CN112101659A

CN112101659A - Complex equipment service life prediction method based on stacking denoising autoencoder

Info

Publication number: CN112101659A
Application number: CN202010965144.6A
Authority: CN
Inventors: 江水; 李军; 陈意; 都竞; 梁天; 胡静; 周靳
Original assignee: Anhui Sanheyi Information Technology Co Ltd
Current assignee: Anhui Sanheyi Information Technology Co Ltd
Priority date: 2020-09-15
Filing date: 2020-09-15
Publication date: 2020-12-18

Abstract

The invention discloses a life prediction method of complex equipment based on a stack denoising autoencoder, which comprises the following steps: firstly, signal acquisition is carried out, historical degradation data of complex equipment acquired based on a sensor is acquired, then self-adaptive feature extraction can be directly carried out on an original monitoring signal based on a stacked denoising self-encoder, finally, feature indexes are intelligently selected to serve as health factor values of the complex equipment, and a life prediction result is output. The method well solves the problem that the prior knowledge is insufficient, and therefore, the category is difficult to label manually or the cost for labeling manually is too high; meanwhile, the denoising self-encoder has the denoising capability, and can effectively improve the prediction precision in the face of noise mixed in monitoring data under a complex working environment. The method does not need to integrate a plurality of feature extraction or dimension reduction methods according to different conditions, has better universality, can reduce the influence of local noise, can avoid resource consumption caused by manually designing training labels in supervised learning, and can intuitively and accurately predict the service life.

Description

Complex equipment service life prediction method based on stacking denoising autoencoder

Technical Field

The invention relates to the technical field of mechanical equipment monitoring, in particular to a life prediction method of complex equipment based on a stacking denoising autoencoder, which is used for evaluating the residual usable life of the complex equipment.

Technical Field

With the trend of informatization, enlargement and complication of mechanical equipment in recent years, demands for sophisticated equipment capability and efficiency are increasing. In practical engineering application, mechanical equipment is inevitably damaged facing to complex load action, severe operation conditions and frequent start-stop working conditions, once a fault occurs, production can be interrupted, efficiency is reduced, huge economic loss can be caused, even normal operation of the whole system is affected, and therefore service life prediction of complex equipment is necessary.

In the traditional mechanical equipment health management, two modes of after-the-fact maintenance and timing maintenance are mainly adopted. After maintenance, namely, the equipment is maintained after being damaged, however, the equipment stops running in partial industrial fields and serious aftereffects are encountered, so that the mode is gradually eliminated; the regular maintenance is to make a maintenance plan of the equipment in advance, and check and maintain each mechanical part on schedule, which needs to consume a large amount of manpower and material resources.

At present, due to the rapid development of the intelligent manufacturing industry and the sensor technology, data generated by equipment is well collected, and mechanical big data also brings a new method for life prediction, namely a life prediction method based on data driving, which becomes a mainstream method in the field nowadays, wherein the machine learning algorithm is most widely applied. The traditional machine learning algorithm is often a shallow model, such as a regression series model, a support vector machine model, a filter model and the like, and has the problems of more manual participation, difficulty in representing a complex functional relation, to-be-improved prediction precision, poor noise reduction capability and the like during prediction;

in summary:

1. the traditional health management of mechanical equipment (mainly including after-maintenance and regular maintenance) is gradually eliminated;

2. the shallow model in the data-driven life prediction method has the problems of to-be-improved prediction precision and poor noise reduction capability;

3. the method comprises the following steps that (1) original sensor information is often subjected to optimization selection by means of a complex sensor information evaluation criterion, and degradation features are extracted and subjected to feature selection;

4. most of data-driven life prediction methods are supervised learning, namely, a real output value corresponding to input needs to be provided as a label in the training process, and the selection of the label needs to depend on the participation of expert experience, so that time is consumed, and no consistent standard exists;

5. most models are difficult to represent complex functional relationships, multiple signal processing methods are often adopted for fusion aiming at specific problems, parameters are often selected depending on manual experience, and the universality and the generalization capability of each device are poor; therefore, a life prediction method of complex equipment based on a stack denoising autoencoder is provided.

