CN113669246A - Intelligent diagnosis method for water pump fault under cross-working condition - Google Patents

Abstract

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.

Description

An intelligent diagnosis method for water pump faults under cross-working conditions

technical field

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

Background technique

Water pump is a basic water conservancy equipment, which is widely deployed in various reservoirs and water inlet/outlet sites. With the continuous intensification of human activities, the pump system must maintain stable operation for a long time. Real-time detection and analysis of the health and faults of the pump, timely detection of hidden faults and maintenance, are the necessary basis for ensuring the health of the pump. Bearing system is the most important basic mechanical component of all kinds of water pump machinery. It is the mechanical component with the most hidden troubles of water pump equipment. The fault diagnosis of the bearing is particularly important for the health status diagnosis of the water pump system. For bearing fault diagnosis, vibration data is the most reliable basis, which can directly correlate with the minor fault types of water pump equipment.

In practical applications, pumps are mostly deployed in field scenes and changing water conditions, often facing complex working conditions and novel operating environments. Under this condition, prior expert knowledge and learned model parameters face ill-posed problems, while re-learning faces the problems of data scarcity and high time consumption. At this time, traditional fault diagnosis methods, such as fault diagnosis methods based on SVM and Bayesian network, are difficult to apply. In recent years, in the field of pump fault diagnosis research, although deep learning models have achieved excellent performance, it is still difficult to achieve intelligent fault diagnosis under cross-working conditions. The fault diagnosis of water pump under the condition of changing three factors of shaft power.

SUMMARY OF THE INVENTION

Purpose of the invention: Aiming at the problems and deficiencies in the prior art, the present invention provides an intelligent diagnosis method for water pump faults under cross-working conditions based on domain confrontation. Domain vibration information and target domain vibration information are aligned in the high-order vector feature space to realize intelligent diagnosis of pump faults under cross-working conditions.

Technical solution: an intelligent diagnosis method for water pump faults under cross-working conditions based on domain confrontation, comprising the following steps:

(1) Building a domain adversarial network framework for water pump fault diagnosis under changing working conditions: A domain adversarial network model framework for pump fault diagnosis under changing conditions of three factors, namely water output, head, and shaft power, is established.

(2) Construct the loss function of category classification and domain classification.

(3) Through the joint optimization of the loss function, the optimization of the domain adversarial network model is realized, and the model training and learning are completed.

(4) Input the pump vibration data under cross-working conditions into the domain adversarial network model, and at the same time complete the classification of the fault type and the working condition domain type, so as to complete the fault diagnosis.

The domain adversarial network framework for pump fault diagnosis under changing working conditions mainly includes pump vibration source domain data and pump vibration vibration domain data input, source/target domain feature extractor based on CNN network, and pump vibration data based on capsule network. High-order vector feature extractor, pump fault class classifier, pump operating condition domain classifier.

(1) Input of pump vibration source domain data and pump vibration vibration domain data under multiple working conditions: A dual-stream network architecture for parallel processing of two input data is constructed to input pump vibration data under multiple working conditions, including: source domain vibration data x ^s and target domain vibration data x ^t . Among them, the vibration data x ^s in the source domain of the pump and the vibration data x ^t in the target domain are collected in different working conditions, and there are differences in the water output, head, and shaft power when the pump is hit.

和目标域高阶矢量特征

其中G_s()为两级特征提取函数。以此，获取一维水泵振动故障特征。(2) Obtain high-order vector features of source domain through two-stage feature extraction of source domain/target domain feature extractor and pump vibration data high-order vector feature extractor

and target domain higher-order vector features

where G _s ( ) is a two-level feature extraction function. In this way, the vibration fault characteristics of the one-dimensional water pump are obtained.

当目标域数据输入时表示为：

其中length()为矢量模长计算函数，squash()为挤压函数。(3) The water pump fault category classifier calculates the vector modulus length of high-order features and takes the maximum value of the modulus length for category classification. When the source domain data is entered, it is expressed as:

When the target domain data is entered, it is expressed as:

Among them, length() is the vector modulus length calculation function, and squash() is the extrusion function.

当目标域数据输入时表示为：

其中W()为全连接计算函数，softmax()为softmax函数。(4) The domain classifier of pump working condition mainly performs domain classification through gradient inversion layer, two-layer full connection and softmax function. Represented as: When the source domain data is input, it is represented as:

When the target domain data is entered, it is expressed as:

Where W() is the full connection calculation function, and softmax() is the softmax function.

The objects of domain adversarial network framework optimization include: ① category classification, ② domain classification.

