CN114791537A - Method and system for predicting residual service life of industrial equipment based on attribute self-adaption - Google Patents

Abstract

The invention discloses an attribute self-adaptive based method and system for predicting the remaining service life of industrial equipment, belonging to the field of fault diagnosis of digital and intelligent manufacturing systems, and comprising the steps of firstly, acquiring power utilization data of the industrial equipment through a sensor; then removing redundant attributes according to the three-phase circuit knowledge; then, distributing higher weight for the important attribute through an attribute self-adaptive network; and finally, sending the weighted time sequence data to a long-term and short-term memory regression network for processing, and predicting the residual service life through a full-connection network.

Description

Method and system for predicting residual service life of industrial equipment based on attribute self-adaption

Technical Field

The invention belongs to the field of fault diagnosis of a digital and intelligent manufacturing system, and particularly relates to a method and a system for predicting Remaining service Life (RUL) of industrial equipment based on attribute self-adaption.

Background

Predictive Maintenance (PdM) is an equipment Maintenance strategy based on industrial big data drive, and because it can maintain bottleneck equipment in advance before a fault occurs by predicting the health condition of production equipment, and prolong the service life of the equipment, it is widely used in the fields of medicine production, equipment manufacturing, industrial control, etc. Common predictive maintenance methods are largely divided into statistical-based predictive maintenance and artificial intelligence-based predictive maintenance.

According to the mode of acquiring the state Monitoring Data (CMD), the predictive maintenance method based on statistics can be further divided into a Direct CMD method and an index CMD method [1 ]. The Direct CM method comprises a Bayesian method [2], a regression model [3], a gamma process [4] and a Markov process; the Indirect CM method includes stochastic filtering, covariant models and hidden Markov models [5 ]. While these methods have achieved superior results, they have shortcomings, such as the difficulty of mathematically building accurate models for complex production systems; furthermore, statistical-based predictive maintenance methods require knowledge in both the mathematical and mechanical domains of expertise. Therefore, it is an urgent need in the industry to be able to simplify the "end-to-end" model of a production system, which is exactly what artificial intelligence based predictive maintenance methods can meet.

As artificial intelligence evolves, more and more machine learning algorithms are applied to the field of predictive maintenance. These algorithms are mainly classified into four categories, namely Support Vector Machines (SVMs), K-Means clustering algorithms, Ensemble Learning (EL), and Neural networks (ANN) [6 ]. SVM is a classical supervised machine learning algorithm, however it lacks timing dependency considerations in the manufacturing system and is difficult to model optimize under complex production conditions. The K-Means clustering algorithm is a classic unsupervised machine learning algorithm, but it is often difficult to determine the number of classes in practical applications, and the clustering algorithm is susceptible to sample order. EL achieves excellent working performance by generating a plurality of weak classifiers and then aggregating them together, and a common method is Random Forest (RFs), but it has disadvantages of difficult realization and long operation time. ANN is a machine learning algorithm developed by the inspiration of biological neurons, and has been widely applied in the field of predictive maintenance, such as: an abnormity detection system of an industrial engine is developed by using a Long Short-Term Memory Network (LSTM) by Aydin and Guldamlasiglu [7 ]; the Lu team uses a Gated Recurrent Unit (GRU) to predict the remaining service life [8 ]; the Souza team applied a Convolutional Neural Network (CNN) to the failure prediction system of spinning devices [9 ]; the Wang group uses a Time Convolutional Network (TCN) for fault detection of a machine manufacturing facility [10 ]. However, these methods have two problems: first, these methods lack the study of feature extraction, either by simply screening features empirically or by coarsely feeding all attributes into the model, which can degrade the performance of the algorithm; on the other hand, each attribute contributes differently to the final prediction result, and therefore a mechanism that can automatically adjust the weight of the attribute is required.

