阅读说明：本技术 一种跨工况条件下水泵故障智能诊断方法 (Intelligent diagnosis method for water pump fault under cross-working condition ) 是由 杨德玮 周克发 杨军 刘晓杰 陈莹颖 陈伟 董凯 于 2021-08-23 设计创作，主要内容包括：本发明公开了一种跨工况条件下水泵故障智能诊断方法,涉及振动信号处理、模式识别等技术领域。该方法构建了一种域对抗网络框架,使得原始工况条件下获取的水泵振动数据(源域)和当前工况条件下获取的水泵振动数据(目标域)在高阶矢量特征空间中对齐,实现自适应的水泵故障诊断。在此网络框架下提出了一种域对抗胶囊网络,能够实现跨工况条件下水泵故障智能诊断。本发明能够稳定、可靠地用于野外复杂工况条件下的水泵故障诊断,具有较高的可靠性和适用性。(The invention discloses an intelligent diagnosis method for water pump faults under a cross-working condition, and relates to the technical fields of vibration signal processing, pattern recognition and the like. According to the method, a domain confrontation network framework is constructed, so that water pump vibration data (source domain) obtained under the original working condition and water pump vibration data (target domain) obtained under the current working condition are aligned in a high-order vector feature space, and self-adaptive water pump fault diagnosis is realized. The domain confrontation capsule network is provided under the network framework, and intelligent diagnosis of water pump faults under the cross-working condition can be realized. The method can be stably and reliably used for diagnosing the water pump fault under the field complex working condition, and has higher reliability and applicability.)

1. An intelligent diagnosis method for water pump faults under cross-working condition based on domain confrontation is characterized in that a domain confrontation network framework for water pump fault diagnosis under the condition of changing working conditions is built; constructing a loss function of classification of water pump fault categories and classification of water pump working condition fields; the optimization of the domain confrontation network model is realized through the combined optimization of the loss function, and the training and learning of the domain confrontation network model are completed; and inputting the water pump vibration data under the cross-working condition into a domain confrontation network model, and simultaneously finishing the classification of the water pump fault type and the working condition field type so as to finish fault diagnosis.

2. The method for intelligently diagnosing the water pump fault under the cross-working-condition based on the domain confrontation as claimed in claim 1, wherein a domain confrontation network model framework for diagnosing the water pump fault under the working condition of three factors of water yield, lift and shaft power during thunderstroke is established.

3. The method for intelligently diagnosing the water pump fault under the cross-working-condition based on the domain countermeasure as claimed in claim 1, wherein the domain countermeasure network framework for diagnosing the water pump fault under the changing working condition mainly comprises water pump vibration source domain data and water pump vibration domain data input, a source domain/target domain feature extractor based on a CNN network, a water pump vibration data high-order vector feature extractor based on a capsule network, a water pump fault category classifier and a water pump working condition domain classifier.

4. The method for intelligently diagnosing water pump faults under cross-working condition based on domain confrontation as claimed in claim 3, wherein the method is characterized by multiple processesUnder the condition, the water pump vibration source domain data and the water pump vibration domain data are input as follows: the method for constructing a double-flow network architecture for processing two paths of input data in parallel and respectively inputting water pump vibration data under the multi-working-condition comprises the following steps: water pump source domain vibration data x^sAnd target domain vibration data x^t(ii) a Wherein, the water pump source domain vibration data x^sAnd target domain vibration data x^tThe collected working conditions and environments are different, and the water yield, the lift and the shaft power are different during thunderstrikes.

5. The method for intelligently diagnosing the water pump fault under the cross-working-condition based on the domain countermeasure as claimed in claim 3, wherein the source domain high-order vector feature is obtained through two-stage feature extraction of a source domain/target domain feature extractor and a water pump vibration data high-order vector feature extractorAnd target domain higher order vector featuresWherein G is_s() Extracting a function for the two-level features; thus, acquiring vibration fault characteristics of the one-dimensional water pump;

the water pump fault category classifier calculates the vector modular length of the high-order features and takes the maximum value of the modular length to classify the categories; when source domain data is entered, it is expressed as:when target domain data is entered, it is expressed as:wherein length () is a vector modulo length calculation function, and square () is a squeeze function;

the water pump working condition field classifier performs field classification through a gradient inversion layer, two layers of full connection and a softmax function; when the source domain data is input, is expressed asWhen target domain data is entered, it is expressed as:where W () is the fully-connected computation function and softmax () is the softmax function.

