A method and system for predicting the remaining life of mechanical equipment

Abstract

The present invention discloses a method and system for predicting the remaining life of mechanical equipment, which belongs to the field of condition monitoring and remaining life prediction. The present invention combines a convolutional neural network with a bidirectional gated recurrent unit to form a hybrid neural network, so as to effectively extract time and space features and improve the remaining life prediction accuracy. On this basis, attention weights are introduced into the hybrid neural network, and a collaborative training mechanism of a hybrid neural network and an optimized genetic algorithm is adopted. The genetic algorithm continuously transmits parameters and feedback losses to find the best attention weight parameters, optimizes the attention weight distribution of different time steps in the remaining life prediction, and accurately reflects the importance of different time steps. Compared with some traditional machine learning methods and deep learning, its root mean square error and prediction score are much lower, which effectively solves the problem of low accuracy of existing remaining life prediction.

Description

Technical Field

[0001] The present invention belongs to the field of condition monitoring and remaining life prediction, and more specifically, relates to a method and system for predicting the remaining life of mechanical equipment. Background Art

[0002] In recent years, with the rapid development of intelligent manufacturing, condition monitoring technology has been widely adopted in various engineering systems, such as wind energy conversion systems, battery systems, rolling bearing systems, etc. Monitoring technology is of great significance for prediction and health management. Through the analysis of monitoring data, people can perform remaining life prediction to estimate the available life of the equipment before failure. Accurate remaining life prediction of industrial equipment such as turbofan engines, bearings, etc. can prevent machine failures in time and effectively reduce maintenance costs. Based on recent research, the remaining life prediction technology of mechanical equipment can be mainly divided into four categories: physical model-based methods, statistical model-based methods, machine learning-based methods, and fusion methods.

[0003] In recent years, deep learning algorithms are often used to automatically obtain deeper abstract features from large-scale data, and are widely used in the field of remaining life prediction. Common methods include convolutional neural networks, recurrent neural networks, long short-term memory artificial neural networks, gated recurrent units, etc. In most existing traditional methods, the same weight is usually assigned to data points with different time steps. But in fact, these data points with different time steps do not provide the same rich information, so their prediction accuracy is not ideal. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention provides a method and system for predicting the remaining service life of mechanical equipment, which aims to assign different weights to data points of different time steps to highlight data points containing more degradation information, thereby solving the technical problem of low accuracy in remaining service life prediction using traditional methods.

[0005] To achieve the above object, the present invention provides a method for predicting the remaining life of mechanical equipment, comprising:

[0006] S1. Obtain and pre-process the operation data of mechanical equipment as training set and verification set;

[0007] S2. Constructing a hybrid neural network; the hybrid neural network includes a convolutional neural network and a bidirectional gated recurrent unit; wherein the convolutional neural network is used to extract data information under the spatial dimension of input data; the input data is obtained by weighting data points of different time steps with corresponding attention weights; the bidirectional gated recurrent unit is used to capture the time correlation in the output data of the convolutional neural network;

[0008] S3. The initialized attention weights are passed to the hybrid neural network. The network is trained using the training set and the loss is fed back to the genetic algorithm. The genetic algorithm continuously learns to obtain the optimal set of attention weights, which are then passed to the hybrid neural network to achieve collaborative training and find the optimal solution. When the loss on the validation set is minimized, the trained prediction model is returned.

[0009] S4. Input the operating status data of the mechanical equipment to be predicted into the trained prediction model to obtain the remaining life prediction result.

[0010] Furthermore, the preprocessing in step S1 includes:

[0011] The data were normalized according to the following formula:

[0012]

[0013] Among them, x i,j is the raw sensor data, is the true value of the standardized data, i represents the i-th sensor, j represents the j-th data point, is the maximum and minimum value of the i-th sensor data;

[0014] All data are segmented using a sliding window.

[0015] Furthermore, the number of channels of the convolutional neural network is set to the time step.