Aiming at the problems, the method adopts a Stacked denoising auto-encoder (SDAE) algorithm, is more suitable for processing one-dimensional input data due to the structural characteristics and has denoising capability on input signals; in addition, the unsupervised learning model of the stacking denoising autoencoder can perform self-adaptive feature extraction on an original monitoring signal, resource consumption caused by a manual design training label in supervised learning can be avoided, a deep layer structure of the stacking denoising autoencoder can directly extract a single node value to serve as an engine health factor value, various feature extraction or dimension reduction methods do not need to be fused according to different situations, and the stacking denoising autoencoder has good universality. Therefore, the SDAE model is applied to the mechanical life prediction of complex equipment, a large number of signal processing technologies and experts in the field are not needed, and the universality is high.

Disclosure of Invention

The invention mainly aims to provide a method for predicting the service life of complex equipment based on a stacking denoising autoencoder, which can effectively solve the problems in the background technology.

In order to achieve the purpose, the invention adopts the technical scheme that: a life prediction method of complex equipment based on a stacking denoising autoencoder comprises the following steps:

step S1, signal acquisition: acquiring historical degradation data of the complex equipment based on sensor acquisition;

step S2, data preprocessing: selecting all sensor data except the monitoring data which are unchanged as input data of the stacking denoising self-encoder, normalizing the original data, and dividing a training set and a test set of the normalized data;

step S3, model building training: the built model comprises a coding network formed by a self-coder and a decoding network formed by the self-coder, and the coding network outputs the extracted health factors; carrying out unsupervised pre-training of denoising self-encoders one by one in an encoding network on a training set, then directly taking the weight of the network as the weight transpose of relative position pre-training in a decoding network, and finally carrying out parameter fine tuning by using a BP algorithm to finish model training;

step S4, verification and evaluation: inputting test data into a trained stacked denoising self-encoder model, performing self-adaptive feature extraction through a plurality of hidden layers to obtain corresponding health factor values, and constructing a HI (health index) curve of the test data to predict the residual life.

Further, the normalization formula in the data preprocessing in step S2 is:

x^*＝(x-x_min)/(x_max-x_min)

wherein x is the original data, x_maxAnd x_minRespectively, a maximum and minimum value for each engine cycle for each sensor.

Further, after normalization, the raw data was divided into training data and test data at 70% and 30%.

Further, in the step S3, the model is set up and trained, the denoising autoencoder encodes the data with noise added through an encoding process, then decodes the encoded result through a decoding process, the difference between the generated reconstructed data and the original input is used as a reconstruction error, and the gradient descent algorithm is used for training, so that the denoising autoencoder can capture the key dependence factor among the input data by adding the noise factor of the algorithm, and learn the characteristics of more robustness.

Furthermore, the algorithm network structure of the built model comprises an input layer, two hidden layers of the coding network, an output layer of the coding network, two hidden layers of the decoding network and an output layer of the decoding network.

Assuming that input data of a first denoising self-encoder of a stacking denoising self-encoder network is x, destroying the input data through a random mapping function to obtain data added with noise

Through f_θ1Generating an output h of a hidden layer₁The formula is as follows:

wherein the content of the first and second substances,

for noisy data, W₁And b₁The weight and the bias of the coding of the first denoising autoencoder are calculated, s is a sigmoid activation function, theta 1 is a coding parameter pre-trained by the first denoising autoencoder, and theta 1 is [ W ═₁，b₁]。

Further, the output h of the hidden layer₁Then go through decoding process g_θ1Generating reconstruction data z, the formula is as follows:

z＝g_θ1′(h₁)＝s(W′₁h₁+b′₁)

wherein, w'₁And b₁Is the decoded weight and bias of the first denoising autoencoder, s is sigmoid activating function, theta 1 is the pre-trained decoding parameter of the first denoising autoencoder,

θ₁′＝[w′₁，b′₁]。

reconstruction error L for encoder_HCalculating, and minimizing an objective function by using a gradient descent algorithm to improve the robustness of the features learned by the denoising autoencoder from the input data x, wherein the formula is as follows:

L_H(x，z)＝||x-z||²

wherein | · | purple sweet²Denotes a 2 norm, n is the number of samples, x⁽ⁱ⁾Is the ith sample data of the first image,

the data is the ith sample data added with noise, i is 1, 2, …, n.

The working principle of the rest denoising autocoders is the same as that of the first denoising autocoder.

