Motor residual life prediction method based on neural network and Wiener process

Abstract

The invention discloses a motor residual life prediction method based on a neural network and a Wiener process, which comprises the following steps: firstly, mapping motor characteristic data to a data space conforming to Wiener process characteristics by using a BP (Back Propagation) neural network, and completing parameter estimation of a Wiener model by using Bayesian and EM (Expectation Maximization) algorithms; according to the residual life prediction method, the influence of the failure threshold on the residual life prediction precision is analyzed, finally, all to-be-estimated parameters are learned by using the genetic neural network, and compared with the residual life prediction effect of a traditional Wiener model, the model provided by the invention achieves a better prediction effect, and the prediction uncertainty is effectively quantified. The Wiener model residual life prediction algorithm based on genetic neural network optimization integrates the advantages of machine learning and a probability statistics method, and has good feasibility.

Description

Technical Field

[0001] The present invention belongs to the technical field of the remaining life of motors, and relates to a method for predicting the remaining life of motors based on neural networks and Wiener processes. Background Art

[0002] Motor equipment is extremely widely used and can be seen everywhere in mechanical equipment in scenarios such as military equipment, construction machinery, aerospace, and industrial production; motor equipment is also a key component of most mechanical equipment and has a relatively high failure rate compared to other components. The performance state of the key mechanical components will determine whether the entire mechanical equipment can operate safely and reliably. Therefore, it is extremely important to perform condition detection and remaining service life prediction on motors.

[0003] The development of artificial intelligence has led to extensive research on remaining life prediction algorithms based on deep learning. However, methods based on deep learning often can only provide point estimates of remaining life prediction and cannot quantify the uncertainty of life prediction. In addition, currently, the Wiener model mainly relies on Bayesian methods for parameter estimation, and the influence of the failure threshold on the remaining life prediction accuracy is not considered. As a key parameter for life prediction in the Wiener model, the failure threshold still needs to be determined based on empirical knowledge, which greatly increases the uncertainty of equipment remaining life prediction. To solve the above problems, the present invention proposes a method for predicting the remaining life of motors based on neural networks and Wiener processes. The Wiener process is very suitable for predicting the remaining life of mechanical equipment that conforms to a linear degradation process, and the probability distribution of the remaining life can also be obtained while obtaining the predicted value of the remaining life. This method uses a genetic neural network to estimate the failure threshold and the prior parameters of the Wiener model simultaneously, effectively improving the remaining life prediction accuracy of the Wiener model. In addition, using a genetic algorithm to estimate the parameters of the Wiener model also effectively solves the problem that the original sensor data does not conform to the Wiener independent increment process and requires additional data feature mining, reduces the difficulty of algorithm implementation, and has better practicability. Summary of the Invention

[0004] Aiming at the problems existing in the prior art, a method for predicting the remaining life of motors based on neural networks and Wiener processes is provided, which includes the following steps:

[0005] (1) Map multi-dimensional features to a one-dimensional independent increment degradation space

[0006] The Wiener process is an independent increment process that conforms to the Brownian motion model. The distribution of the increments is only related to time. After the sensor signal data undergoes feature extraction and PCA dimensionality reduction, the feature degradation trend does not conform to the independent increment process of Wiener degradation characteristics. It is necessary to establish a model to mine the deep information of the signal and map it to the state space that conforms to the Wiener process. For convenience, it is assumed here that the mapped data increases linearly in one dimension, and the data change range is from 0 to 1. Since the operating cycles of each motor are not exactly the same, the slopes of the label data for each motor are not consistent. The formula for generating label data is:

[0007]

[0008] In formula (1), t is the corresponding time of the data, T is the entire operating duration of the motor, y is the generated label data, and k represents the slope of the straight line. Using the publicly available motor dataset, the fitting of the data is completed using a BP neural network.