[0016] Furthermore, the bidirectional gated recurrent unit is calculated as follows:

[0017]

[0018] Where X t ,h t-1 ,h t ,r t ,z t , y t Respectively represent the input vector, the state storage variable at the previous moment, the state storage variable at the current moment, the state of the reset gate, the state of the update gate, the state of the current candidate set, and the state of the output vector at the current moment; W r ,W z , W O Represents the reset gate, update gate, candidate set, output vector and x respectively. t and h t-1 The weight parameter of the connection matrix formed; I represents the identity matrix; [] represents vector connection; · represents matrix dot product; × represents matrix product; σ represents the S-shaped activation function; tanh represents the tanh activation function; b r , bz 、b h Represents the bias parameter learned during training.

[0019] Furthermore, the steps of the genetic algorithm are specifically as follows:

[0020] Population initialization: The attention weights are encoded into the chromosomes of the genetic space through binary codes, and each attention weight corresponds to a string of binary codes;

[0021] Weight transfer: Convert the binary code into a decimal as the attention weight value, and transfer the attention weight to the hybrid neural network. The hybrid neural network starts training and learning and returns the corresponding loss value generated by the prediction error.

[0022] Loss sorting: Randomly select each group of binary codes and divide them into several groups. According to the returned loss sorting, select the weighted code with the smallest error value in each group;

[0023] Crossover recombination: Randomly combine the minimum error sets selected in the previous step in pairs to obtain multiple sets of set 1 and set 2. Randomly select some gene points in set 1, and the remaining gene points that are not selected are provided by set 2. Recombine the gene points in these two sets into a new gene set to obtain the recombined gene fragments.

[0024] Mutation: Perform the inversion operation on the original gene according to the mutation probability of the genotype after crossover recombination;

[0025] Generate a new population: Rebuild the population based on crossover and mutation, and start new training.

[0026] In general, compared with the prior art, the above technical solution conceived by the present invention can achieve the following beneficial effects.

[0027] The present invention combines a convolutional neural network with a bidirectional gated recurrent unit to form a hybrid neural network, so as to effectively extract time and space features and improve the remaining life prediction accuracy. On this basis, attention weights are introduced into the hybrid neural network, and a collaborative training mechanism of a hybrid neural network and an optimized genetic algorithm is adopted. The two share and transfer parameters and loss feedback. The genetic algorithm finds the optimal attention weight parameters by continuously transferring parameters and feeding back losses, thereby optimizing the attention weight distribution of different time steps in the remaining life prediction, so that the importance of different time steps can be accurately reflected. Compared with some traditional machine learning methods and deep learning (multilayer perceptron, support vector regression, random forest, convolutional neural network, recurrent neural network, deep belief network, etc.), its root mean square error and prediction score are much lower, which effectively solves the problem of low accuracy of existing remaining life prediction.

[0028] The present invention also improves the traditional genetic algorithm to prevent the algorithm from falling into a local optimal state, and introduces the loss feedback of the hybrid neural network into the improved genetic algorithm, so that the process can significantly improve the performance by controlling the direction of random search. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A schematic diagram of a method for predicting the remaining life of a turbofan engine based on a genetically optimized hybrid neural network provided by an embodiment of the present invention;

[0030] Figure 2 The measurements are from 14 selected sensors for all engines of the FD001 subset;

[0031] Figure 3 The true remaining life and predicted remaining life obtained by the proposed method on two datasets;

[0032] Figure 4 Attention distribution diagram in different experimental datasets. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0034] The present invention provides a method for predicting the remaining life of mechanical equipment, such as Figure 1 As shown, it includes: preprocessing the operation data of mechanical equipment; model establishment and training: first, the initialized attention weight is passed to the convolution-bidirectional gated unit hybrid neural network, and then the model is trained and the loss is fed back to the improved genetic algorithm. The genetic algorithm continuously learns to obtain the optimal set of attention weights, and then passes it to the hybrid neural network to achieve collaborative training and find the optimal solution. When the loss on the validation set is minimized, the model is returned; the returned model is predicted on the test set, and the remaining life prediction result is returned. The model established by the present invention is composed of a convolutional neural network and a bidirectional gated recurrent unit to form a framework, wherein the convolutional neural network extracts the spatial features of the data points, and the bidirectional gated recurrent unit obtains the temporal correlation of the data. The combination of the two can extract both spatial features and temporal features, and learn the complete features in the data.