Disclosure of Invention

In order to solve the technical problems, the invention provides an attribute-adaptive-based method and system for predicting the remaining service life of industrial equipment, aiming at bottleneck equipment of an industrial control system, based on the Internet of things and artificial intelligence. The method comprises the steps of firstly, acquiring power consumption data of industrial equipment through a sensor; then removing redundant attributes according to the three-phase circuit knowledge; then, distributing higher weight for the important attribute through an attribute self-adaptive network; and finally, sending the weighted time sequence data to a long-term and short-term memory regression network for processing, and predicting the remaining service life through a full-connection network.

The technical scheme of the invention is as follows:

an attribute-adaptive-based method for predicting the remaining service life of industrial equipment comprises the following steps:

a data acquisition step: acquiring time sequence data in the processing process through an intelligent sensor fixed on industrial equipment; the time sequence data comprises 5 types of attributes of each moment, wherein the 5 types comprise a voltage type, a current type, a power factor type and a frequency type;

and (3) attribute fusion: removing redundant attributes in the 5 types of attributes according to the three-phase circuit knowledge, combining the rest attributes, and taking a time window with a preset size to form an attribute characteristic;

a weight distribution step: processing the attribute characteristics through an attribute adaptive network, and distributing weights for the combined residual attributes to obtain attribute adaptive characteristics;

a fault diagnosis step: and processing the attribute self-adaptive characteristics and outputting a result through the long-short term memory regression network, and processing the output result through the full-connection network to predict the residual service life of the industrial equipment.

Further, in the data acquisition step, a mutual inductor is respectively arranged on each phase of the three-phase circuit to convert high voltage into low voltage.

Furthermore, in the data acquisition step, a fuse and leakage protection equipment are additionally arranged on the intelligent sensor for acquiring data.

Furthermore, in the data acquisition step, the data acquisition method is Run-to-Fail, namely, the data acquisition of the current operation and maintenance period is stopped when a systematic machining fault is encountered; when the debugging is normal, starting to acquire data of the next operation and maintenance period; the acquisition frequency was 50 hz.

Further, the time series data includes 28 attributes of 5 types per time instant, namely:

voltage class: a phase voltage, B phase voltage, C phase voltage, AB line voltage, BC line voltage and CA line voltage;

current class: phase A current, phase B current and phase C current;

power class: active power of phase A, active power of phase B, active power of phase C, total active power, reactive power of phase A, reactive power of phase B, reactive power of phase C, total reactive power, apparent power of phase A, apparent power of phase B, apparent power of phase C, total apparent power;

power factor class: phase A power factor, phase B power factor, phase C power factor, total power factor;

frequency class: phase A frequency, phase B frequency, and phase C frequency.

Further, removing redundant attributes in the class 5 attributes includes:

removing redundant attributes in voltage classes: AB line voltage, BC line voltage, CA line voltage;

removing redundant attributes in the power class: active power of phase A, active power of phase B, active power of phase C, total active power, total reactive power, apparent power of phase A, apparent power of phase B, apparent power of phase C and total apparent power;

removing redundant attributes in the power factor class: the total power factor.

Further, the remaining attributes are 15: phase A voltage, phase B voltage, phase C voltage, phase A current, phase B current, phase C current, phase A reactive power, phase B reactive power, phase C reactive power, phase A power factor, phase B power factor, phase C power factor, phase A frequency, phase B frequency, and phase C frequency.

Further, in the step of assigning weights, the step of assigning weights by the attribute adaptive network includes:

a compression step: inputting the attribute features into a pooling layer to obtain intermediate features;

self-adapting: inputting the intermediate features into a fully-connected network to obtain a weight vector;

a weight distribution step: and inputting the attribute features and the weight vector, and distributing a new weight to each attribute feature to obtain the attribute adaptive feature.

Further, in the fault diagnosis step, the step of processing the attribute adaptivity characteristics through the long-short term memory regression network and outputting the result comprises the following steps:

inputting the attribute adaptive characteristic into a forgetting gate, and calculating to obtain a quantity f;

inputting the attribute adaptive characteristic into a memory gate, and calculating to obtain a quantity i;

calculating quantities c and h from quantities f and i;

and generating an output result according to the quantities c and h.