[0009] (2) Wiener model parameter estimation

[0010] Using the degradation data obtained by the BP neural network, the Bayesian and EM algorithms are used to estimate the parameters of the Wiener model at each moment. At the same time, a method is designed to traverse the optimal failure threshold of each device, and the failure threshold is also used as a parameter to be estimated in the Wiener model, and the parameter estimation values at each moment are saved. The specific implementation process is as follows:

[0011] Step1: For each motor, since the motor is a mechanical device and its performance degradation is gradual, the HI interval is 0 - 1. Without loss of generality, its failure threshold is defined at 0.9. For safety, a conservative failure threshold is used, and its range is assumed to be in the range of 0.7 - 0.9. Then the failure threshold is in the interval [0.7, 0.9], and the initial failure threshold is 0.7;

[0012] Step2: Calculate the remaining life of the Wiener model under the current threshold condition, compare it with the true value, and calculate the total RUL prediction error;

[0013] Step3: Increase the failure threshold by 0.01, calculate the remaining life under this threshold condition, and calculate the error at the same time;

[0014] Step4: Repeat Step2 and Step3 until all values in the interval are traversed, and select the value with the smallest RUL prediction error among all failure thresholds as the m estimated value.

[0015] Repeat the above steps for different motors.

[0016] (3) Genetic neural network training

[0017] The genetic neural network is also a network for supervised learning. Using the Wiener model estimation parameters at each moment saved in the previous step as label data and the sensor feature data at the corresponding moment as input data, train the genetic neural network model until the model reaches the preset accuracy or the number of iterations is completed, and save the relevant network parameters.

[0018] (4) Real-time life prediction

[0019] For the sensor data obtained in real time, after preprocessing and feature extraction steps, directly send it into the trained genetic neural network model to obtain the parameter estimation value at the corresponding moment, and substitute the estimation value into the Wiener life prediction model to obtain the remaining life of the device.

[0020] A method for predicting the remaining life of an electric motor based on a neural network and a Wiener process is implemented through the following device. The life prediction framework is mainly divided into three parts. As a supervised learning model, the genetic neural network requires input data and corresponding labels for training: In the first step, use the Bayesian and EM algorithms to estimate the parameters of each device at each moment; in the second step, use the historical monitoring data of the device as the input of the genetic neural network, and the estimation parameters at each moment saved in the previous step as label data to train the genetic neural network; in the last step, use the trained model to predict the remaining life of the relevant device.

[0021] The present invention has the following beneficial effects:

[0022] While obtaining the remaining life value, the fusion model can also obtain the probability distribution of the device life, effectively quantifies the prediction uncertainty, and at the same time effectively reduces the complexity of the implementation process of predicting the remaining life of the Wiener model, proving the superiority and effectiveness of the algorithm invented by the present invention. Description of the drawings

[0023] Figure 1 It is a schematic flow chart of a method for predicting the remaining life of an electric motor based on a neural network and a Wiener process according to the present invention;

[0024] Figure 2 It is a schematic structural diagram of a BP neural network in one embodiment of the present invention;

[0025] Figure 3 It is an experimental result diagram of a BP neural network establishing a health index [0, 1] for the deterioration process in one embodiment of the present invention;

[0026] Figure 4 It is an experimental result diagram of verifying the influence of the failure threshold setting on the remaining life prediction error in one embodiment of the present invention;

[0027] Figure 5It is the motor probability density and remaining life graph provided in one of the embodiments of the present invention;

[0028] Figure 6 It is the training flow chart of the genetic neural network in one of the embodiments of the present invention;

[0029] Figure 7 It is the visualization graph of the genetic neural network structure finally trained and completed in one of the embodiments of the present invention;

[0030] Figure 8 It is the life prediction result graph after loading the trained parameters into the Wiener process model in one of the embodiments of the present invention; Detailed implementation manners

[0031] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0032] As Figure 1-8 shown, a method for predicting the remaining life of a motor based on a neural network and a Wiener process includes the following steps:

[0033] S1: Estimate the parameters of each device at each moment by using the Bayesian and EM algorithms.

[0034] The Wiener process can well describe the randomness of the device during the degradation process, and can model both the monotonic and non-monotonic random processes of the degradation trend. The Wiener process model is:

[0035] X(t) = X(0) + ηt + σB(t) (2)

[0036] In the formula, t is the corresponding moment of the data, X(0) is the initial state, which is generally set to 0; η is the drift coefficient, representing the degradation rate of the device, and is a random variable obeying the Gaussian distribution; σ is the diffusion coefficient, representing the volatility of the device's degradation process due to various random factors; B(t) is the standard Brownian motion, satisfying B(t) ~ N(0, t).

