CN113887571A - Electronic equipment fault prediction method for improving SVR algorithm - Google Patents

Abstract

The invention relates to an electronic equipment fault prediction method for improving SVR algorithm, which comprises the steps of extracting sample data capable of representing the whole state degradation process of electronic equipment from an electronic equipment state monitor; carrying out data preprocessing on the sample characteristic data; establishing a Support Vector Regression (SVR) prediction model; performing parameter optimization on the SVR prediction model through a particle swarm algorithm; designing a recursive multi-step prediction method based on the SVR model after the particle swarm optimization improvement, and drawing an equipment state degradation curve according to the recursive multi-step prediction method; and designing a fault early warning based on a sliding time window according to the equipment state degradation curve. The method for predicting the fault objectively and accurately reflects the running state of the monitored electronic equipment, predicts the future state change trend of the equipment, and timely performs fault early warning, thereby providing a basis for guiding subsequent equipment fault management and maintenance work.

Description

Electronic equipment fault prediction method for improving SVR algorithm

Technical Field

The invention relates to an equipment fault judgment technology, in particular to an electronic equipment fault prediction method based on multi-step prediction and particle swarm optimization SVR algorithm improvement.

Background

With the wide application and the continuous enrichment and enhancement of functions and performances of new technologies, the complexity and the precision of modern large-scale equipment systems are higher and higher. In order to allow a stable operation of the plant for a long time, meeting the requirements of production development, it is therefore required that the tolerance of the plant to faults is extremely low, which is in contradiction to the risks that may arise with increasingly complex plant systems. At present, the traditional system maintenance means such as periodic investigation, after-affair management and maintenance and the like cannot adapt to the development trend of modern equipment, and the requirements of an equipment monitoring and maintenance system, a scientific maintenance and management technology and an optional maintenance method which are efficient, accurate and low in cost are more and more strong.

With the development of modern sensor technology and information technology, more efficient state maintenance-based technology is more and more favored by people. The method adopts advanced fault diagnosis and prediction technology to detect according to the actual operation condition of the equipment, predicts the health degree change trend of the equipment and the equipment fault according to the dynamic information of the equipment during operation by measurement and statistics, and provides reasonable repair measures according to the actual operation condition of the equipment. The state-based maintenance method is particularly based on fault prediction, which is a key link in state maintenance, and can avoid the problem of excessive overhaul in preventive maintenance, thereby greatly reducing the maintenance times, reducing the expenditure and improving the maintenance efficiency.

A common method for predicting equipment failure in the industrial field is failure prediction based on expert knowledge experience, such as "a system for predicting equipment failure based on knowledge base" (CN 107966942A). However, the method depends on a domain expert knowledge base, has strong subjectivity and limitation, and has high requirements on the richness of expert knowledge and the expert knowledge level. Therefore, the demand for a data-driven failure prediction method is increasing, and the main ideas are as follows: and establishing a fault early warning model based on the historical data of the equipment, and setting a threshold value according to the generated early warning model, such as 'a fault prediction method based on machine learning' (CN 108304941A). The method establishes an accurate model conforming to the mechanism of the equipment, and has stronger universality and adaptability. Commonly used methods are regression models, time series models, gray models, neural network models, and the like. However, in consideration of the situation that the electronic device is in a complex working environment and the historical data of the electronic device is limited, the common mode cannot meet the requirement, and the Support Vector Regression (SVR) model has the advantages of wide application range, good generalization capability, strong robustness, simplicity in operation and the like, and is more suitable for learning of small samples. Therefore, the SVR prediction model can well meet the problems of more noise interference and less historical data amount in electronic equipment fault prediction, but the SVR prediction model still has some defects, such as the problem that a smooth parameter C and a kernel parameter sigma 2 in an algorithm model are difficult to determine, and the problem that the prediction process of the SVR prediction model lacks interpretability and the prediction stability is poor.

Disclosure of Invention

Aiming at the problems existing in the equipment fault prediction when the support vector regression algorithm is applied, the electronic equipment fault prediction method for improving the SVR algorithm is provided, the running state of the monitored electronic equipment is objectively and accurately reflected, the future state change trend of the equipment is predicted, the fault early warning is timely carried out, and a basis is provided for guiding the follow-up equipment fault management and maintenance work.