An attribute-based adaptive industrial equipment remaining useful life prediction system comprising:

the data acquisition module is used for acquiring time sequence data in the processing process through an intelligent sensor fixed on industrial equipment; the time sequence data comprises 5 types of attributes of each moment, wherein the 5 types comprise a voltage type, a current type, a power factor type and a frequency type;

the attribute fusion module is used for removing redundant attributes in the 5 types of attributes according to the three-phase circuit knowledge, combining the rest attributes and taking a time window with a preset size to form an attribute characteristic;

the neural network model is composed of an attribute self-adaptive network, a long-term and short-term memory regression network and a full-connection network, and is used for processing attribute characteristics through the attribute self-adaptive network and distributing weights for the combined residual attributes to obtain attribute self-adaptive characteristics; and processing the attribute adaptive characteristics and outputting a result through the long-term and short-term memory regression network, and then processing the output result through the full-connection network to predict the residual service life of the industrial equipment.

Further, the neural network model is optimized by using an Adam optimizer, and parameters in the neural network model are updated according to a mean square error loss function and an error back propagation algorithm of the residual service life until convergence.

The invention has the beneficial effects that:

the invention provides an attribute self-adaptive industrial equipment remaining service life prediction method, which can be used for automatically improving the calculation weight of important attributes by removing redundant features and an attribute self-adaptive network, and improving the evaluation indexes such as the accuracy, precision, recall rate, F1 value and the like of RUL prediction in a real production scene; the invention establishes a time sequence dependence and prediction regression model of the production system through the long-term and short-term memory regression network, and simultaneously considers the current state and the historical information of the equipment, thereby establishing a more accurate and reliable production system model. Thus, the present invention goes beyond existing machine learning-based predictive maintenance algorithms.

Drawings

Fig. 1 is a flow chart of fault diagnosis of an industrial device according to an embodiment of the present invention.

Fig. 2 is a schematic diagram of attribute adaptive network processing in an embodiment of the present invention.

FIG. 3 is a diagram illustrating an exemplary long term memory regression network process.

Detailed Description

The technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The embodiment provides an attribute-adaptive-based method for predicting the remaining service life of industrial equipment, and the input of the method is a sequence with a time window size of tau

Each time instant contains 28 attributes, namely a-phase voltage U _A B-phase voltage U _B C-phase voltage U _C AB line voltage U _AB BC line voltage U _BC CA line voltage U _CA Phase I of current A _A Phase I of current B _B Phase I of current C _C Active power P of phase A _A B phase active power P _B Active power P of C phase _C Total active power P, A phase reactive power Q _A B phase reactive power Q _B C-phase reactive power Q _C Total reactive power Q, A apparent power S _A Apparent power S of B-camera _B C apparent power S of the camera _C Total apparent power S, A phase power factor

B phase power factor

C-phase power factor

Total power factor

Frequency f of phase A _A Phase B frequency f _B And C phase frequency f _C I.e. by

The output is the current remaining service life of the industrial equipment

Fig. 1 shows a processing flow chart of the method, which specifically includes the steps of:

(1) and (4) data acquisition. In real plants, the wiring of industrial equipment often employs a "three-phase four-wire system". It is dangerous to directly collect the electricity data of the industrial equipment. Therefore, a mutual inductor is required to be respectively installed on each phase of the three-phase circuit to convert high-voltage electricity into low-voltage electricity and then perform data acquisition. For an intelligent sensor (measuring electric meter) for collecting data, a fuse protector and an electric leakage protection device are additionally arranged to ensure the safety of personnel and equipment.

(2) And fusing the attributes. The method is characterized in that redundant attributes in original data are removed by using three-phase circuit knowledge, and other attributes are subjected to feature fusion, and the specific method comprises the following steps:

from equations (1) through (3), it can be seen that the line voltage can be represented by the phase voltage, so when the property U is chosen _A 、U _B And U _C In time, U can be put _AB 、U _BC And U _CA Removed as a redundant attribute.

P＝P _A +P _B +P _C (7)

From equations (4) through (6), it can be seen that when I is continuously selected _A 、I _B 、I _C 、

Then, P can be substituted _A 、P _B 、P _C And P is removed as a redundant attribute.