[0037] The degradation amount Y(t) of the device monitored at any moment is:

[0038] Y(t) = X(t) - X(0) = ηt + σB(t) (3)

[0039] Assume W k is the remaining life of the device at time t k Based on the concept of the first passage time, there is:

[0040] W k = inf{wk : Y(w k +t k ) ≥ m} (4)

[0041] Y(w k +t k ) = η(w k +t k ) + σB(w k +t k ) = y k +ηw k +H(w k ) (5)

[0042] where m is the threshold representing when the device fails, usually set according to expert knowledge or industry standards; y k is the degradation amount of the device at time t k ; H(w k ) = B(w k +t k ) - B(t k ). According to the properties of Brownian motion, H(w k ) is also a standard Brownian motion. The stochastic process represented by equation (4) can be regarded as a Brownian motion with an initial value of Y(t k ) and a drift coefficient of η.

[0043] From the above definitions, the conditional probability density function corresponding to the remaining life W k is:

[0044]

[0045] Then, by the total probability formula, the probability density of the remaining life W k of the device can be obtained as:

[0046]

[0047] (1) Bayesian posterior estimation of random parameters

[0048] Assume that the prior distribution of the parameter η is p(η), then:

[0049]

[0050] Since η|Y 0:k is a conditional random variable and p(η|Y 0:k ) is a probability density, according to the characteristics of random variables with a normal distribution, it can be known that η|Y 0:k is also normally distributed, and there is:

[0051]

[0052] Wherein:

[0053]

[0054] After obtaining new device monitoring data each time, the posterior estimation of the random parameter η can be updated according to Equation (9).

[0055] (2) Determine the estimation of unknown parameters by the EM algorithm

[0056] In order to update the parameters, at the beginning, we need to estimate the unknown parameters according to the obtained monitoring data. Assume the parameter to be estimated To reflect the time-update characteristic of Θ, denote based on Y 0:k the unknown parameter to be estimated, denote the result of its maximum likelihood estimation. Log-likelihood function:

[0057]

[0058] Assume that the EM algorithm is based on Y 0:k the estimation value at the i-th step obtained Take the expectation of the likelihood function, then there is:

[0059]

[0060] Let Then the likelihood estimation of the parameter can be obtained as:

[0061]

[0062] In the parameter update based on the EM algorithm, each step of recursion has an explicit formula. Therefore, the maximization of each M-step in the EM algorithm only requires one calculation, which makes the process of parameter estimation and update simple and fast. According to the above, substituting the parameters can obtain the probability density function and distribution function of the remaining life.

[0063] The prediction error values of the remaining life of each motor change with the failure threshold as Figure 4 shown. In the regional range, the probability density function and the prediction of the remaining service life corresponding to the optimal failure threshold Wiener model are as Figure 5 shown. So far, the estimation of all the parameters to be estimated in the model has been completed, and the estimated values of the motor parameters at each moment are saved for the next step of neural network model training.

[0064] S2: Use the historical monitoring data of the device as the input of the genetic neural network, and use the estimated parameters at each moment saved in the previous step as the label data to train the genetic neural network.

[0065] As Figure 6As shown in the figure, for the convenience of network training and to improve the generalization ability of the network, only two significant digits are retained for the Wiener estimation parameters saved in the previous step. A motor dataset is randomly selected as the test set, and other motor data are used as the training set for network training. The genetic neural network is used to fit all the parameters to be estimated. The network input dimension is set to 6, and the network output dimension is set to 4. The genetic algorithm parameters are set as follows: the population size is 50, the number of genetic generations is 100, the selection probability is 90%, the crossover probability is 70%, and the mutation probability is 2%.

