CN110866314B - Rotating Machinery Remaining Lifetime Prediction Method Based on Multilayer Bidirectionally Gated Recurrent Unit Networks - Google Patents

Abstract

The invention discloses a method for residual life of a rotary machine of a multilayer bidirectional gating circulation unit network, which comprises the steps of collecting vibration signals; constructing a health index; constructing a network training set; constructing a multi-layer bidirectional gating circulation unit network; training a multi-layer bidirectional gating circulation unit network; network testing, estimation of residual life and acquisition of a confidence interval; and (4) predicting and evaluating the residual life. The method combines the advantages of strong feature extraction capability of deep learning, utilizes the neural network of the bidirectional gating circulation unit to carry out regression prediction, and obtains the confidence interval of the residual life by a Bootstrap method. Aiming at the problems that the model precision of a recurrent neural network model is sensitive to the value of the learning rate in the training process, and the prediction performance of the model is influenced by overhigh and overlow, the natural index is utilized to attenuate the learning efficiency to train the neural network efficiently. The method can accurately predict the residual life and the confidence interval of the rotary machine, can greatly reduce expensive unscheduled maintenance, and avoids the occurrence of disasters.

Description

Rotating Machinery Remaining Lifetime Prediction Method Based on Multilayer Bidirectionally Gated Recurrent Unit Networks

technical field

The invention relates to a technology for predicting the remaining life of a rotating machine, in particular to a method for predicting the remaining life of a rotating machine based on a multi-layer bidirectional gated cycle unit network.

Background technique

Due to the development of advanced sensor and computer technology, a large amount of condition monitoring data has been accumulated in industrial production, and data-driven methods have been widely used in the prediction of mechanical equipment remaining life, because they can use condition monitoring data to quantify the degradation process, while Rather than building an accurate system model that is not readily available.

Fault prediction and health management includes four stages: fault detection, fault diagnosis, remaining life prediction and health management. When a fault is discovered or diagnosed, the machine is usually shut down as quickly as possible to avoid catastrophic consequences. Performing such actions usually takes place at inconvenient times and often results in considerable time and financial loss. Therefore, maintenance strategies must be scheduled in a predictive rather than diagnostic fashion. Consider accurate remaining life predictions for efficient machinery predictive maintenance, reducing costly unscheduled machinery repairs. From this perspective, it is crucial to consider residual life prediction for efficient machinery predictive maintenance. Common rotating parts in industrial sites, such as bearings, gears, rotors, etc., are important components of rotating machinery, and their health directly affects the normal operation of rotating machinery. Severe damage to these key components will lead to production shutdown and huge economic losses. Therefore, accurate prediction of the remaining life of mechanical equipment is of great significance for the safe and reliable operation of equipment.

The existing remaining life prediction methods mainly have the following problems: (1) Deep learning methods can effectively mine the hidden features of sensor data and provide better point estimates for remaining life prediction. Predictions often vary widely due to measurement noise and model parameters. Only point estimation of remaining life cannot meet the actual requirements. In order to express the uncertainty of the prediction results, it is necessary not only to calculate the determined predicted value of remaining life, but also to calculate the confidence interval of remaining life; (2) The remaining life prediction with deep learning method is to use fixed learning efficiency to train the network, which is inefficient (3) The model-based method tries to establish a mathematical or physical model to describe the degradation process of the machinery, and uses the measured data to update the model parameters. In practice, it is difficult to find an accurate model to describe the degradation process of the rotating machinery.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method for predicting the remaining life of a rotating machine with a network of multi-layer bidirectional gated cyclic units. The prediction method can improve the prediction accuracy, effectively obtain the confidence interval of the remaining life, and then provide reliable suggestions for the operation and maintenance of mechanical equipment to avoid catastrophic consequences.

In order to solve the above-mentioned technical problems, the technical solution adopted in the present invention is: a method for predicting the remaining life of rotating machinery in a multi-layer bidirectional gated cyclic unit network, comprising the following steps:

Step 1, collecting vibration signals: collecting vibration signals of key components of rotating machinery.

Step 2, construction of health indicators: construct health indicators describing the degradation process of key components of rotating machinery.

Step 3, using sliding window technology to construct training and testing data sets.

Step 4, build the network: construct the MBi-GRU neural network by stacking a single bidirectional GRU network.

Step 5, build an integrated learning network through the Bootstrap method to obtain the uncertainty expression of the prediction results. The input and output of the model are the data obtained in step (3).

Step 6, training the network: the training of the entire model is performed by minimizing the loss function through error reverse transmission. The present invention proposes to train the MBi-GRU neural network with natural attenuation learning efficiency.