Q＝Q _A +Q _B +Q _C (8)

S ² ＝P ² +Q ² (9)

From the equations (7) and (8), it can be seen that Q is continuously selected _A 、Q _B 、Q _C Then Q, S can be added _A 、S _B 、S _C And S is removed as a redundant attribute. In addition, the frequency f _A 、f _B And f _C Are related to power supply attributes and are therefore not redundant attributes.

In conclusion, by applying the three-phase circuit related knowledge, the redundant attribute in the original data can be removed, and only U is reserved _A 、U _B 、U _C 、I _A 、I _B 、I _C 、

Q _A 、Q _B 、Q _C 、f _A 、f _B 、f _C A total of 15 attributes constituting a time window ofAttribute characteristics of tau

(i.e., two-dimensional space vector, d attributes, τ time).

(3) And (4) weight distribution. As shown in fig. 2, the attribute-based adaptive network takes the attribute characteristics as input, and assigns a higher weight to the fused important attributes. The method comprises the following specific steps:

compressing, inputting the attribute feature F into the pooling layer, calculating the intermediate feature as formula (10)

Adaptive, inputting the intermediate features F' into a fully-connected network, e.g. formula (11) calculates the weight vector

A＝softmax(ω ₂ ·σ(ω ₁ ·F'+b ₁ )+b ₂ ) (11)

Wherein

The invention adopts a 3-layer fully-connected network as a network parameter, the dimensionalities of the layer 1 and the layer 3 are d, the dimensionality of the layer 2 is k which is d/alpha (alpha is a hyper parameter), and sigma (·) is a sigmoid activation function.

Distribution, namely performing bitwise operation on F and A according to a formula (12) to obtain attribute adaptive characteristics

Wherein

And is

Is a bit-wise multiplication operation, i.e.

(4) And (5) fault diagnosis. As shown in FIG. 3, the characteristic Q is input into the long-short term memory regression network, and then the remaining service life of the production equipment is predicted through the full-connection network

The method comprises the following specific steps:

enter forget gate, as formula (13):

f _t ＝σ(ω _f ·[h _t-1 ,Q _t ]+b _f ) (13)

wherein omega _f And b _f Is a forgetting gate parameter, h _t Is a hidden variable; σ is the sigmoid activation function.

Entering a memory gate, as in equation (14) and equation (15):

i _t ＝σ(ω _i ·[h _t-1 ,Q _t ]+b _i ) (14)

wherein omega _i 、b _i 、ω _c 、b _c Is to memorize the gate parameter, h _t-1 Is a hidden variable, tanh (-) is a tanh activation function; σ is the sigmoid activation function.

Entering the output gate, as in equation (16) and equation (17):

wherein ω is _o 、b _o Is the output gate parameter; output gate based on calculated c _t 、h _t To obtain the final output result LSTM (Q).

Entering a fully connected network, as in equation (18):

wherein ω is ₃ 、b ₃ 、ω ₄ 、b ₄ Are parameters of the layers of a fully connected network.

The embodiment provides an attribute-adaptive-based system for predicting remaining service life of industrial equipment, which includes:

the data acquisition module is used for realizing the data acquisition in the step (1);

the attribute fusion module is used for realizing the attribute fusion in the step (2);

and (4) a neural network model which consists of an attribute self-adaptive network, a long-short term memory regression network and a fully-connected network and is used for realizing the weight distribution of the step (3) and the fault diagnosis of the step (4).

The neural network model is optimized by using an Adam optimizer, and parameters in the neural network model are updated according to a mean square error loss function and an error back propagation algorithm of the remaining service life until convergence.

Experimental design and results

(1) The experimental environment is as follows:

the system environment is as follows: windows 10, Python 3.7, Keras 2.3, Tensorflow 1.13

Hardware environment: memory: 15GB, GPU: NVIDIA RTX 2080Ti, CPU 4.00GHz Intel (R) Xeon (R) W-2125, hard disk: 2 TB.