6. The intelligent diagnosis method for water pump faults under cross-working-condition conditions based on domain confrontation, as claimed in claim 1, wherein the optimization objects of the domain confrontation network model comprise: classifying the categories, and classifying the fields;

the water pump fault category classifier optimization is based on an edge loss function:

L_c＝T_c max(0,m⁺-p_c)²+λ(1-T_c)max(0,p_c-m^-)²

wherein c represents the output c-th label; p is a radical of_cA set of probability values representing the output of the class classifier; t is_kRepresenting a classification indication function, assuming that the output kth label represents class K, i.e. the label is responsible for predicting the probability of class K, when the input sample is class K and c is K, T is_c1, otherwise T_c＝0；m⁺For upper bounds, when probability value p_c>m⁺Setting the loss function to 0; m is^-To the lower boundary, when the probability value p_c<m^-Setting the loss function to 0; λ is a scaling factor used to adjust the two terms.

7. The method for intelligently diagnosing the water pump fault under the cross-working-condition based on the domain confrontation as claimed in claim 6, wherein the task of the water pump fault working condition field classifier is a two-classification task, and a cross entropy loss function is used:

wherein the content of the first and second substances,andrespectively representing a first numerical value and a second numerical value of a final layer full-connection layer output vector input by an ith source domain data sample;andrespectively representing a first numerical value and a second numerical value of a final layer full-connection layer output vector input by an ith target domain data sample; the first output value represents the source domain, the second represents the target domain, and the domain label of the source domain data sample is [1,0 ]]The domain label of the target domain data sample is [0,1 ]]。

8. The method for intelligently diagnosing the water pump fault under the cross-working-condition based on the domain countermeasure as claimed in claim 6, wherein in the optimization stage, the classification training of the water pump fault source domain is firstly carried out: source domain feature extractor G using tagged source domain data_sWater pump vibration data high-order vector feature extractor G based on capsule network_capPerforming class classification training with a water pump fault class classifier C, and extracting a source domain feature extractor G_sWater pump vibration data high-order vector feature extractor G based on capsule network_capAnd the water pump fault category classifier C parameter is expressed asθ_CParameter optimization is performed by minimizing class classification loss:

wherein the content of the first and second substances,is composed ofθ_CAn optimized value of (d);

then, carrying out classification training in the field of water pump fault conditions: target domain feature extractor G by maximizing domain classification loss_tParameter(s)Optimizing and classifying D parameter theta of water pump working condition field classifier by minimizing field classification loss_DOptimizing:

wherein the content of the first and second substances,is composed ofθ_DThe optimum value of (c).

9. The intelligent diagnosis method for the water pump fault under the cross-working condition based on the domain countermeasure as claimed in claim 1, characterized in that in the fault diagnosis stage:

firstly, calculating vibration data of a target domain by adopting a multilayer convolutional neural network, and extracting depth characteristics of the data of the target domain;

subsequently, a two-stage feature extraction target domain high-order vector feature is formed by combining a water pump vibration data high-order vector feature extractor

Then, classifying the high-order vector characteristics, and classifying the fault category and the fault working condition field; and the fault diagnosis of the water pump across working conditions is completed.

Technical Field

The invention relates to a water pump fault intelligent diagnosis method under a cross-working condition based on domain confrontation, in particular to a method for realizing alignment of source domain vibration information and target domain vibration information in a high-order vector characteristic space through model learning and optimization so as to realize intelligent diagnosis of water pump faults under the cross-working condition, and belongs to the technical field of vibration signal processing and mode recognition.