The technical scheme of the invention is as follows: an electronic equipment fault prediction method for improving an SVR algorithm specifically comprises the following steps:

1) data preparation and selection: selecting a plurality of groups of relatively complete full-life-cycle monitoring data from the state data set of the electronic equipment, and selecting characteristic parameters which can represent different fault types of the electronic equipment and can be continuously monitored and recorded as degradation characteristic variables of the equipment;

2) carrying out data preprocessing on the complete full-life-cycle monitoring data selected in the step 1): the data are measured state data and time sequence two-dimensional data, the two-dimensional data are divided into a training set and a test set by referring to a degeneration characteristic variable, and every four adjacent data are divided into one group in the training set and the test set to prepare for subsequent multi-step prediction;

3) establishing an SVR prediction model: sending the training set in the step 2) into an SVR prediction model for model training, and verifying the trained model by using a test set to obtain a prediction effect;

4) performing parameter optimization on the SVR prediction model established in the step 3) through a particle swarm algorithm to obtain an optimal smooth parameter C and a kernel function coefficient sigma²；

5) Designing a recursive multi-step prediction method based on the improved SVR model of the particle swarm algorithm, and drawing an equipment state degradation curve according to the method:

the idea of recursive multi-step prediction is shown by the following formula:

prediction(t)＝model(obs(t-1),obs(t-2),...,obs(t-m))

prediction(t+1)＝model(prediction(t),obs(t-1),obs(t-2),...,obs(t-m+1))

prediction data obtained at the time t according to the SVR prediction model, model is the SVR prediction model with the optimal parameters substituted, obs (t-1), obs (t-2), and obs (t-n) are actual training data measured by a monitor before the time t, prediction (t +1) is prediction data obtained after combining recursive multi-step prediction according to the SVR prediction model, every four adjacent data in the step 2) are divided into one group, wherein m is 4, namely the data at the time t +1 are predicted by taking the data at the time 4 before the time t and the prediction data at the time t as the training data, and the prediction data are generated by analogy;

drawing an equipment state degradation curve according to the prediction data;

7) and designing a fault early warning based on a sliding time window according to the data degradation curve.

Further, the step 2) adds white gaussian noise to the data set to realize data enhancement on the basis of dividing the data set, and specifically, the following expression is expressed:

wherein x (N) is an actual signal in the data set, s (N) is a signal added with noise, N is a signal sampling time, and N is a total signal sampling number; the magnitude of the input noise is set by setting the value of the signal-to-noise ratio SNR.

Further, the step 4) specifically comprises the following steps:

4.1) population initialization: initializing SVR model parameters including penalty factor C and kernel function parameter sigma²Generating an initial population;

4.2) constructing a fitness function: the fitness function f (x) is constructed by using the fitting error of the SVR prediction model;

4.3) particle swarm algorithm parameter setting: setting population scale, maximum genetic algebra and algorithm termination conditions in the algorithm;

4.4) calculating the individual extreme value and the global optimal solution: inputting an SVR prediction model for training and calculating an individual extreme value and a global optimal solution according to a fitness function;

4.5) update speed and position: updating the speed and the position of each particle according to the calculated individual extreme value and the global optimal solution;

4.6) judging whether the termination condition is met: whether the fitness value meets the requirement or not can be used as a termination condition of the algorithm, and a set maximum genetic algebra can also be used as the termination condition; if the algorithm termination condition is not met, returning to 4.4) to continue executing the operation until the algorithm termination condition is met;

4.7) obtaining the optimal smoothing parameter C and the kernel coefficient sigma²。

Further, the step 6) realizes the method: in the data inferiorAnd (3) curve generation, namely selecting the health degree value on the time sequence by adopting a time frame with a specified unit length, replacing a value corresponding to specified time by a calculated average value in the time frame, and setting a threshold model on the basis of a sliding time window:

wherein E_maxIs the maximum value of the average health value of the equipment, k is the early warning threshold coefficient, E_anFor the fault early warning threshold value, the value output according to each time window and E_anAnd comparing to judge whether to alarm.