[0066] During the training process, the genetic algorithm will continuously search for the network with the best fitting effect, and find the one with the best fitness in the population through operations such as crossover, mutation, and inheritance. During the training process of the genetic neural network, as the number of iterations increases, the population fitness also gradually increases. Here, the larger the fitness value, the better the network training effect for us. As Figure 7 shown in the figure, the neural network structure after the final training is visualized. The network contains 6 input neurons, 5 hidden layer neurons, and 4 output layer neurons. The solid lines in the figure represent that the connections are in the activated state, and the dashed lines represent the non-activated state. The red lines indicate that the connection weights are less than 0, and the green lines indicate that the weights are greater than 0. The thickness of the lines represents the magnitude of the weights.

[0067] S3: Use the trained model to predict the lifespan of related equipment.

[0068] As Figure 8 shown in the figure, use the trained model to predict the lifespan. Substitute the test set parameters obtained by the genetic neural network into the Wiener model, calculate the lifespan probability density function and the remaining lifespan of the equipment according to the formula, and compare the remaining lifespan prediction effects with those obtained by the estimation parameters based on the Bayesian and EM algorithms. Compare the remaining lifespan prediction effects of the two under the condition of using the best failure threshold.

[0069] In the case of using the mean square error, the remaining lifespan prediction error using the estimation parameters of the genetic neural network is less than the prediction error of the Wiener model. It can be concluded that the remaining lifespan prediction model proposed in the present invention has a good remaining lifespan prediction effect. And based on the characteristics of the neural network method that can effectively mine the deep features of data and reduce noise interference, the Wiener process lifespan prediction model optimized by the genetic neural network has a smaller prediction error and is superior to the traditional Wiener process remaining lifespan prediction model.

Claims

1. A method for predicting the remaining useful life of an electric machine based on a neural network and a Wiener process, characterized in that, It includes the following steps: (1) Mapping multi-dimensional features to a one-dimensional independent incremental degradation space The Wiener process is an independent incremental process that conforms to the Brownian motion model. The distribution of the increments only depends on time. After the sensor signal data undergoes feature extraction and PCA dimensionality reduction, the feature degradation trend does not conform to the independent incremental process of Wiener degradation characteristics. It is necessary to establish a model to mine the deep information of the signal and map it to the state space that conforms to the Wiener process. For convenience, it is assumed here that the mapped data increases linearly in one dimension, and the data change range is from 0 to 1. Since the operating cycles of each motor are not exactly the same, the slopes of the label data for each motor are not consistent. The formula for generating label data is: In formula (1), t is the corresponding time of the data, T is the entire operating duration of the motor, y is the generated label data, and k represents the slope of the straight line; the fitting of the data is completed using a BP neural network; (2) Wiener model parameter estimation Using the degradation data obtained by the BP neural network, the Bayesian and EM algorithms are used to estimate the parameters of the Wiener model at each moment. At the same time, a method is designed to traverse the optimal failure threshold of each device, and the failure threshold is also used as a parameter to be estimated in the Wiener model, and the parameter estimation values at each moment are saved. The specific implementation process is as follows: Step1: For each motor, since the motor is a mechanical device and its performance degradation is gradual, the HI interval is 0 - 1. Without loss of generality, its failure threshold is defined as 0.

9. For safety, a conservative failure threshold is used, and its range is assumed to be in the range of 0.7 - 0.

9. Then the failure threshold is in the interval [0.7, 0.9], and the initial failure threshold is 0.7; Step2: Calculate the remaining life of the Wiener model under the current threshold condition, compare it with the true value, and calculate the total error of RUL prediction; Step3: Increase the failure threshold by 0.01, calculate the remaining life under this threshold condition, and calculate the error at the same time; Step4: Repeat Step2 and Step3 until all values in the interval are traversed, and select the value with the smallest RUL prediction error among all failure thresholds as the m estimated value; (3) Genetic neural network training The genetic neural network is also a supervised learning network. The Wiener model estimation parameters at each moment saved in the previous step are used as label data, and the sensor feature data at the corresponding moment is used as input data to train the genetic neural network model until the model reaches the preset accuracy or the number of iterations is completed, and the relevant network parameters are saved; (4) Real-time life prediction For the sensor data obtained in real time, after preprocessing and feature extraction steps, it is directly sent into the trained genetic neural network model to obtain the parameter estimation values at the corresponding moment, and the estimated values are substituted into the Wiener life prediction model to obtain the remaining life of the device.

Images