Step 7. Acquisition of the mean value of remaining life and confidence interval: input the data of the test set into the network trained in step (5), and the corresponding output of the test set is the desired remaining life prediction value and confidence interval.

Step 8, evaluating the prediction result of the remaining life: the present invention uses two indexes to evaluate: mean square error and prediction absolute error.

A further improvement of the present invention lies in that: in the step 2, the original features are extracted from the vibration signal. Select the features with a large trend value to form a feature subset, and finally use the unsupervised SOM algorithm to construct a health index for evaluating the life of the bearing from the selected feature subset;

A further improvement of the present invention is that: in the step 4, the update gate controls the current input data x _t and the previous memory information h _t-1 at the same time, and outputs a value z _t between 0 and 1, and the calculation formula is

z _t ＝σ(W _z [h _t-1 ,x _t ]+b _x ) (1)

In the formula, x is the input data, h is the output of the GRU unit, r is the reset gate, z is the update gate, r and z jointly control how to calculate the new hidden state h _t from the previous hidden state h _t-1 .

z _t decides how much to transfer h _t-1 to the next state, which can be obtained from formula (1). In the formula, σ is the sigmoid function, W _z is the update gate weight, and b _z is the bias. The reset gate controls how important ht _-1 is to the outcome _ht . If the previous memory h _t-1 is completely unrelated to the new memory, the reset gate can work to remove the influence of the previous memory, i.e.

r _t ＝σ(W _r [h _t-1 ,x _t ]+b _r ) (2)

即Generate new memory information according to the update gate

which is

The output at the current moment is h _t , namely

和反向的隐层状态的输出

三个部分共同决定。The basic unit of the Bi-GRU model consists of a GRU unit for forward propagation and a GRU unit for backward propagation. In a unidirectional neural network structure, states are always output from front to back. However, in the remaining life prediction, if the output at the current moment can be related to the state at the previous moment and the state at the next moment. The current hidden layer state of Bi-GRU is determined by the current input x _t , the hidden layer state forward at time t-1

and the output of the reversed hidden layer state

The three parts are jointly determined.

和反向的隐层状态的输出

所对应的权重，b_t表示t时刻隐层状态所对应的偏置。Among them: the GRU() function represents the nonlinear transformation of the input mechanical equipment degradation index, and encodes the degradation index into the corresponding GRU hidden layer state. w _t and v _t respectively represent the forward hidden layer state corresponding to the bidirectional GRU at time t

and the output of the reversed hidden layer state

The corresponding weight, b _t represents the bias corresponding to the state of the hidden layer at time t.

Three Bi-GRU layers and a fully connected regression layer are used to perform regression prediction on the constructed health indicators. Using a 3-layer model can increase the parameters of the model and improve the learning ability of the model. The hidden state of each layer has two information flow directions, 1 is transmitted to the next moment, and 2 is used as the input of the next layer at the current moment. The MBi-GRU model can make full use of the past and future relevant information of the degradation state of mechanical equipment and effectively transfer between layers, thereby improving the accuracy of remaining life prediction.

The further improvement of the present invention is that: in the step 5, the health index has N values, namely (z(t ₁ ), z(t ₂ ),..., z(t _N )), n=1, 2, ..., N. Let L be the length of the sliding window. (z(t _i ),z(t _i+1 ),...,z(t _i+L-1 )) a degenerated state with a window length, the corresponding output is z(t _i+L ), N A degraded state data can generate NL samples, and the predicted system state can be expressed by the following functional relationship

z(t _i+L )=φ(z(t _i ),z(t _i+1 ),...,z(t _i+L-1 )) (8)

Confidence intervals for forecast results are obtained to quantify the uncertainty of point estimates. Resample K times from the original training data with an alternative method. The model φ _k (k=1, 2, . . . , K) is trained with resampled data S _k (t ₁ :t _i ) (k=1, 2, . . . , K) each time. Finally, the ensemble operation of multiple models will produce the mean and variance of the remaining life prediction results. The above description is expressed by the formula as follows

Z _k (l _i +t _i )＝φ _k (S _k (t ₁ :t _i )) (9)

Among them, Z _k (l _i +t _i ) represents the state prediction value of the system obtained by the Bootstrap method.

The remaining lifetime l _i at time t _i is defined as follows:

l _i ＝inf{l _i :Z _k (l _i +t _i )≥τ|z(t ₁ :t _i )} (10)

Among them, τ is the preset failure threshold; z _0:i is the estimated system state value from t ₀ to t _i time, Z _k (t _i + l _i ) is the estimated system state at t _i + l _i time value. The final confidence interval of the remaining life can be obtained by calculating the percentile of the remaining life l _i at time t _i .