(2) Experimental data:

the experiment collects real data of the horizontal machining center in the manufacturing process to demonstrate the practical performance of the model. In the experiment, 10 operation and Maintenance Cycles (MC) are collected in a Run-to-Fail mode, namely the collection of the operation and Maintenance Cycle is stopped when a systematic processing fault is met, and the collection of the next operation and Maintenance Cycle is started after the system is debugged normally; the sampling frequency was 50 hz.

(3) The optimization mode is as follows:

the model is optimized using an Adam optimizer; selecting a Mean Square Error (MSE) loss function of the residual service life,

n is the value of the batch size,

is the true remaining useful life value; the model parameters are updated using an error back-propagation algorithm (back-propagation) until convergence.

(4) The experimental results are as follows:

the experiment adopts a ten-fold cross validation method (9 operation and maintenance cycles are used as a training set in sequence, the rest 1 operation and maintenance cycle is used as a test set, and the test is performed for 10 times in total) to predict the residual service life of the system at each moment

When the temperature is higher than the set temperature

And true RUL are all above or all below some threshold ξ, then the prediction is considered correct; otherwise, the prediction is considered as wrong. The experiments are based on the confusion matrix in terms of Accuracy (Accuracy [% ])]) Precision [% of Precision [ ])]) Recall [% ]]) And F1 value (F1-Score) the model was evaluated on 4 indices. Epoch 140, batchsize 64, hidden variable dimension 16, and threshold ξ 0.3 (empirical values to ensure that sufficient operation and maintenance preparation time remains).

As can be seen from table 1, the model performance is best when τ is 8. The Accuracy, Precision, Recall and F1-Score of the model exhibit the same trend of variation, i.e. rising between τ -1 to τ -8 with increasing time window; then, from τ -8 to τ -16, decrease; then, the rate is increased from tau to 16 to tau to 32, but the rate is not increased greatly; finally, the value of tau is decreased from 32 to 64. It can be seen that selecting an appropriate time window does help to improve various performances of the model.

TABLE 1 comparison of model Performance at different window sizes τ

τ

1

2

4

8

16

32

64

Accuracy

75.24±1.20

78.73±1.56

79.24±1.68

84.60±1.23

76.13±2.20

78.19±2.31

77.02±1.01

Precision

85.20±1.54

86.20±1.25

86.29±1.16

87.43±1.51

85.52±1.50

86.00±1.42

74.78±8.04

Recall

84.04±5.70

87.28±5.48

87.75±5.72

94.43±1.22

84.45±8.36

85.58±9.30

81.90±9.84

F1-Score

0.81±0.02

0.83±0.02

0.84±0.02

0.89±0.01

0.79±0.05

0.80±0.07

0.77±0.08

Table 2 model performance comparison under different attribute fusion modes (τ ═ 8)

Model (model)	Accuracy[％]	Precision[％]	Recall[％]	F1-Score
					Original attribute (same weight)	74.66±1.39	85.36±1.17	81.37±6.05	0.79±0.02
Original attribute (weight adaptive)	77.40±1.66	86.06±1.26	85.36±6.08	0.82±0.03
					Removing redundant Properties (same weight)	76.55±1.25	86.94±1.13	82.93±2.56	0.81±0.01
Removing redundant Properties (weight adaptive)	84.60±1.23	87.43±1.51	94.43±1.22	0.89±0.01

As can be seen from Table 2, when the number of attributes is the same, the average of Accuracy, Precision, Recall and F1-Score can be respectively improved by 6.77%, 0.59%, 7.75% and 0.06 by using the attribute adaptive network designed by the invention.

In addition, the present invention is compared to existing machine learning algorithms applied in the field of predictive maintenance. The models involved in the comparison are SVM, K-Means, RFs, LSTM, GRU and TCN, respectively. In comparative experiments, SVMs use gaussian kernel functions; the number of K-Means categories is set to 2; RFs consist of 20 decision trees; the lengths of hidden variables and the window sizes of LSTM and GRU are respectively set as 16 and 8; the size of the convolution kernel of the CNN is set to 3; the time window size of the TCN is set to 8. The experimental results of table 3 demonstrate the superiority of the present invention.