The pump fault class classifier is optimized based on the marginal loss function:

L _c =T _c max(0,m ⁺ -p _c ) ² +λ(1-T _c )max(0,p _c -m ^- ) ²

Among them, c represents the c-th label of the output; p _c represents a set of probability values output by the category classifier; T _k represents the classification indicator function, assuming that the K-th label of the output represents the category K, that is, the label is responsible for predicting the category K. probability, then when the input sample is category K and c=K, T _c =1, otherwise T _c =0; m ⁺ is the upper boundary, taking a fixed value of 0.9, when the probability value p _c >0.9, the loss function Set to 0; m ^- is the lower boundary, and takes a fixed value of 0.1. When the probability value p _c <0.1, the loss function is set to 0;

The task of the classifier in the field of pump failure conditions is a binary classification task, using the cross-entropy loss function:

和

分别表示以第i个源域数据样本输入，最后一层全连接层输出向量的第一和第二个数值；

和

分别表示以第i个目标域数据样本输入，最后一层全连接层输出向量的第一和第二个数值。第一个输出数值表示源域，第二个表示目标域，源域数据样本的领域标签为[1,0]，目标域数据样本的领域标签为[0,1]。in,

and

Respectively represent the first and second values of the output vector of the last fully connected layer with the i-th source domain data sample input;

and

Respectively represent the first and second values of the output vector of the last fully connected layer with the i-th target domain data sample input. The first output value represents the source domain, the second represents the target domain, the domain label of the source domain data sample is [1,0], and the domain label of the target domain data sample is [0,1].

θ_C，通过最小化类别分类损失来进行参数优化：In the optimization stage, the first is the classification training of pump fault source domain: using the labeled source domain data to classify the source domain feature extractor G _s , the high-order vector feature extractor G _cap of the pump vibration data based on the capsule network, and the pump fault category The parameters of the source domain feature extractor G _s , the high-order vector feature extractor G _cap of the pump vibration data based on the capsule network and the pump fault category classifier C are expressed as

θ _C , parameter optimization by minimizing the class classification loss:

为类别分类损失，

为

θ_C的优化值。in,

classification loss for the class,

for

The optimized value of θ _C.

进行优化，并且通过最小化领域分类损失对水泵工况领域分类器D参数θ_D进行优化：Then, domain classification training for pump fault conditions is performed: by maximizing the domain classification loss, the G _t parameter of the target domain feature extractor is

is optimized, and the pump condition domain classifier D parameter θ _D is optimized by minimizing the domain classification loss:

为领域分类损失，

为

θ_D的优化值。in,

is the domain classification loss,

for

The optimized value of _θD .

Beneficial effects: Compared with the prior art, the intelligent diagnosis method for water pump faults under cross-working conditions based on domain confrontation provided by the present invention proposes a water pump vibration data depth feature extractor, a water pump vibration data high-order vector feature extractor, The pump fault intelligent diagnosis domain adversarial network model composed of the pump fault category classifier and the pump working condition domain classifier; through the optimization process of model training, the model parameters of the high-order vector feature extractor are optimized, and the high difference between the data of different working conditions is reduced. The difference of order vector characteristics solves the problem of model adaptability under cross working conditions in bearing fault diagnosis; the method of the invention can be applied to complex and novel working conditions, realizes adaptive intelligent diagnosis of water pump faults, and improves the diagnosis model. adaptability and promotion capabilities.

Description of drawings

Fig. 1 is the domain confrontation network framework diagram of pump vibration data and fault diagnosis under changing working conditions in the embodiment of the present invention;

FIG. 2 is a comparison diagram of the performance of an embodiment of the present invention and a method in the prior art.

Detailed ways

Below in conjunction with specific embodiments, the present invention will be further illustrated, and it should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. The modifications all fall within the scope defined by the appended claims of this application.

Taking the fault diagnosis of water pump as an example: the vibration signal mode of the pump station changes significantly under the conditions of three factors: water output, head, and shaft power, resulting in a very different mapping relationship between vibration characteristics and fault states. This makes it difficult to apply prior expert experience and learned models under cross-working conditions. It seriously affects the adaptability and generalization ability of the model. Aiming at this problem, the present invention designs a domain adversarial network, which can realize feature alignment in the learned high-order vector feature space, so as to realize the self-adaptation of the fault diagnosis model under cross-working conditions.

As shown in Figure 1, based on the intelligent diagnosis method of water pump faults under the condition of domain confrontation, a domain confrontation network model framework is established for the fault diagnosis of water pumps under the conditions of three factors: water output, head, and shaft power. , a water pump fault intelligent diagnosis domain adversarial network model is proposed, which is composed of pump vibration data depth feature extractor, pump vibration data high-order vector feature extractor, pump fault category classifier and pump operating condition domain classifier. Through the optimization process of model training, the model parameters of the high-order vector feature extractor are optimized, the difference of high-order vector features between different working conditions data is reduced, and the model adaptability problem under cross working conditions in bearing fault diagnosis is solved.