The further improvement of the present invention is: in the step 6, at the initial stage of network training, the learning efficiency is set higher, and when gradually approaching the optimal solution, the learning efficiency is gradually reduced so as to get closer to the optimal value of the network. Using the natural exponential decay learning efficiency to train the deep recurrent neural network can effectively train the neural network. The calculation method of the natural exponential decay learning efficiency is as follows:

Among them, lr: the current learning rate, lr ₀ : the initial learning rate, r _decay : the decay rate of each round of learning, 0<r _decay <1, S _global : the current number of global learning steps, S _decay : each round of learning The number of steps, S _decay =N _sample /N _batch , that is, the total number of samples divided by the size of each batch.

The further improvement of the present invention is: in step 8, mean square error, i.e. Mean Squared Error, MSE and prediction absolute error, i.e. Mean Absolute Percentage Error, MAPE; Calculation formula is as follows:

Among them, HI _Act is the real bearing degradation state, HI _Pre is the predicted bearing degradation state, and N _p is the number of predicted points.

Beneficial effect:

1. Data-driven remaining life prediction methods do not rely on fixed degradation models;

2. The confidence interval of the remaining life can be effectively obtained through the Bootstrap method, providing reliable knowledge for the operation and maintenance of equipment;

3. The MBi-GRU model can make full use of the past and future relevant information of the degradation state of mechanical equipment and effectively transfer between layers, thereby improving the accuracy of remaining life prediction;

4. The natural attenuation learning efficiency can efficiently train the deep neural network, which greatly saves time and effectively prevents the network from not converging;

5. The invention can accurately predict the remaining life of mechanical equipment, and is simple and easy to implement. It can be widely used in the health assessment and remaining life prediction of rotating machinery in the fields of chemical industry, metallurgy, electric power, aviation and other fields. It can greatly reduce expensive unplanned maintenance and avoid The occurrence of a catastrophe.

Description of drawings

FIG. 1 is a flow chart of the method for predicting the remaining life of a rotating machine based on a multi-layer bidirectional gated cyclic unit network of the present invention.

Figure 2 is a schematic diagram of 6308 bearing installation;

Figure 3 is a bearing spalling diagram;

Figure 4 is a time-domain waveform diagram of the bearing vibration signal;

Figure 5 is a constructed bearing health index map;

Fig. 6 is an effect diagram of remaining life prediction by the method of the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific preferred embodiments.

As shown in FIG. 1 , the method for predicting the remaining life of a rotating machine with a multi-layer bidirectional gated cyclic unit network in this embodiment includes the following steps.

Step 1, vibration signal collection: collect the vibration signals of the key parts of the rotating machinery.

The key components of rotating machinery include bearings, gears or rotors, etc.; the acquisition method of vibration signals of key components is an existing technology. In this application, taking bearings as an example, ABLT-1A bearing life enhancement test bench. The test bench consists of test heads, test It consists of headstock, transmission system, loading system, lubrication system, electrical control system, testing and data acquisition system.

The test bench can install 4 bearings at the same time for accelerated fatigue life test.

The original test system of the test bench is composed of 4 thermocouples and 1 acceleration sensor, which respectively pick up the temperature signals of the 4 bearing outer rings and the vibration signal of the entire test bench. In order to effectively monitor the running status of each bearing, the test system has been adjusted and increased to 4 acceleration sensors to pick up the vibration signals on the 3 rigid shells respectively. The vibration sensors of channels 1, 2, 3, and 4 correspond to the bearing data of stations 1, 2, 3, and 4, respectively. The loading conditions of the test are shown in Table 1. The rated dynamic load of the bearing is 42.3kN, and the actual added weight is 35kg, that is, the rated dynamic load on each bearing is 17.5kN. The radial load loading conditions are shown in Table 2. After the bearing was fully loaded and operated for 16 hours, the final test machine was shut down because the root mean square vibration reached the shutdown threshold. The vibration rms at full load was 11.0 and the shutdown threshold was set at 45.0. After cutting the bearing with wire cutting, it can be seen that the outer ring of the bearing at the third station has obvious peeling phenomenon, as shown in Figure 3.

Table 1 Test conditions of the whole life test

Table 2 Radial load loading conditions

Step 2, build health indicators, as shown in Figure 5.

Step 21, the time-domain waveform of the 6308 bearing vibration signal is shown in Figure 4. The original features extracted from the vibration signal include 16 time-domain features, 13 frequency-domain features, 17 time-frequency domain features and 2 features based on trigonometric functions , the trigonometric functions are characterized by the inverse trigonometric hyperbolic cosine standard and the inverse trigonometric hyperbolic sine standard deviation, respectively.