Background

The water pump is a basic water conservancy device and is widely arranged at various reservoirs and water inlet/outlet sites. With the increasing human activities, the water pump system must be kept in stable operation for a long time. The system can detect and analyze the health and the fault of the pump in real time, find the hidden fault trouble and overhaul in time, and is a necessary basis for guaranteeing the health state of the water pump. The bearing system is the most important basic mechanical element of various water pump machines, is large in quantity and rolls for a long time, and is the most frequent mechanical element of the hidden trouble of water pump equipment. The fault diagnosis for the bearing is particularly important for the health state diagnosis of the water pump system. For bearing fault diagnosis, vibration data is the most reliable basis and can be directly related to tiny fault types of water pump equipment.

In practical application, the water pumps are arranged in outdoor scenes and changeable water conditions, and are often subjected to complex working conditions and novel operating environments. Under the condition, the prior expert knowledge and the learned model parameters face the problem of unsuitability, and the relearning faces the problems of data deficiency and high time loss. At this time, the conventional fault diagnosis method, such as the fault diagnosis method based on SVM and bayesian network, is difficult to apply. In recent years, in the field of water pump fault diagnosis research, although a deep learning model has acquired excellent performance, fault intelligent diagnosis under a cross-working condition is still difficult to realize, and a priori deep learning model is difficult to be suitable for water pump fault diagnosis under the condition of changing working conditions of three factors, namely water yield, lift and shaft power during thunderstroke.

Disclosure of Invention

The purpose of the invention is as follows: aiming at the problems and the defects in the prior art, the invention provides a water pump fault intelligent diagnosis method under the cross-working condition based on domain confrontation, which realizes the alignment of source domain vibration information and target domain vibration information in a high-order vector characteristic space through fault classification and working condition domain classification model learning and joint optimization so as to realize the intelligent diagnosis of the water pump fault under the cross-working condition.

The technical scheme is as follows: a water pump fault intelligent diagnosis method under a cross-working condition based on domain confrontation comprises the following steps:

the method comprises the following steps of (I) building a domain countermeasure network framework for water pump fault diagnosis under the condition of variable working conditions: a domain confrontation network model framework for diagnosing the water pump fault under the condition of three factors of water yield, lift and shaft power during thunderstroke is established.

And (II) constructing loss functions of category classification and field classification.

And (III) realizing optimization of the domain confrontation network model through combined optimization of the loss function, and finishing model training and learning.

And (IV) inputting the water pump vibration data under the cross-working condition into the domain confrontation network model, and simultaneously finishing the classification of the fault type and the working condition domain type so as to finish fault diagnosis.

The domain countermeasure network framework for water pump fault diagnosis under the condition of changing working conditions mainly comprises water pump vibration source domain data and water pump vibration domain data input, a source domain/target domain feature extractor based on a CNN network, a water pump vibration data high-order vector feature extractor based on a capsule network, a water pump fault category classifier and a water pump working condition domain classifier.

(1) Inputting water pump vibration source domain data and water pump vibration domain data under the multi-working condition: the method has the advantages that a double-flow network architecture for processing two paths of input data in parallel is constructed to respectively input water pump vibration data under the multi-working-condition, and the method comprises the following steps: source domain vibration data x^sAnd target domain vibration data x^t. It is composed ofIn, water pump source domain vibration data x^sAnd target domain vibration data x^tThe collected working conditions and environments are different, and the water yield, the lift and the shaft power are different during thunderstrikes.

(2) Obtaining source domain high-order vector characteristics through two-stage characteristic extraction of a source domain/target domain characteristic extractor and a water pump vibration data high-order vector characteristic extractorAnd target domain higher order vector featuresWherein G is_s() A function is extracted for the two-level features. Therefore, the vibration fault characteristics of the one-dimensional water pump are obtained.

(3) And the water pump fault class classifier calculates the vector modular length of the high-order features and takes the maximum value of the modular length to classify the classes. When source domain data is entered, it is expressed as:when target domain data is entered, it is expressed as:where length () is the vector modulo length computation function and square () is the squeeze function.

(4) The water pump working condition field classifier is mainly used for carrying out field classification through a gradient inversion layer, two layers of full connection and a softmax function. Expressed as: when source domain data is entered, it is expressed as:when target domain data is entered, it is expressed as:where W () is the fully-connected computation function and softmax () is the softmax function.