[0035] The following is a specific description in conjunction with an embodiment. For example, the measurement values ​​of the sensors of all engines in the FD001 subset are obtained, and 14 sensor data with appropriate variances and sufficient degradation information are selected for subsequent use, such as Figure 2 , where the horizontal axis represents time, and the unit is cycle. The data is normalized to eliminate the influence between different sensor characteristics. At the same time, for data with a long time, a sliding window is used to segment the data. The specific implementation method is as follows:

[0036] (11) The normalization operation is performed by the following formula:

[0037]

[0038] Among them, x i,j is the raw sensor data, is the true value of the standardized data, i represents the i-th sensor, j represents the j-th data point, is the maximum and minimum value of the i-th sensor data.

[0039] (12) Sliding windows are used for segmentation. In order to capture the temporal correlation between data, sliding windows are used to encapsulate data at adjacent time points. The remaining life prediction of the last data point in the time window is used as the remaining life prediction of the time window. Through analysis, the data sliding window time length is set to 30 and the step size is set to 1.

[0040] Preferably, the hybrid neural network introduced with attention weights used for training is composed of a convolutional neural network and a bidirectional gated recurrent unit, and the whole is:

[0041] Y=f H (x)

[0042] in, is a vector containing m time steps and n features, Y is the network output representing the remaining life prediction, and f H (·) is a neural network consisting of nonlinear functions that map input to output.

[0043] The specific training method is as follows:

[0044] (21) First, the original input data is selected through a sliding window, and then the attention weights at different time steps are defined as follows:

[0045]

[0046] i represents the i-th weight set, l represents the l-th time step, represents the weight of the lth time step in the i-th weight set;

[0047] By weighted calculation of the attention weights at different time steps, the following input data can be obtained:

[0048]

[0049] represents the value of the lth data point in the tth sample;

[0050] Thus, the data value X to which the attention weight is assigned is X=[X 1 ,X 2 ,...,X T ] is transferred to the convolutional-bidirectional recurrent gated hybrid neural network, X T Represents the value of the data point of the Tth sample multiplied by the attention weight.

[0051] (22) The convolutional neural network is used to obtain the spatial features of the attention-weighted input data. In order not to compress the temporal correlation of the data, the number of channels is set to the time step. Therefore, the convolutional neural network does not interfere with the temporal relationship in the data, but only compresses the spatial information. The output data is then input into the bidirectional gated recurrent neural network.

[0052] (23) The gated recurrent unit learns the spatial features extracted by the convolutional neural network and learns the temporal correlation dependency. The calculation of the unit is as follows:

[0053]

[0054] Where X t ,h t-1 ,h t ,z t ,r t , t Respectively represent the input vector, the state storage variable at the previous moment, the state storage variable at the current moment, the state of the update gate, the state of the reset gate, the state of the current candidate set, and the state of the output vector at the current moment. z ,W r , o Represent the update gate, reset gate, candidate set, output vector and x respectively. t and h t-1 The weight parameters of the connection matrix formed. I represents the identity matrix; [] represents vector connection; · represents matrix dot product; × represents matrix product; σ represents the S-shaped activation function; tanh represents the tanh activation function. b r 、b z 、b h is the bias learned during training;

[0055] The genetic algorithm sends the learned attention weights back to the hybrid neural network. At the same time, the hybrid network returns the training loss to the genetic algorithm to guide its further learning and training. The present invention can use a traditional genetic algorithm and a hybrid neural network for collaborative training. Preferably, the present invention also makes the following improvements to the traditional genetic algorithm. The main steps are as follows:

[0056] Population initialization: The attention weights are encoded into the chromosomes of the genetic space through binary code, and each weight corresponds to a string of binary gene codes;

[0057] Weight transfer: Convert the binary code into a decimal, that is, the initial attention weight value, and transfer the attention weight to the previous hybrid neural network. The neural network starts training and learning, and returns the corresponding loss value generated by the prediction error of the neural network.