Step 22, selecting features with a trend value greater than 0.8 to form a feature subset.

Step 23, using the unsupervised SOM algorithm to construct a health index for evaluating the life of the bearing from the selected feature subset, as shown in Figure 5.

Step 3, construct the network training set and test set, the sliding window length is set to 30, and the data after 1500 points is the test set.

In step 4, the MBi-GRU neural network is constructed by stacking a single bidirectional GRU network, the number of layers is 3, and the number of network neurons is 300.

Step 5, the state prediction value of the system obtained through the Bootstrap method.

Step 6, training network: the training of the whole model minimizes the loss function to train by error reverse transmission, and the present invention proposes to train the deep convolutional neural network with polynomial attenuation learning efficiency; set the learning efficiency according to formula 7, the initial learning efficiency is 0.01, the decay rate of learning efficiency is 0.5, S _decay is 40 steps, and S _global is 200 steps.

Step 7. Acquisition of the mean value and confidence interval of the remaining life: input the data of the test set into the network trained in step (5), the corresponding output of the test set is the predicted bearing degradation state value, through the formula (8) ~ (10 ) can calculate the predicted value of the remaining life of the bearing to be 81min, the confidence interval is [76.0,85.3], and the error percentage with the true value is 5.88%.

z(t _i+L )=φ(z(t _i ),z(t _i+1 ),...,z(t _i+L-1 )) (8)

Z _k (l _i +t _i )＝φ _k (S _k (t ₁ :t _i )) (9)

l _i ＝inf{l _i :Z _k (l _i +t _i )≥τz(t ₁ :t _i )} (10)

Step 8, evaluate the predicted remaining life, and compare with the prediction results based on GRU network, LSTM network, and fully connected neural network NN:

The remaining life value predicted by the method announced by the invention is smaller than MSE and MAPE of the other three methods, and can more accurately predict the remaining life of the bearing. Confidence intervals for remaining life were obtained by Bootstrap methods. The learning efficiency of natural decay can effectively train the network, and then learn the features in the vibration signal.

In short, the invention combines the advantages of deep learning's powerful feature extraction capabilities, uses a bidirectional gated recurrent unit neural network for regression prediction, and obtains the confidence interval of the remaining life through the Bootstrap method. Aiming at the problem that the model accuracy of the recurrent neural network model is sensitive to the value of the learning rate during the training process, too high or too low will affect the predictive performance of the model, the natural exponential decay learning efficiency is used to efficiently train the neural network.

This embodiment can accurately predict the remaining life of rotating machinery and the confidence interval, and can be widely used in the prediction of remaining life of rotating machinery in the fields of chemical industry, metallurgy, electric power, aviation, etc., can greatly reduce expensive unplanned maintenance, and avoid the occurrence of catastrophes.

The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be carried out to the technical solutions of the present invention. These equivalent transformations All belong to the protection scope of the present invention.

Claims (6)

1. The method for predicting the remaining life of rotating machinery of multi-layer bidirectional gated cyclic unit network, characterized in that: comprising the following steps: Step 1, collecting vibration signals: collecting vibration signals of key components of rotating machinery; Step 2, construction of health indicators; construct health indicators describing the degradation process of key components of rotating machinery; Step 3, constructing a network training set: For the vibration signals collected in step 1, construct a network training set, use the degradation state of key components of the rotating machinery as the input of the bidirectional gated recurrent unit neural network, and use sliding window technology to construct training and testing data sets ; Step 4, constructing the MBi-GRU (multi-layer bidirectional gated recurrent unit, MBi-GRU) neural network: constructing the MBi-GRU neural network from a single bidirectional GRU network; Step 5, constructing an integrated learning network through the Bootstrap method to obtain the uncertainty expression of the prediction result; Step 6, train the network: train by minimizing the loss function through error backpropagation, and train the MBi-GRU neural network with natural attenuation learning efficiency; Step 7. Acquisition of the mean value and confidence interval of the remaining life: input the data of the test set into the network trained in step (2), and the corresponding output of the test set is the predicted value of the remaining life to be obtained; Step 8, Evaluate the prediction result of the remaining life: use the mean square error and the absolute prediction error to evaluate the prediction result of the remaining life. 2. The method for predicting the remaining life of rotating machinery of multi-layer bidirectional gated cyclic unit network according to claim 1, characterized in that: in the step 2, the original features are extracted from the vibration signal, and the features with a large trend value are selected to form the features Finally, the unsupervised SOM algorithm is used to construct the health index of the selected feature subset to evaluate the life of the bearing. 3. The method for predicting the remaining life of rotating machinery with a multi-layer bidirectional gated cyclic unit network according to claim 1, characterized in that: in step 4, the update gate simultaneously controls the current input data x _t and the previous memory information h _t-1 , output a value z _t between 0 and 1, the calculation formula is z _t ＝σ(W _z [h _t-1 ,x _t ]+b _x ) (1) In the formula, x is the input data, h is the output of the GRU unit, r is the reset gate, z is the update gate, r and z jointly control how to calculate the new hidden state h _t from the previous hidden state h _t-1 ; z _t decides how much to transfer h _t-1 to the next state, which can be obtained from formula (1); In the formula, σ is the sigmoid function, W _z is the update gate weight, and b _z is the bias. The reset gate controls how important h _t-1 is to the outcome h _t ; if the previous memory h _t-1 is completely irrelevant to the new memory, the reset gate can function to remove the influence of the previous memory, i.e. r _t ＝σ(W _r [h _t-1 ,x _t ]+b _r ) (2) Generate new memory information according to the update gate