The object of the domain confrontation network framework optimization comprises the following objects: classification of category and classification of field.

The water pump fault category classifier optimization is based on an edge loss function:

L_c＝T_cmax(0,m⁺-p_c)²+λ(1-T_c)max(0,p_c-m^-)²

wherein c represents the output c-th label; p is a radical of_cA set of probability values representing the output of the class classifier; t is_kRepresenting a classification indication function, assuming that the output kth label represents class K, i.e. the label is responsible for predicting the probability of class K, when the input sample is class K and c is K, T is_c1, otherwise T_c＝0；m⁺For the upper boundary, take a fixed value of 0.9 as the probability value p_c>At 0.9, setting the loss function to 0; m is^-For the lower boundary, take a fixed value of 0.1 as the probability value p_c<At 0.1, setting the loss function to 0; λ is a proportionality coefficient, which is used to adjust the two terms, usually 0.5.

The task of the water pump fault working condition field classifier is a two-classification task, and a cross entropy loss function is used:

wherein the content of the first and second substances,andrespectively representing a first numerical value and a second numerical value of a final layer full-connection layer output vector input by an ith source domain data sample;andrespectively representing the first and second values of the output vector of the final layer of the fully-connected layer input with the ith target domain data sample. The first output value represents the source domain, the second represents the target domain, the source domain dataThe field label of the sample is [1,0 ]]The domain label of the target domain data sample is [0,1 ]]。

In the optimization stage, firstly, classification training of the water pump fault source domain is carried out: source domain feature extractor G using tagged source domain data_sWater pump vibration data high-order vector feature extractor G based on capsule network_capPerforming class classification training with a water pump fault class classifier C, and extracting a source domain feature extractor G_sWater pump vibration data high-order vector feature extractor G based on capsule network_capAnd the water pump fault category classifier C parameter is expressed asθ_CParameter optimization is performed by minimizing class classification loss:

wherein the content of the first and second substances,in order to classify the loss for a class,is composed ofθ_CThe optimum value of (c).

Then, carrying out classification training in the field of water pump fault conditions: target domain feature extractor G by maximizing domain classification loss_tParameter(s)Optimizing and classifying D parameter theta of water pump working condition field classifier by minimizing field classification loss_DOptimizing:

wherein the content of the first and second substances,in order to be a loss of domain classification,is composed ofθ_DThe optimum value of (c).

Has the advantages that: compared with the prior art, the intelligent diagnosis method for the water pump fault under the cross-working condition based on the domain confrontation, provided by the invention, provides a water pump fault intelligent diagnosis domain confrontation network model which consists of a water pump vibration data depth feature extractor, a water pump vibration data high-order vector feature extractor, a water pump fault category classifier and a water pump working condition field classifier; through the optimization process of model training, the model parameters of the high-order vector feature extractor are optimized, the high-order vector feature difference between different working condition data is reduced, and the problem of model adaptability under the cross-working condition in the bearing fault diagnosis is solved; the method provided by the invention can be suitable for complex and novel working conditions, realizes adaptive intelligent diagnosis of water pump faults, and improves the adaptability and popularization capability of a diagnosis model.

Drawings

FIG. 1 is a domain countermeasure network framework diagram of water pump vibration data and fault diagnosis under varying conditions according to an embodiment of the present invention;

FIG. 2 is a graph comparing the performance of the method of the present invention with that of the prior art.

Detailed Description

The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.

The water pump fault diagnosis is taken as an embodiment: the vibration signal mode of the pump station is obviously changed under the condition that three factors of water yield, lift and shaft power are changed during thunderstroke, so that the mapping relation between the vibration characteristics and the fault state is different. Leading to prior expert experience and learned models that are difficult to apply under cross-regime conditions. The adaptability and popularization capability of the model are seriously influenced. Aiming at the problem, the invention designs a domain countermeasure network which can realize feature alignment in a high-order vector feature space obtained by learning, thereby realizing the self-adaptation of a fault diagnosis model under the cross-working condition.