Further, the step of unsupervised pre-training for denoising the self-encoder one by one in the encoding network on the training set comprises:

(1) randomly initializing the network parameters of the stacking denoising self-encoder;

(2) and the training set data sequentially passes through a first denoising autoencoder, a second denoising autoencoder and a third denoising autoencoder in the coding network.

Specifically, the output h of the hidden layer is reserved after the training of the first denoising self-encoder is finished₁And the output of the hidden layer is used as the input of a second denoising self-encoder, the output of the hidden layer is reserved after the pre-training of the second denoising self-encoder is finished, and the output h of the hidden layer is used₂And the third denoised self-encoder is used as the input of the third denoised self-encoder to finish the unsupervised pre-training of the training set data in the encoding network. And after the training is finished, setting the weight of the decoding network as the transposition of the corresponding weight of the coding network.

Further, after training is completed, original data of a training set is used as a label, and a BP algorithm is used for fine tuning of network parameters of the stacking denoising autoencoder, so that the nonlinear function mapping capability of a multilayer network can be further improved, and better characteristic expression can be obtained, wherein the formula is as follows:

wherein x is^mIs the raw data of the training set, y^mIf so, the stacked denoised is output from the hidden layer of the encoder network. Theta ═ theta₁,θ₂,…,θ₆The updating mode of the parameters is

Wherein alpha is the learning rate in the parameter fine adjustment process, and the fine-adjusted network parameters are used for extracting the multidimensional characteristics output by the coding network.

Further, in the test set verification in step S4, for the trained model, the test set is input into the model, adaptive feature extraction is performed through multiple hidden layers to obtain corresponding health factor values, a hi (health index) curve of the test data is constructed, and smoothing filtering is performed for 15 times, so that local noise can be reduced, and the remaining service life of the complex equipment can be predicted more accurately.

Compared with the prior art, the method for predicting the service life of the complex equipment based on the stacking denoising autoencoder has the following beneficial effects:

1. the SDAE has the noise reduction capability, and can effectively improve the prediction precision in the case of noise mixed in monitoring data under a complex working environment;

2. SDAE is unsupervised learning, can directly carry out self-adaptive feature extraction on an original monitoring signal, and better solves the problem that the prior knowledge is insufficient, and therefore, the category is difficult to label manually or the cost for labeling the category manually is too high;

3. strong generalization ability: a single node value can be directly extracted to serve as an engine health factor value, and a plurality of feature extraction or dimension reduction methods do not need to be fused according to different conditions, so that the universality is good;

4. carrying out smoothing filtering processing on the output health factor for 15 times to reduce local noise influence;

5. beneficial information contained in each sensor does not need to be evaluated, feature extraction is not needed, so that a feature set is obtained, more advanced characterization is carried out on original data, and resource consumption caused by artificial design training labels in supervised learning can be avoided.

Drawings

FIG. 1 is a flowchart of a life prediction method for complex equipment based on a stacked denoising autoencoder according to the present invention.

FIG. 2 is a structural diagram of a denoising autoencoder in the life prediction method of complex equipment based on a stacked denoising autoencoder of the present invention;

FIG. 3 is a structural diagram of a stacked denoising autoencoder in the method for predicting the service life of complex equipment based on the stacked denoising autoencoder of the present invention

FIG. 4 is a schematic diagram illustrating training of a stacked denoising autoencoder in the method for predicting the life of complex equipment based on the stacked denoising autoencoder.

Detailed Description

In order to make the technical means, the creation features, the achievement purposes and the effects of the invention easy to understand, the following detailed description is made in conjunction with the accompanying drawings to further explain the invention.

The invention provides a life prediction method of complex equipment based on a stacking denoising autoencoder, which can be used for maintaining the complex equipment in time and reducing unnecessary maintenance by providing an accurate and reliable residual life prediction method of the complex equipment, so that the maintenance of the complex equipment is targeted, and meanwhile, the residual life of the complex equipment is accurately predicted, so that equipment accidents are reduced, and a data stream is shown in figure 1.