The invention has the beneficial effects that: the invention improves the electronic equipment failure prediction method of SVR algorithm, 1) adopt the failure prediction mode based on data drive, have got rid of expert's knowledge experience of confirming the weight is difficult to obtain and the disadvantage with stronger subjectivity, through the analysis and excavation of the latent law of historical data, predict the deterioration condition of the apparatus state, the invention fully utilizes the apparatus state data, the failure prediction mode constructed can relatively objectively, accurately, scientifically evaluate the health condition of the industrial equipment; 2) particularly aiming at the problems that the working environment of the electronic equipment is complex and the required historical data is relatively less, compared with a common data-driven fault prediction mode-neural network, the method disclosed by the invention is more suitable for learning with small samples, and compared with a time sequence model, the method disclosed by the invention has stronger noise resistance and better robustness, so that the fault prediction mode disclosed by the invention can better meet the requirement of electronic equipment fault prediction; 3) parameter optimization is carried out on the SVR prediction model by utilizing a particle swarm algorithm, so that the SVR prediction model can obtain parameters more suitable for a used data set, and the prediction effect of the SVR prediction model is improved; 4) a recursive multi-step prediction method is designed on the basis of the SVR model after the particle swarm optimization is improved, so that the SVR model has better interpretability during training and prediction, and the prediction result is more stable; 5) compared with the traditional fault early warning, only two states of health and fault exist, the threshold model set by the method can carry out equipment state grading management according to different K value selections, informs the state of the equipment in advance, is beneficial to a user to take measures in advance, makes adjustment to prolong the service life of the equipment, reduces the fault rate and saves the cost.

Drawings

FIG. 1 is an operation flow chart of the electronic equipment fault prediction method based on multi-step prediction and particle swarm optimization SVR algorithm;

FIG. 2a is a graph of the effect of the SVR prediction model under voltage data;

FIG. 2b is a graph of the effect of the SVR prediction model on temperature data;

FIG. 3 is a flow chart of parameter optimization of the SVR prediction model by particle swarm optimization in the method of the present invention;

FIG. 4a is a graph showing the effect of the voltage data comparing with other prediction methods;

FIG. 4b is a graph showing the effect of the temperature data of the present invention compared to other prediction methods;

FIG. 5 is a flow chart of a sliding time window based fault warning in accordance with an embodiment of the present invention;

fig. 6 is a graph of a fault warning result based on a sliding time window according to an embodiment of the present invention.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.

The operation flow of the electronic equipment fault prediction method based on the multi-step prediction and particle swarm optimization SVR algorithm is shown in figure 1, and mainly comprises the following parts:

s1, selecting several groups of complete full-life-cycle monitoring data from the state data set of the electronic equipment, and selecting characteristic parameters which can represent different fault types of the electronic equipment and can be continuously monitored and recorded as degradation characteristic variables of the equipment. Different industrial equipment can select respective parameters to be monitored according to the characteristic types of the equipment, and the parameters mainly comprise current, temperature, voltage and the like.

S2, preprocessing the data aiming at the selected data set;

the invention selects data processing on the python platform. The method comprises the steps of inputting complete equipment state data of the full life cycle collected by a state monitor into a two-dimensional array through a python platform, wherein the two-dimensional array is respectively measured state data and a time sequence, dividing the two-dimensional array into a training set and a testing set according to experimental needs by referring to degradation characteristic variables, taking the first 80% of a group of data reflecting the full life cycle of the equipment as the training set, defining the training set of the measured state data as Y _ train, and defining the time training set as X _ train. The last 20% of the data reflecting the full life cycle of the device is taken as a test set, the measurement data test set is defined as Y _ test, and the time test set is defined as X _ test. On the basis, the data of the training set and the data of the test set are divided into a group according to the principle that four adjacent data are divided into a group, for example, the 1 st to 4 th data in the two-dimensional array are used as a first group, the 2 nd to 5 th data are used as a second group, and the like, so that the data division is completed.

In order to better check the robustness of the model and more truly reflect the actual working environment of the equipment during working, Gaussian white noise is added into the data set on the basis of dividing the data set to realize data enhancement. The specific operation is represented by the following expression:

where x (N) is the actual signal in the data set, s (N) is the signal after adding noise, N is the time of signal sampling, and N is the total number of signal samples. The magnitude of the input noise is set by setting the value of the signal-to-noise ratio (SNR). The present invention selects a signal-to-noise ratio of 30.

S3, setting parameters aiming at the SVR prediction model, wherein the parameters comprise the kernel function type, the smoothing parameter C and the kernel function coefficient sigma of the SVR model²；

In the python platform, an SVR prediction model required by an experiment is established by using a scinit-leann library, and the established model is like SVR (kernel, C, gamma). The initial SVR model parameters are set to SVR1 ═ SVR (kernel ═ rbf', C ═ 20, and gamma ═ 0.0001), where kernel is kernel type, C is smoothing parameter, and gamma is kernel coefficient, and the commonly used kernel is shown in table 1.