As shown in figure 1, a domain confrontation network model framework for water pump fault diagnosis under the condition of three factors of water yield, lift and shaft power change during thunderstroke is established based on the domain confrontation intelligent diagnosis method for water pump faults under the cross-working condition, and a water pump fault intelligent diagnosis domain confrontation network model consisting of a water pump vibration data depth feature extractor, a water pump vibration data high-order vector feature extractor, a water pump fault category classifier and a water pump working condition field classifier is provided. Through the optimization process of model training, the model parameters of the high-order vector feature extractor are optimized, the high-order vector feature difference between different working condition data is reduced, and the problem of model adaptability under the cross-working condition in bearing fault diagnosis is solved.

The network framework mainly comprises water pump vibration source domain data, water pump vibration domain data input, a source domain/target domain feature extractor based on a CNN (CNN network), a water pump vibration data high-order vector feature extractor based on a capsule network, a water pump fault category classifier and a water pump working condition domain classifier, and is shown in figure 1.

In the model (domain confrontation network framework) training phase:

firstly, vibration data of a source domain and vibration data of a target domain are calculated by adopting a multilayer Convolutional Neural Network (CNN), and depth characteristics of the source domain data and depth characteristics of the target domain data are extracted.

Subsequently, a two-stage characteristic extraction is formed by combining a water pump vibration data high-order vector characteristic extractor to obtain source domain high-order vector characteristicsAnd target domain higher order vector features

And then, classifying the high-order vector characteristics, and classifying the fault category and the fault working condition field. Wherein the class classifier isWherein the working condition field classifier of the fault is

Subsequently, the classifier is optimally trained.

(1) Optimizing a water pump fault source domain category classifier: source domain feature extractor G using tagged source domain data_sWater pump vibration data high-order vector feature extractor G based on capsule network_capPerforming class classification training with a water pump fault class classifier C, and extracting a source domain feature extractor G_sWater pump vibration data high-order vector feature extractor G based on capsule network_capAnd the water pump fault category classifier C parameter is expressed asθ_CParameter optimization is performed by minimizing class classification loss:

wherein the content of the first and second substances,is composed ofθ_CThe optimum value of (c).

(2) Optimizing a classifier in the field of water pump fault conditions: target domain feature extractor G by maximizing domain classification loss_tParameter(s)Optimizing and classifying D parameter theta of water pump working condition field classifier by minimizing field classification loss_DOptimizing:

wherein the content of the first and second substances,is composed ofθ_DThe optimum value of (c).

In the fault diagnosis stage:

firstly, the vibration data of a target domain is calculated by adopting a multilayer Convolutional Neural Network (CNN), and the depth feature of the target domain data is extracted.

Subsequently, a two-stage feature extraction target domain high-order vector feature is formed by combining a water pump vibration data high-order vector feature extractor

And then, classifying the high-order vector characteristics, and classifying the fault category and the fault working condition field. Wherein the class classifier isWherein the working condition field classifier of the fault is

Therefore, the fault diagnosis of the water pump across working conditions is completed. The tests are respectively carried out on A, B water pump devices, wherein A1, A2 and A3 are respectively three working condition conditions of 1HP, 2HP and 3HP of the load, and B1, B2 and B3 are respectively three working condition conditions of 600RPM, 800RPM and 1000 RPM.

The results are shown in tables 1 and 2, and the fault diagnosis result with better accuracy is realized: for the A type of water pumps, the fault diagnosis accuracy rate under the cross-working condition is kept above 96%, and for the B type of water pumps, the fault diagnosis accuracy rate under the cross-working condition is kept above 95%, so that the field application can be met.

FIG. 2 is a comparison of the performance of the disclosed method with a prior art method, wherein FIG. 2(a) is a comparison method, and wherein FIG. 2(b) is a comparison method. Wherein lighter color of the squares indicates a greater number of troubleshooting errors. It can be seen that: the number of fault diagnosis errors of the compared method is obviously larger than that of the method provided by the invention, and the performance advantage of the method provided by the invention is proved.

TABLE 1 diagnosis accuracy of Water Pump A Cross-working Condition

TABLE 2 diagnosis accuracy of B-span working condition of water pump