In the specific embodiment, based on CMAPSS, which is an aircraft engine simulation state monitoring public data set of NASA (science Center of Excellence, PCoE), the implementation steps are as follows:

step S1, signal acquisition: acquiring historical degradation data acquired by a plurality of sensors of the turbine engine, wherein specific information is shown in the following table 1;

serial number	Symbol	Description of the invention
			1	T2	Total temperature of fan inlet
2	T24	Low pressure compressor temperature
			3	T30	High pressure compressor temperature
4	T50	Low pressure turbine temperature
			5	P2	Fan inlet pressure
6	P15	Bypass line pressure
			7	P30	Air pressure of high pressure compressor
8	Nf	Physical fan speed
			9	Nc	Physical core speed of rotation
10	Epr	Engine pressure ratio
			11	Ps30	Static pressure of high pressure compressor
12	Phi	Fuel flow to high pressure compressor static pressure ratio
			13	NRf	Correcting fan speed
14	NRc	Correcting core rotational speed
			15	BPR	Bypass ratio
16	farB	Fuel to air ratio in combustion chamber
			17	htBleed	Heat content of exhaust valve
18	Nf_dmd	Demand fan speed
			19	PCNfR_dmd	Corrected demanded fan speed
20	W31	High pressure turbine coolant discharge
			21	W32	Low pressure turbine coolant discharge

Step S2, data preprocessing: directly observing original monitoring information of a plurality of sensors, and selecting all sensor data except monitoring data which are unchanged as input data of the stacking denoising self-encoder; and carrying out data normalization on the original data to be in a [0, 1] interval, dividing the original data into training data and test data by 70% and 30%, wherein the training set and the test set respectively comprise 100 monitoring units, and the time sequence length of each unit is different. The training and test sets included 20631 and 13096 engine cycles, respectively. All cells start to degrade from slightly lossy states of varying degrees. Only 14-dimensional information among the 21-dimensional sensor information has different degrees of tendencies (#2, #3, #4, #7, #8, #9, #11, #12, #13, #14, #15, #17, #20, #21), and the remaining 7-dimensional monitoring information is unchanged and has no useful information.

Step S3, model building training: the built model comprises a coding network formed by a self-coder and a decoding network formed by the self-coder, and the coding network outputs the extracted health factors; the method comprises the steps of firstly carrying out unsupervised pre-training of denoising self-encoders one by one in an encoding network on a training set, then directly taking the weight of the network as the weight transpose of relative position pre-training in a decoding network, finally carrying out parameter fine tuning by using a BP algorithm, completing model training, wherein the structure of the denoising self-encoder is shown in figure 1.

Preferably, the built model comprises 4 denoising automatic coding machines in total, wherein the first 2 self-coders form a coding network, the second 2 self-coders form a decoding network, the coding network outputs the extracted health factors, and the structure of the stacking denoising automatic coding machine is shown in fig. 2.

Preferably, unsupervised pre-training of denoising from an encoder one by one in an encoding network is performed on a training set, then weights of the network are directly taken as weight transpositions which are pre-trained in relative positions in a decoding network, finally parameter fine tuning is performed by using a BP algorithm, model training is completed, and the training process is as shown in FIG. 3.

Preferably, the structure of the noise automatic coding machine adopted by the invention is 14-7-1-7-14, wherein 14 is a 14-dimensional sensor value of engine data, 14-7-14 and 7-1-7 respectively form 2 noise automatic coding machines corresponding to an input layer to an output layer of a coding network in the figure, and the characteristic number finally extracted by the coding network is 1 and directly used as an engine health factor value at a single time point. The pre-training times of a single denoising autoencoder are 20, the noise rate is 0.05, the learning rate is 1, and the reverse tuning times of the global network are 20.

And step S4, verifying and evaluating. After the model training is finished, inputting the normalized test data into a stacking denoising self-encoder for self-adaptive feature extraction, and taking a single feature extracted from each sampling point in the engine degradation process as an HI value, thereby obtaining a finally constructed HI curve. In order to make the HI curve smoother and reduce the influence of local noise, the obtained HI curve is subjected to sliding filtering processing with a window size of 15.

Preferably, in order to better evaluate the method, relevant parameters are set, and analysis and comparison are performed by using methods such as a Deep Belief Network (DBN), an Autoencoder (AE), a Convolutional Neural Network (CNN), a BP neural network (BPNN), a Support Vector Machine (SVM), a Random Forest (RF), and the like.