TABLE 1

Among them, the 'rbf' kernel, i.e., the gaussian kernel, has the best model fitting effect, so the rbf kernel is selected. The 'rbf' kernel function is

Can be derived from

Therefore, the kernel function coefficient can be changed only by adjusting the size of gamma.

After the SVR prediction model is established, training sets Y _ train and X _ train in the selected divided temperature data set and voltage data set (for example, as shown in table 2) are input to svr.fit (X _ train, Y _ train) to train the model. After the SVR prediction model training is completed, the time data of the last 20% data set is input to SVR prediction (test _ X) as X _ test to generate prediction data. And finally, comparing the predicted data with the actual data, and judging the prediction effect of the SVR model. The prediction effect of the SVR prediction model under two different data sets is shown in FIGS. 2a and 2 b. It can be seen from fig. 2a that the SVR model performs very poorly under the voltage data set. Therefore, the problem of parameter selection in the SVR prediction model is urgently needed to be solved.

TABLE 2

The invention selects Mean Square Error (MSE) and coefficient of determination (R)²) To quantitatively measure SVR prediction modelThe effect of the mold.

The mean square error is the average of the sum of squares of the deviations of the respective data from the true value, so the sum of squares of the whole of its estimated values is subtracted from the true value to average it, i.e.:

in the invention, the estimated value is the equipment state data fitted by the SVR model, and the fitted data f (x)_i) And true data y_iThe mean square error can be obtained by substituting the formula. The smaller the mean square error value is, the strong ability of the model to fit experimental data is indicated, but the mean square error value cannot be 0, when the mean square error value is 0, the model is indicated to completely fit the listed conditions, at the moment, the model is in an overfitting state, and the model effect in the practical application process is poor.

The decision coefficient indicates the magnitude of the correlation between the independent variable Y and the independent variable X. R²The formula of (1) is:

where SSE is the sum of squared residuals SSE ═ Σ (y)_i-f(x_i))²SST is the sum of the squares

The magnitude of the decision coefficient determines the closeness of the correlation. When R is²When the equation is closer to 1, the higher the reference value of the equation expressing the correlation is, the better the fitting effect is, the higher the interpretation degree of the independent variable on the dependent variable is, the higher the percentage of the variation caused by the independent variable in the total variation is, and the denser the observation points are near the generated curve; conversely, closer to 0, the lower the reference value, the poorer the fitting effect, the lower the degree of interpretation of the independent variable by the dependent variable, the lower the percentage of the total variation by the variation caused by the independent variable, and the more sparse the observation points are in the vicinity of the generated curve. And the MSE and R of the SVR prediction model under the voltage data set and the temperature data set²As shown in table 3.

TABLE 3

It can be seen from the table that the SVR model shows quite different prediction effects in different data sets under the same parameter, and hardly has the prediction effect in the voltage data set. Therefore, the reasonable selection of the parameters of the SVR prediction model becomes an urgent problem to be solved by the SVR model. The invention selects a particle swarm algorithm to improve the SVR prediction model, optimizes the smooth parameter C and the kernel function parameter sigma of the SVR²。

S4, performing parameter optimization on the SVR prediction model through a particle swarm algorithm, wherein the method specifically comprises the following steps:

1) population initialization: initializing SVR model parameters including a penalty factor C and a kernel function parameter sigma 2, and generating an initial population;

2) constructing a fitness function: the fitness function f (x) is constructed by using the fitting error of the SVR prediction model;

3) particle swarm algorithm parameter setting: setting population scale, maximum genetic algebra and algorithm termination conditions in the algorithm;

4) calculating individual extrema and global optimal solution: inputting an SVR prediction model for training and calculating an individual extreme value and a global optimal solution according to a fitness function;

5) update speed and position: updating the speed and the position of each particle according to the calculated individual extreme value and the global optimal solution;

6) judging whether a termination condition is met: whether the fitness value meets the requirement or not can be used as a termination condition of the algorithm, and a set maximum genetic algebra can also be used as the termination condition. If the algorithm termination condition is not met, returning to 4) to continue executing the operation until the algorithm termination condition is met;

7) the obtained optimal smoothing parameter C and the kernel coefficient sigma²。

The specific flow chart is shown in fig. 3.