Domain confrontation network framework, the network framework mainly includes pump vibration source domain data and pump vibration vibration domain data input, source domain/target domain feature extractor based on CNN network, high-order vector feature extractor for pump vibration data based on capsule network, The pump fault category classifier and the pump working condition field classifier are shown in Figure 1.

In the model (Domain Adversarial Network Framework) training phase:

First, a multi-layer convolutional neural network (CNN) is used to calculate the vibration data of the source and target domains, and the depth features of the source domain data and the depth features of the target domain data are extracted.

和目标域高阶矢量特征

Then, combined with the pump vibration data high-order vector feature extractor, a two-stage feature extraction was formed to obtain the source-domain high-order vector features.

and target domain higher-order vector features

其中故障的工况领域分类器为

Subsequently, the higher-order vector features are classified to classify the fault category and the operating condition domain of the fault. where the class classifier is

where the fault condition domain classifier is

Then, the classifier is optimized for training.

θ_C，通过最小化类别分类损失来进行参数优化：(1) Pump fault source domain category classifier optimization: use the labeled source domain data to analyze the source domain feature extractor G _s , the high-order vector feature extractor G _cap of the pump vibration data based on capsule network, and the pump fault category classifier C For class classification training, the source domain feature extractor G _s , the high-order vector feature extractor G _cap of the pump vibration data based on the capsule network, and the pump fault class classifier C parameters are expressed as

θ _C , parameter optimization by minimizing the class classification loss:

为

θ_C的优化值。in,

for

The optimized value of θ _C.

进行优化，并且通过最小化领域分类损失对水泵工况领域分类器D参数θ_D进行优化：(2) Domain classifier optimization for pump failure conditions: By maximizing the domain classification loss, the G _t parameters of the target domain feature extractor are

is optimized, and the pump condition domain classifier D parameter θ _D is optimized by minimizing the domain classification loss:

为

θ_D的优化值。in,

for

The optimized value of _θD .

During the troubleshooting phase:

First, a multi-layer convolutional neural network (CNN) is used to calculate the vibration data in the target domain to extract the deep features of the target domain data.

Then, combined with the pump vibration data high-order vector feature extractor, a two-level feature extraction target domain high-order vector feature was formed.

其中故障的工况领域分类器为

Subsequently, the higher-order vector features are classified to classify the fault category and the operating condition domain of the fault. where the class classifier is

where the fault condition domain classifier is

So far, the fault diagnosis of the pump across the working conditions has been completed. Tests were carried out on two types of pump equipment, A and B, respectively. A1, A2, and A3 were three working conditions with loads of 1HP, 2HP, and 3HP, respectively. condition.

The results are shown in Tables 1 and 2, and the fault diagnosis results with good accuracy are achieved: for the A type of water pump, the fault diagnosis accuracy rate under the cross-working condition remains above 96%, for the B type of water pump, the cross-working condition The fault diagnosis accuracy rate is kept above 95%, which can satisfy the field application.

Fig. 2 is a performance comparison between the method disclosed in the present invention and the existing method, wherein Fig. 2(a) is a comparative method, and Fig. 2(b) is a method proposed by the present invention. The lighter the color of the square, the higher the number of troubleshooting errors. It can be seen that the number of fault diagnosis errors of the compared methods is significantly larger than that of the method proposed in the present invention, which proves the performance advantage of the method proposed in the present invention.

Table 1. Diagnostic accuracy rate of pump A across working conditions

Table 2. Diagnostic accuracy rate of pump B across working conditions

Claims (9)