Preferably, analysis of experimental accuracy results shows that due to shallow models such as BPNN, SVM and RF, performance depends on artificial feature extraction, influence of subjective factors is large, generalization capability is poor, and accuracy is 83.8%, 82.95% and 87.67% respectively; the DBN, AE, and CNN have stronger nonlinear mapping capability than the shallow model, but are susceptible to noise and have not high enough prediction accuracy, which is: 91.2%, 93.98%, 92.78%; according to the method, a deeper model is established, the good characteristic learning capability of the deep model is fully utilized, the deep characteristics are extracted to be used as the health of complex equipment, so that the service life is predicted, and the prediction accuracy is 99.08%. The method is proved to be capable of accurately predicting the residual service life of the complex equipment.

According to the embodiment, the SDAE has the noise reduction capability, and can effectively improve the prediction precision in the case of noise mixed in monitoring data in a complex working environment; the SDAE is unsupervised learning, can directly carry out self-adaptive feature extraction on the original monitoring signal, and better solves the problem that the prior knowledge is insufficient, and therefore, the category is difficult to label manually or the cost for labeling the category manually is too high; the generalization capability is strong, a single node value can be directly extracted to serve as an engine health factor value, multiple feature extraction or dimension reduction methods do not need to be fused according to different conditions, and the universality is good; carrying out smoothing filtering processing on the output health factor for 15 times to reduce local noise influence; beneficial information contained in each sensor does not need to be evaluated, feature extraction is not needed, so that a feature set is obtained, more advanced characterization is carried out on original data, and resource consumption caused by a manual design training label in supervised learning can be avoided.

The foregoing shows and describes the general principles and broad features of the present invention and advantages thereof. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are merely illustrative of the principles of the present invention, but that various changes and modifications may be made without departing from the spirit and scope of the invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims

1. The invention mainly aims to provide a life prediction method of complex equipment based on a stacking denoising autoencoder, which comprises the following steps:

step S3, model building training: the built model comprises a self-coding mechanism coding network and a self-coding mechanism decoding network, and the coding network outputs the extracted health factors; carrying out unsupervised pre-training of denoising self-encoders one by one in an encoding network on a training set, then directly taking the weight of the network as the weight transpose of relative position pre-training in a decoding network, and finally carrying out parameter fine tuning by using a BP algorithm to finish model training;

2. The method of claim 1, wherein the method comprises: the normalization formula in the data preprocessing is as follows:

x^*＝(x-x_min)/(x_max-x_min)

3. The method of claim 2, wherein the method comprises: after normalization, the raw data were divided into training data and test data at 70% and 30%.

4. The method of claim 1, wherein the method comprises: in the model building training, the denoising self-coding opportunity encodes the data added with noise through an encoding process, then decodes an encoding result through a decoding process, the difference between the generated reconstruction data and the original input is used as a reconstruction error, and a gradient descent algorithm is used for training.

5. A built model structure according to claim 4, comprising an input layer, two hidden layers of the coding network, an output layer of the coding network, two hidden layers of the decoding network and an output layer of the decoding network.

6. The method of claim 1, wherein the method comprises: the unsupervised pre-training step of carrying out one-by-one denoising self-encoder in an encoding network on a training set comprises the following steps:

(2) the training set data sequentially passes through a first denoising self-coding machine, a second denoising self-coding machine and a third denoising self-coding machine in the coding network.

7. The method of claim 6, wherein the method comprises: and after the training of the first denoising self-coding machine is finished, the output of the hidden layer is reserved, the output of the hidden layer is used as the input of a second denoising self-coding machine, the output of the hidden layer is reserved after the pre-training of the second denoising self-coding machine is finished, the output of the hidden layer is used as the input of a third denoising self-coding machine, and the weight of the decoding network is set as the transposition of the corresponding weight of the coding network after the training is finished.

8. The method of claim 7, wherein the method for predicting the life of the complex equipment based on the stacked denoising self-encoder comprises: after training is finished, original data of a training set is used as a label, and the BP algorithm is utilized to finely adjust network parameters of the stacking denoising autoencoder, so that the nonlinear function mapping capacity of the multilayer network can be further improved, and better characteristic expression can be obtained.

9. The method of claim 1, wherein the method comprises: and performing smooth filtering treatment on all constructed HI curves for 15 times, so that local noise can be reduced, and the residual service life of the complex equipment can be predicted more accurately.