which is

The output at the current moment is h _t , namely

The basic unit of the Bi-GRU model consists of a forward-propagating GRU unit and a backward-propagating GRU unit; in a unidirectional neural network structure, the state is always output from front to back; However, in the remaining life prediction, if the output at the current moment can be related to the state at the previous moment and the state at the next moment; the current hidden layer state of Bi-GRU is from the current input x _t , t-1 time forward hidden state of

and the output of the reversed hidden layer state

The three parts are jointly determined.

Among them: the GRU() function represents the nonlinear transformation of the input mechanical equipment degradation index, and encodes the degradation index into the corresponding GRU hidden layer state; w _t and v _t respectively represent the forward hidden layer state corresponding to the two-way GRU at time t

and the output of the reversed hidden layer state

The corresponding weight, b _t represents the bias corresponding to the state of the hidden layer at time t. 4. The method for predicting the remaining life of rotating machinery of multi-layer bidirectional gated cyclic unit network according to claim 1, characterized in that: the health index in step 5 has N values, i.e. (z(t ₁ ), z(t ₂ ),..., z(t _N )), n=1, 2,..., N; let L be the length of the sliding window. (z(t _i ),z(t _i+1 ),...,z(t _i+L-1 )) a degenerated state with a window length, the corresponding output is z(t _i+L ), N A degraded state data can generate NL samples, and the predicted system state can be expressed by the following functional relationship z(t _i+L )=φ(z(t _i ),z(t _i+1 ),...,z(t _i+L-1 )) (8) Obtaining the confidence interval of the prediction result is to quantify the uncertainty of the point estimate; use an alternative method to resample K times from the original training data; the model φ _k (k=1,2,...,K) uses re The sampled data S _k (t ₁ :t _i )(k=1,2,...,K) training; finally, the integrated operation of multiple models will produce the mean and variance of the remaining life prediction results; the above description uses the formula expressed as follows Z _k (l _i +t _i )＝φ _k (S _k (t ₁ :t _i )) (9) Among them, Z _k (l _i +t _i ) represents the state prediction value of the system obtained by the Bootstrap method; the remaining life l _i at time t _i is defined as follows: l _i ＝inf{l _i :Z _k (l _i +t _i )≥τ|z(t ₁ :t _i )} (10) Among them, τ is the preset failure threshold; z _0:i is the estimated system state value from t ₀ to t _i time, Z _k (t _i + l _i ) is the estimated system state at t _i + l _i time value; the confidence interval of the final remaining life can be obtained by calculating the percentile of the remaining life l _i at time t _i . 5. the method for predicting the remaining life of rotating machinery of multi-layer bidirectional gated recurrent unit network according to claim 1, is characterized in that: utilize natural exponential decay learning efficiency to train deep recurrent neural network in described step 6, can effectively train Neural network, natural exponential decay learning efficiency calculation method is as follows:

Among them, lr: the current learning rate, lr ₀ : the initial learning rate, r _decay : the decay rate of each round of learning, 0<r _decay <1, S _global : the current number of global learning steps, S _decay : each round of learning The number of steps, S _decay =N _sample /N _batch , that is, the total number of samples divided by the size of each batch. 6. The method for predicting the remaining life of rotating machinery of a multi-layer bidirectional gated cyclic unit network according to claim 1, characterized in that: in step 8, the calculation formulas of mean square error and predicted absolute error are as follows:

Among them, HI _Act is the real bearing degradation state, HI _Pre is the predicted bearing degradation state, and N _p is the number of predicted points.

Images