1. An intelligent diagnosis method for water pump faults under cross-working conditions based on domain confrontation, characterized in that, a domain confrontation network framework for water pump fault diagnosis under changing working conditions is constructed; Loss function; through the joint optimization of the loss function, the optimization of the domain adversarial network model is realized, and the training and learning of the domain adversarial network model is completed; the pump vibration data under cross-working conditions is input into the domain adversarial network model, and the types of pump faults are completed at the same time. And the classification of the field of working conditions, in order to complete the fault diagnosis. 2. The intelligent method for diagnosing water pump faults under cross-working conditions based on domain confrontation according to claim 1, is characterized in that, it is characterized in that, the faults of water pumps under the working conditions of three factors of changing water output, lift, and shaft power are established. A framework for domain adversarial network models for diagnosis. 3. The method for intelligent diagnosis of water pump faults under cross-working conditions based on domain confrontation according to claim 1, wherein the domain confrontation network framework for water pump fault diagnosis under said changing working conditions mainly comprises water pump vibration source domain data And pump vibration and vibration domain data input, source/target domain feature extractor based on CNN network, high-order vector feature extractor for pump vibration data based on capsule network, pump fault category classifier, and pump working condition domain classifier. 4. The intelligent method for diagnosing water pump faults under cross-working conditions based on domain confrontation according to claim 3, wherein the input of the water pump vibration source domain data and the water pump vibration vibration domain data under the multi-working condition is as follows: The dual-stream network architecture that processes two input data in parallel inputs the pump vibration data under multiple working conditions, including: the pump source domain vibration data x ^s and the target domain vibration data x ^t ; among them, the pump source domain vibration data x ^s and the target domain vibration data x s The working conditions and environments collected by the vibration data x ^t are different, and there are differences in the water output, head, and shaft power of the thunder. 5. The method for intelligently diagnosing water pump faults under cross-working conditions based on domain confrontation according to claim 3, is characterized in that, through the source domain/target domain feature extractor and pump vibration data high-order vector feature extractor two-level feature Extract and obtain high-order vector features in the source domain

and target domain higher-order vector features

Among them, G _s ( ) is a two-level feature extraction function; in this way, the one-dimensional pump vibration fault feature is obtained; The water pump fault classification classifier calculates the vector modulus length of high-order features and takes the maximum value of the modulus length for classification; when the source domain data is input, it is expressed as:

When the target domain data is entered, it is expressed as:

Among them, length() is the vector modulus length calculation function, and squash() is the extrusion function; The water pump working condition domain classifier performs domain classification through gradient inversion layer, two-layer full connection and softmax function; when the source domain data is input, it is expressed as

When the target domain data is entered, it is expressed as:

Where W() is the full connection calculation function, and softmax() is the softmax function. 6. The method for intelligently diagnosing water pump faults under cross-working conditions based on domain confrontation according to claim 1, wherein the optimization object of the domain confrontation network model comprises: ① category classification, ② domain classification; The pump fault class classifier is optimized based on the marginal loss function: L _c =T _c max(0,m ⁺ -p _c ) ² +λ(1-T _c )max(0,p _c -m ^- ) ² Among them, c represents the c-th label of the output; p _c represents a set of probability values output by the category classifier; T _k represents the classification indicator function, assuming that the K-th label of the output represents the category K, that is, the label is responsible for predicting the category K. probability, when the input sample is category K and c=K, T _c =1, otherwise T _c =0; m ⁺ is the upper boundary, when the probability value p _c >m ⁺ , the loss function is set to 0; m ^- is the lower boundary, when the probability value p _c < m ^- , the loss function is set to 0; λ is a scale coefficient, which is used to adjust the two proportions. 7. The method for intelligently diagnosing water pump faults under cross-working conditions based on domain confrontation according to claim 6, wherein the task of the water pump fault working condition domain classifier is a two-classification task, and a cross-entropy loss function is used:

in,

and

Respectively represent the first and second values of the output vector of the last fully connected layer with the i-th source domain data sample input;

and

Respectively represent the first and second values of the i-th target domain data sample input, and the output vector of the last fully connected layer; the first output value represents the source domain, the second represents the target domain, and the value of the source domain data sample The domain label is [1,0], and the domain label of the target domain data sample is [0,1]. 8. The method for intelligent diagnosis of water pump faults under cross-working conditions based on domain confrontation according to claim 6, characterized in that, in the optimization stage, the first is the classification training of water pump fault source domain categories: using the labeled source domain data to The source domain feature extractor G _s , the high-order vector feature extractor G _cap of the water pump vibration data based on the capsule network and the water pump fault classification classifier C are used for class classification training, the source domain feature extractor G _s , the water pump vibration data based on the capsule network The parameters of the high-order vector feature extractor G _cap and the water pump fault category classifier C are expressed as

θ _C , parameter optimization by minimizing the class classification loss:

in,

for

The optimized value of θ _C ; Then, domain classification training for pump fault conditions is performed: by maximizing the domain classification loss, the G _t parameter of the target domain feature extractor is

is optimized, and the pump condition domain classifier D parameter θ _D is optimized by minimizing the domain classification loss:

in,

for

The optimized value of _θD . 9. The intelligent diagnosis method for water pump faults under cross-working conditions based on domain confrontation according to claim 1, characterized in that, in the fault diagnosis stage: First, the multi-layer convolutional neural network is used to calculate the vibration data of the target domain, and the depth features of the target domain data are extracted; Then, combined with the pump vibration data high-order vector feature extractor, a two-level feature extraction target domain high-order vector feature was formed.

Then, the high-order vector features are classified to classify the fault category and the working condition field of the fault; the fault diagnosis of the pump across working conditions is completed.

Images