In the python platform, the invention chooses to use scinit-opt library to establish a particle swarm algorithm, and the parameters of the particle swarm algorithm are PSO ═ PSO (func ═ phere, dim ═ 2, pop ═ 5, max _ iter ═ 1000, lb ═ 1,0.000001, ub ═ 100, 100). And setting the minimum value of the difference sum of the data which is obtained by fitting the SVR and the real data as a fitness function. In the particle swarm algorithm parameter setting, the invention sets the population size to be 5, the maximum iteration number to be 1000, and the upper and lower bounds of the two parameters to be 1-100 and 0.000001-100. And obtaining the optimal C and gamma values after iteration is finished.

S5, designing a recursive multi-step prediction method based on the improved SVR model of the particle swarm optimization, and drawing an equipment state degradation curve according to the method, wherein the method comprises the following specific steps:

substituting the optimal parameters into an SVR prediction model, adopting the idea of recursive multi-step prediction, combining data predicted from the SVR model at the time t with data actually measured before the time t as training data, obtaining predicted data at the time t +1 by using an original SVR model, and the like. The specific prediction idea is shown in the following formula:

prediction(t)＝model(obs(t-1),obs(t-2),...,obs(t-m))

prediction(t+1)＝model(prediction(t),obs(t-1),obs(t-2),...,obs(t-m+1))

the prediction (t) is prediction data obtained according to the SVR prediction model at the time t, the model is the SVR prediction model with the optimal parameters substituted, obs (t-1), obs (t-2), and obs (t-n) are actual training data measured by a monitor before the time t, and the prediction (t +1) is prediction data obtained according to the SVR prediction model after the recursive multi-step prediction is combined. In the present invention, m is set to 4, that is, data at the t +1 th time is predicted by using data at 4 times before t and predicted data at t as training data, and so on to generate predicted data. And finally, drawing a device state degradation curve according to the prediction data. After multiple experiments, the invention selects the most representative 4 experiments, and the degradation curves are shown in fig. 4a and 4 b.

It can be seen from FIGS. 4a and 4b that the SVR model has better prediction effect under small sample learning compared with BP neural network, and the improved SVR prediction model (Multi-PSO-SVR) has better prediction effect compared with the initial SVR prediction modelThe prediction effect and stability of the method are greatly improved compared with a SVR prediction model (PSO-SVR) improved by a particle swarm algorithm without multi-step prediction. And MSE and R under the voltage data set of Table 4²Comparative table 5 MSE and R under temperature data set²The comparison quantitatively shows the prediction effect of the improved SVR prediction model compared with the initial SVR prediction model.

TABLE 4

TABLE 5

As can be seen from tables 4 and 5, the prediction effect of the SVR prediction model improved based on the multi-step prediction and the particle swarm optimization is greatly improved compared with that of a single SVR prediction model under different data sets, and the prediction effect of the SVR prediction model improved based on the multi-step prediction and the particle swarm optimization is more stable. Therefore, the improved SVR prediction model based on the multi-step prediction and the particle swarm optimization selected by the invention can really show good prediction performance in the small sample learning of the electronic equipment.

S6, designing a fault early warning based on a sliding time window according to the data degradation curve, wherein the method specifically comprises the following steps:

and (4) adopting a time frame with a specified unit length to check the health degree value on the time sequence, and replacing the value corresponding to the specified time by the calculated average value in the time frame. Setting a threshold model on the basis of a sliding time window:

wherein E_maxIs the maximum value of the average health value of the equipment, k is the early warning threshold coefficient, E_anAnd a fault early warning threshold value is obtained. Then, the value and E outputted according to each time window_anComparing to determine whether to performAnd (6) alarming. The flow chart of the fault early warning is shown in fig. 5.

The method selects a degradation curve generated by improving an SVR prediction model under a temperature data set based on multi-step prediction and a particle swarm optimization as an example. The alarm is set when the temperature is higher than 35 ℃, namely the threshold value is set to 35, the window width is set to 4, the sliding size is 1 each time, and the number of the windows represents the time size. The use of the average value in the time window instead of the corresponding value for the specified time is intended to avoid the occurrence of false alarms due to external disturbances. The alarm result is shown in fig. 6. It can be seen from fig. 6 that the time window for its alarm is 137, which coincides with the time node above 35 ℃ in the actual curve. Therefore, the fault early warning design based on the sliding time window is accurate and effective.