[0058] Loss sorting: Randomly select each group of binary codes and divide them into several groups. At the same time, sort them according to the returned loss (prediction error) and select the weighted code with the smallest error value in each group.

[0059] Crossover recombination: Randomly combine the minimum error sets selected in the previous step in pairs to obtain multiple sets of set one and set two, randomly select some gene points in set one, and the remaining gene points that are not selected are provided by set two, and the gene points in these two sets are recombined into a new gene set to obtain the recombined gene fragments; the gene points of the selected set here are not fixed, and 50% are selected in this embodiment; ordinary genetic algorithms randomly select gene points for exchange, but this method is not an exchange but a recombination, randomly select gene points in set one, and recombine the selected gene points of set one with the unselected gene points of set two to obtain a new genotype. For example: there are two six-bit gene fragments, and the gene points at the three positions 1, 3, and 5 in set one are randomly selected (246 is not selected, and 246 is provided by set two), then the gene point 135 of set one and the gene point 246 of set two are recombined into a new 6-bit gene fragment;

[0060] Mutation: Perform the inversion operation on the original gene according to the mutation probability of the genotype after crossover recombination;

[0061] Generate a new population: Rebuild the population based on crossover and mutation, and start new training.

[0062] The differences between the improved genetic algorithm of the present invention and the common genetic algorithm include:

[0063] The selection strategy in ordinary genetic algorithms is generally to first calculate the probability of each individual appearing in the offspring to randomly select individuals to form a population, while the present invention randomly combines the selected minimum error sets in pairs to obtain new sets for the next step of crossover mutation, so that the crossover mutation process has more possible rows.

[0064] The crossover operator in the common genetic algorithm is mainly single-point or multi-point crossover, which exchanges genes by randomly selecting crossover points. The new gene fragments finally obtained still include part of the original genotype. The present invention improves the crossover operator, mainly by randomly recombinating two randomly combined gene fragments to obtain new gene fragments. Compared with the common method, the present method increases the difference between offspring, expands the diversity of its global search, can exploit more position space, and avoid the algorithm's early convergence to fall into the local optimum.

[0065] From the above two points, it can be seen that the improved genetic algorithm of the present invention mainly adds a random mechanism in the selection strategy and crossover operator to reconstruct the final population space compared to the ordinary genetic algorithm, so that the algorithm has a stronger global search capability.

[0066] In addition, the present invention introduces the loss feedback of the previous hybrid neural network into the improved genetic algorithm, so that the improvement can make the process significantly improve the performance by controlling the direction of random search.

[0067] Figure 3 The predicted remaining life and the actual remaining life prediction curves obtained by the above method on the two data sets FD001 and FD004 are shown. It can be found from the figure that the actual value and the predicted value are very close, which shows that this method has a high prediction accuracy.

[0068] Figure 4 The attention distribution diagrams of different experimental datasets are shown to facilitate a better understanding of the importance sampling of different time steps in the time series.

[0069] Table 1 compares the effects of convolutional neural networks plus several different recurrent neural networks by controlling the type of hybrid neural network and whether to introduce attention weights. It also compares the ordinary genetic algorithm and the improved genetic algorithm. By comparing the root mean square error and score of each method, it can be found that the combination of convolutional neural network and bidirectional gated recurrent unit plus improved genetic algorithm has significant superiority. Table 2 compares the method of the present invention with other commonly used methods for remaining life prediction. The results are the same. The method of the present invention is better than other methods in both evaluation indicators. Among them, the two model evaluation indicators are the root mean square error (RMSE) and score. The score is a model evaluation method commonly used in the field of remaining life prediction. The two evaluation indicators are calculated as follows:

[0070]

[0071]

[0072] Among them, n is the number of data, y i is the true value of the remaining life, The remaining life prediction value, a 1 =10, a 2 =13, where a 1 、a 2 It is a commonly used specified value.

