a continuous sequence, a vanishing gradient problem occurs when time series data becomes long. The technique to solve this problem is LSTM.

LSTM is a sigmoid neural network layer that selectively transmits information to the next step by adding or removing information and the cell state that can transmit the data applied with the structure shown in FIG. 6 without loss or deformation ( The sigmoide neural net layer and the tanh layer interact and consist of three gates.

The first step of LSTM, the forget gate layer (f _t ), determines which information to store and remove from the cell state through σ (sigmoid layer), and can be expressed as Equation (3).

<Equation 3>

(candidate vector)을 만드는 tanh layer로 구성되어 식 (4) 및 식 (5)와 같이 나타낼 수 있다.The second step is an input gate layer (i _t ) that determines which data to store among future data, a sigmoid layer that determines which values to update, and a cell state to be added.

It is composed of a tanh layer that creates a (candidate vector) and can be expressed as Equations (4) and (5).

<Equation 4>

<Equation 5>

??)를 이전의 셀상태 C _t-1 (previous cell memory)에 적용하여 새로운 셀 상태C_t(current cell memory)로 갱신하기 위해 식 (6)을 사용하여 최신화 한다.The data obtained in the previous step (f _t , i _t and

??) is applied to the previous cell state C _t-1 (previous cell memory ) and updated using Equation (6) to update to the new cell state C _t (current cell memory).

<Equation 6>

Here, ⊙ is element wise multiplication.

The last step is the process of determining the output data using the sigmoid layer and the tanh layer as the output gate layer (Ot), and it can be expressed as Equations (7) and (8).

<Equation 7>

<Equation 8>

LSTM can perform abnormal signal detection and prediction, etc. by memorizing and learning information for a long period of time through such an iterative process. However, there is a disadvantage that the analysis takes a lot of time, and to compensate for this disadvantage, the Vanilla LSTM with a simple structure with a single hidden layer and an output layer is attracting attention. It analyzes simple sequence data and specific data faster and more accurately than the existing LSTM. And it has the characteristics of being able to predict.

- Al electrolytic capacitor failure prediction

7 is a capacitance graph of moving average data, and FIG. 8 is a correlation graph between collected information (ground truth) and actual data.

In order to perform machine learning for failure prediction of aluminum electrolytic capacitors, LSTM was performed using keras as a framework using a Jupyter program. First, a data set was constructed as shown in FIG. 7 through a preprocessing operation of averaging five pieces of data using a centered moving average in order to reduce errors with respect to the acquired data.

Of the data set, 70% (115,416) of data was used as training data, and the remaining 30% (49,464) of data was used as test data. To check whether overfitting or underfitting occurs during training, the validation data was tested using 15% (17,312) of the training data. In addition, when the ground truth is expressed as a graph with the x-axis and the actual data as the y-axis, as shown in FIG. 9, the correlation coefficient (r) is calculated as 0.8945, confirming that the research result is reliable (r≥0.7).

9 is a capacitance graph of actual data and predicted data, and FIG. 10 is a graph showing the loss of LSTM and Vanilla LSTM.

The results of performing the LSTM with the number of hidden layers set to 100, learning late = 0.001, epoch = 100, batch size = 1000, etc. and Vanila LSTM with one hidden layer are shown as shown in FIG. 9 . When looking at the application results of LSTM and Vanilla LSTM, overfitting or underfitting, which is a phenomenon in which the actual value and the predicted value were the same or significantly different, was not observed, and it was difficult to confirm a large difference in comparing the two graphs.

The epoch means the number of times the data set is learned, and the loss indicates the difference between the predicted value and the actual value. That is, as the train loss and validation loss approach 0, it can be said that learning is good, and it can be confirmed through FIG. 10 how many times the training was performed through the epoch value. In the figure, the train loss of LSTM and Vanilla LSTM approaches 0 as the epoch increases, and the validation loss also decreases as the epoch increases. 0.043 is shown. As a result, even though Vanila LSTM prevented overfitting by early stopping at epoch 14, validation loss was 0.044 smaller than that of existing LSTM, and it proceeded quickly and accurately. Through this, it can be seen that Vanila LSTM is more advantageous in terms of time and accuracy than LSTM if machine learning by derivation of the optimal epoch is performed.

Root Mean Square Error (RMSE) and Mean Absolute (MAE), as in Equations (9) and (10), which are used to evaluate the accuracy of LSTM and Vanila LSTM models, and are used to obtain the error between the predicted value and the actual measured value. Error) was applied to compare.

<Equation 9>

<Equation 10>

<Table 4>

11 and Table 4 show the results of calculating the RMSE and MAE by performing machine learning on the optimal LSTM (epoch 100) and Vanila LSTM (epoch 14).

RMSE shows that Vanila LSTM is 8.6% higher than LSTM and MAE is 1.7% higher, so Vanilla LSTM can proceed faster and more accurately than existing LSTM when predicting aluminum electrolytic capacitor failure.

- conclusion

Big data and machine learning were performed on aluminum electrolytic capacitors used in most electronic circuits to predict failure. The results obtained are summarized as follows.

1) By measuring the capacitance of aluminum electrolytic capacitors at high temperatures, failures can be predicted through machine learning.

2) LSTM and Vanila LSTM decrease loss as the epoch increases, but Vanila LSTM minimizes loss faster than LSTM.

3) Vanila LSTM decreased RMSE by 8.6% and MAE by 1.7% compared to LSTM.

As such, those skilled in the art to which the present invention pertains will understand that the present invention may be embodied in other specific forms without changing the technical spirit or essential characteristics thereof. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. The scope of the present invention is indicated by the following claims rather than the above detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

Claims (2)

detecting a capacitance by measuring a charging voltage and a discharging voltage for a predetermined period above the limit temperature of the capacitor, and determining a failure prediction by applying a Long Short-Term Memory (LSTM) or Vanila LSTM based on the detected data; Aluminum electrolytic capacitor failure prediction method using machine learning, including a.
measuring the charging voltage and the discharging voltage for a predetermined period above the limit temperature of the capacitor, and converting it into a capacitance through Equations 1 and 2;
constructing a data set by performing a pre-processing operation of averaging data using a centered moving average in order to reduce an error with respect to the acquired data of the capacitor; and
In the data set, 70% of data is used as training data, and the remaining 30% of data is used as test data to perform Long-Short Term Memory (LSTM);
Aluminum electrolytic capacitor failure prediction method using machine learning, including a.
<Equation 1>

<Equation 2>

Images