The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims (4)

1. An electronic equipment fault prediction method for improving an SVR algorithm is characterized by comprising the following steps:

1) data preparation and selection: selecting a plurality of groups of relatively complete full-life-cycle monitoring data from the state data set of the electronic equipment, and selecting characteristic parameters which can represent different fault types of the electronic equipment and can be continuously monitored and recorded as degradation characteristic variables of the equipment;

2) carrying out data preprocessing on the complete full-life-cycle monitoring data selected in the step 1): the data are measured state data and time sequence two-dimensional data, the two-dimensional data are divided into a training set and a test set by referring to a degeneration characteristic variable, and every four adjacent data are divided into one group in the training set and the test set to prepare for subsequent multi-step prediction;

3) establishing an SVR prediction model: sending the training set in the step 2) into an SVR prediction model for model training, and verifying the trained model by using a test set to obtain a prediction effect;

4) performing parameter optimization on the SVR prediction model established in the step 3) through a particle swarm algorithm to obtain an optimal smooth parameter C and a kernel function coefficient sigma²；

5) Designing a recursive multi-step prediction method based on the improved SVR model of the particle swarm algorithm, and drawing an equipment state degradation curve according to the method:

the idea of recursive multi-step prediction is shown by the following formula:

prediction(t)＝model(obs(t-1),obs(t-2),...,obs(t-m))

prediction(t+1)＝model(prediction(t),obs(t-1),obs(t-2),...,obs(t-m+1))

prediction data obtained at the time t according to the SVR prediction model, model is the SVR prediction model with the optimal parameters substituted, obs (t-1), obs (t-2), and obs (t-n) are actual training data measured by a monitor before the time t, prediction (t +1) is prediction data obtained after combining recursive multi-step prediction according to the SVR prediction model, every four adjacent data in the step 2) are divided into one group, wherein m is 4, namely the data at the time t +1 are predicted by taking the data at the time 4 before the time t and the prediction data at the time t as the training data, and the prediction data are generated by analogy;

drawing an equipment state degradation curve according to the prediction data;

6) and designing a fault early warning based on a sliding time window according to the data degradation curve.

2. The method for predicting failure of electronic device with improved SVR algorithm as claimed in claim 1, wherein said step 2) adds white gaussian noise to the data set to enhance data on the basis of dividing the data set, and the data enhancement is represented by the following expression:

wherein x (N) is an actual signal in the data set, s (N) is a signal added with noise, N is a signal sampling time, and N is a total signal sampling number; the magnitude of the input noise is set by setting the value of the signal-to-noise ratio SNR.

3. The method for predicting the failure of the electronic device based on the improved SVR algorithm as claimed in claim 1 or 2, wherein the step 4) comprises the following steps:

4.1) population initialization: initializing SVR model parameters including penalty factor C and kernel function parameter sigma²Generating an initial population;

4.2) constructing a fitness function: the fitness function f (x) is constructed by using the fitting error of the SVR prediction model;

4.3) particle swarm algorithm parameter setting: setting population scale, maximum genetic algebra and algorithm termination conditions in the algorithm;

4.4) calculating the individual extreme value and the global optimal solution: inputting an SVR prediction model for training and calculating an individual extreme value and a global optimal solution according to a fitness function;

4.5) update speed and position: updating the speed and the position of each particle according to the calculated individual extreme value and the global optimal solution;

4.6) judging whether the termination condition is met: whether the fitness value meets the requirement or not can be used as a termination condition of the algorithm, and a set maximum genetic algebra can also be used as the termination condition; if the algorithm termination condition is not met, returning to 4.4) to continue executing the operation until the algorithm termination condition is met;

4.7) obtaining the optimal smoothing parameter C and the kernel coefficient sigma²。

4. The method for predicting the failure of the electronic device based on the improved SVR algorithm of claim 3, wherein said step 6) comprises the steps of: on the data degradation curve, the health degree value on the time series is selected by adopting a time frame with a specified unit length, the value corresponding to the specified time is replaced by the average value obtained by calculation in the time frame, and the value corresponding to the specified time is obtained in a sliding time windowSetting a threshold model on the basis of:

wherein E_maxIs the maximum value of the average health value of the equipment, k is the early warning threshold coefficient, E_anFor the fault early warning threshold value, the value output according to each time window and E_anAnd comparing to judge whether to alarm.

Images