TABLE 3 comparison of the model with existing machine learning algorithms

Model (model)	Accuracy[％]	Precision[％]	Recall[％]	F1-Score
					SVM	49.85±3.56	74.92±15.79	32.42±9.93	0.40±0.10
K-Means	63.69±4.28	80.15±4.05	67.47±3.87	0.72±0.02
					RFs	76.57±1.23	85.15±1.19	84.92±5.17	0.82±0.02
LSTM	77.25±2.02	88.48±1.70	81.44±6.05	0.82±0.03
					GRU	77.98±1.66	87.81±1.41	79.06±5.59	0.81±0.02
CNN	69.85±1.44	87.30±1.58	70.74±4.98	0.75±0.02
					TCN	64.43±2.62	82.99±2.65	67.51±9.55	0.69±0.04
The invention	84.60±1.23	87.43±1.51	94.43±1.22	0.89±0.01

Reference documents:

[1]Si X S,Wang W,Hu C H,et al.Remaining useful life estimation–a review on the statistical data driven approaches[J].European journal of operational research,2011,213(1):1-14.

[2]Mosallam A,Medjaher K,Zerhouni N.Data-driven prognostic method based on Bayesian approaches for direct remaining useful life prediction[J].Journal of Intelligent Manufacturing,2016,27(5):1037-1048.

[3]Liao H,Zhao W,Guo H.Predicting remaining useful life of an individual unit using proportional hazards model and logistic regression model[C]//RAMS'06.Annual Reliability and Maintainability Symposium,2006.IEEE,2006:127-132.

[4]van Noortwijk J M.A survey of the application of gamma processes in maintenance[J].Reliability Engineering&System Safety,2009,94(1):2-21.

[5]Kumar A,Chinnam R B,Tseng F.An HMM and polynomial regression based approach for remaining useful life and health state estimation of cutting tools[J].Computers&Industrial Engineering,2019,128:1008-1014.

[6]Carvalho T P,Soares F A,Vita R,et al.A systematic literature review of machine learning methods applied to predictive maintenance[J].Computers&Industrial Engineering,2019,137:106024.

[7]Aydin O,Guldamlasioglu S.Using LSTM networks to predict engine condition on large scale data processing framework[C]//2017 4th International Conference on Electrical and Electronic Engineering(ICEEE).IEEE,2017:281-285.

[8]Lu Y W,Hsu C Y,Huang K C.An autoencoder gated recurrent unit for remaining useful life prediction[J].Processes,2020,8(9):1155.

[9]Souza R M,Nascimento E G S,Miranda U A,et al.Deep learning for diagnosis and classification of faults in industrial rotating machinery[J].Computers&Industrial Engineering,2021,153:107060.

[10]Wang Y,Deng L,Zheng L,et al.Temporal convolutional network with soft thresholding and attention mechanism for machinery prognostics[J].Journal of Manufacturing Systems,2021,60:512-526.

the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and a person skilled in the art may make modifications or equivalent substitutions to the technical solutions of the present invention without departing from the scope of the present invention, and the scope of the present invention should be determined by the claims.

Claims (10)

1. An attribute-adaptive-based method for predicting the remaining service life of industrial equipment is characterized by comprising the following steps of:

a data acquisition step: collecting time sequence data in the processing process through an intelligent sensor fixed on industrial equipment; the time series data comprises 5 types of attributes at each moment, wherein the 5 types comprise a voltage type, a current type, a power factor type and a frequency type;

and (3) attribute fusion: removing redundant attributes in the 5 types of attributes according to the three-phase circuit knowledge, combining the rest attributes, and taking a time window with a preset size to form an attribute characteristic;

a weight distribution step: processing the attribute characteristics through an attribute adaptive network, and distributing weights for the combined residual attributes to obtain attribute adaptive characteristics;

a fault diagnosis step: and processing the attribute self-adaptive characteristics and outputting a result through the long-short term memory regression network, and processing the output result through the full-connection network to predict the residual service life of the industrial equipment.