[0073] Table 1: Comparative performance of different remaining life prediction methods

[0074]

[0075]

[0076] Table 2: Experimental results on different datasets of FD001 and FD004

[0077]

[0078] In summary, compared with the prior art, the above technical solution conceived by the present invention can achieve higher prediction accuracy with relatively limited technical complexity.

[0079] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of mechanical equipment, characterized in that: include: S1. Obtain and pre-process the operation data of mechanical equipment as training set and verification set; S2. Construct a hybrid neural network; the hybrid neural network includes a convolutional neural network and a bidirectional gated recurrent unit; wherein the convolutional neural network is used to extract data information under the spatial dimension of the input data; the input data is obtained by weighting the data points of different time steps with corresponding attention weights; the bidirectional gated recurrent unit is used to capture the time correlation in the output data of the convolutional neural network; the calculation of the bidirectional gated recurrent unit is as follows: Where X t ,h t-1 ,h t ,r t ,z t , y t Respectively represent the input vector, the state storage variable at the previous moment, the state storage variable at the current moment, the state of the reset gate, the state of the update gate, the state of the current candidate set, and the state of the output vector at the current moment; W r ,W z , W O Represents the reset gate, update gate, candidate set, output vector and x respectively. t and h t-1 The weight parameter of the connection matrix formed; [] represents vector connection; · represents matrix dot product; × represents matrix product; σ represents S-shaped activation function; tanh represents tanh activation function; b r , b z , b h Represents the bias parameter learned during the training process; S3. The initialized attention weights are passed to the hybrid neural network. The network is trained using the training set and the loss is fed back to the genetic algorithm. The genetic algorithm continuously learns to obtain the optimal set of attention weights, which are then passed to the hybrid neural network to achieve collaborative training and find the optimal solution. When the loss on the validation set is minimized, the trained prediction model is returned. S4. Input the operating status data of the mechanical equipment to be predicted into the trained prediction model to obtain the remaining life prediction result; The specific steps of the genetic algorithm are: Population initialization: The attention weights are encoded into the chromosomes of the genetic space through binary codes, and each attention weight corresponds to a string of binary codes; Weight transfer: Convert the binary code into a decimal as the attention weight value, and transfer the attention weight to the hybrid neural network. The hybrid neural network starts training and learning and returns the corresponding loss value generated by the prediction error. Loss sorting: Randomly select each group of binary codes and divide them into several groups. According to the returned loss sorting, select the weighted code with the smallest error value in each group; Crossover recombination: Randomly combine the minimum error sets selected in the previous step in pairs to obtain multiple sets of set 1 and set 2. Randomly select some gene points in set 1, and the remaining gene points that are not selected are provided by set 2. Recombine the gene points in these two sets into a new gene set to obtain the recombined gene fragments. Mutation: Perform the inversion operation on the original gene according to the mutation probability of the genotype after crossover recombination; Generate a new population: Rebuild the population based on crossover and mutation, and start new training.

2. A method for predicting the remaining life of mechanical equipment according to claim 1, characterized in that: The preprocessing in step S1 includes: The data was normalized according to the following formula: Among them, x i,j is the raw sensor data, is the true value of the standardized data, i represents the i-th sensor, j represents the j-th data point, is the maximum and minimum value of the i-th sensor data; All data are segmented using a sliding window.

3. A method for predicting the remaining life of mechanical equipment according to claim 1 or 2, characterized in that: The number of channels of the convolutional neural network is set to the time step.

4. A mechanical equipment remaining life prediction system, characterized in that: It includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement a method for predicting the remaining life of mechanical equipment as described in any one of claims 1 to 3.

5. A machine-readable storage medium, characterized in that: The machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement a method for predicting the remaining life of a mechanical device as described in any one of claims 1 to 3.

Images