2. The method of claim 1, wherein in the data collecting step, a transformer is installed at each phase of the three-phase circuit to convert the high voltage into the low voltage.

3. The method of claim 1, wherein in the data acquisition step, fuses and earth leakage protection devices are added to the smart sensors for acquiring data.

4. The method as set forth in claim 1, wherein in the data acquisition step, the data acquisition method is Run-to-Fail, that is, the data acquisition of the current operation and maintenance cycle is stopped when a systematic machining fault is encountered; when the debugging is normal, starting to acquire data of the next operation and maintenance period; the acquisition frequency was 50 hz.

5. The method of claim 1, wherein the time series data comprises 28 attributes of 5 classes per time instant, namely: voltage class: a phase voltage, B phase voltage, C phase voltage, AB line voltage, BC line voltage and CA line voltage;

current class: phase A current, phase B current and phase C current;

power class: phase A active power, phase B active power, phase C active power, total active power, phase A reactive power, phase B reactive power, phase C reactive power, total reactive power, phase A apparent power, phase B apparent power, phase C apparent power and total apparent power;

power factor class: phase A power factor, phase B power factor, phase C power factor, total power factor;

frequency class: phase A frequency, phase B frequency, and phase C frequency.

6. The method of claim 5, wherein removing redundant attributes from the class 5 attributes comprises:

removing redundant attributes in voltage classes: AB line voltage, BC line voltage, CA line voltage;

removing redundant attributes in the power class: active power of phase A, active power of phase B, active power of phase C, total active power, total reactive power, apparent power of phase A, apparent power of phase B, apparent power of phase C and total apparent power;

removing redundant attributes in the power factor class: the total power factor.

7. The method of claim 6, wherein the remaining attributes are 15: phase A voltage, phase B voltage, phase C voltage, phase A current, phase B current, phase C current, phase A reactive power, phase B reactive power, phase C reactive power, phase A power factor, phase B power factor, phase C power factor, phase A frequency, phase B frequency, and phase C frequency.

8. The method of claim 1, wherein in the weight assigning step, the attribute adaptive network assigning weights step comprises:

a compression step: inputting the attribute features into a pooling layer to obtain intermediate features;

self-adapting: inputting the intermediate features into a fully-connected network to obtain a weight vector;

a weight distribution step: and inputting the attribute features and the weight vector, and distributing a new weight to each attribute feature to obtain the attribute adaptive feature.

9. The method of claim 1, wherein in the fault diagnosing step, the step of processing the attribute adaptivity characteristics through the long-short term memory regression network and outputting the results comprises:

inputting the attribute adaptive characteristic into a forgetting gate, and calculating to obtain a quantity f;

inputting the attribute adaptive characteristic into a memory gate, and calculating to obtain a quantity i;

calculating quantities c and h from quantities f and i;

and generating an output result according to the quantities c and h.

10. An attribute-based adaptive residual service life prediction system for industrial equipment, which is used for implementing the method of any one of claims 1 to 9, and is characterized by comprising the following steps:

the data acquisition module is used for acquiring time sequence data in the processing process through an intelligent sensor fixed on industrial equipment; the time series data comprises 5 types of attributes at each moment, wherein the 5 types comprise a voltage type, a current type, a power factor type and a frequency type;

the attribute fusion module is used for removing redundant attributes in the 5 types of attributes according to the three-phase circuit knowledge, combining the rest attributes and taking a time window with a preset size to form an attribute characteristic;

the neural network model is composed of an attribute self-adaptive network, a long-short term memory regression network and a full-connection network, and is used for processing attribute characteristics through the attribute self-adaptive network and distributing weights for the combined residual attributes to obtain attribute self-adaptive characteristics; processing the attribute self-adaptive characteristics and outputting a result through a long-term and short-term memory regression network, and then processing the output result through a full-connection network to predict the remaining service life of the industrial equipment; the neural network model is optimized by using an Adam optimizer, and parameters in the neural network model are updated according to a mean square error loss function and an error back propagation algorithm of the remaining service life until convergence.

Images