CN116028847A - Universal method and system for automatic intelligent diagnosis of turbine mechanical faults - Google Patents

Abstract

The invention relates to a universal method and a universal system for automatic intelligent diagnosis of turbine mechanical faults, comprising the following steps: collecting and storing vibration signals in x and y directions at the axis through a sensor arranged on the turbomachine; carrying out data preprocessing and data cleaning on the vibration signals, and coupling in the x-direction and the y-direction to generate an axis track image; training to generate an OrbitNet model according to the data of the extracted vibration signals and verifying the model; and carrying out fault identification on the large-scale turbomachinery part by using an OrbitNet model. The method provided by the invention has the advantages that the obtained axial locus is subjected to image enhancement, the image representativeness is further improved, and the general type and the robustness of the built model are improved, wherein the axial locus diagram effectively represents the failure mechanism of the typical failure of the turbomachine component, so that the data and the failure mechanism are effectively combined, and the accuracy and the universality of the method are improved.

Description

Universal method and system for automatic intelligent diagnosis of turbine mechanical faults

Technical Field

The invention relates to the technical field of fault identification, in particular to a universal method and system for automatic intelligent diagnosis of turbine mechanical faults.

Background

Large turbine rotary machines such as gas turbines, steam turbines, and turbine compressors are consumer and industrial core power equipment for electricity, petroleum, chemical, and metallurgy. Once the key components such as the bearing, the rotating shaft and the like in the turbine equipment are out of order, the machine is not scheduled to stop running, so that the reliability and the productivity of the equipment are reduced, the running and maintenance cost of the equipment is increased, and the casualties and huge economic losses are caused when the equipment is serious.

In recent years, with the rapid development of sensor technology, information and communication technology, internet of things and other technologies, state monitoring and diagnosis technology has been widely applied to various products. During the monitoring process, various sensors are installed on a large-scale equipment group, each equipment is provided with a plurality of measuring points, the frequency of data sampling is high, and the data collection duration from the start of service to the end of service life of the equipment is long, which inevitably brings us into an industrial big data era. The data-driven fault diagnosis method is more and more important, and the main purpose of the data-driven fault diagnosis method is to hope to learn the expression form of faults from a large amount of monitoring data, so that the health condition of the machine is autonomously identified, the safe and reliable operation of the machine is facilitated, and a large amount of manual operation and maintenance cost is saved. Currently, two types of methods, namely feature extraction and fault identification, are mainly used for a data-driven fault diagnosis method.

The feature extraction can remove irrelevant data and redundant data, so that the calculation complexity is reduced, the calculation time is saved, the machine learning efficiency and effect are improved, and the high-quality features are beneficial to improving the overall performance and accuracy of the model. The common data feature extraction method mainly focuses on three aspects of time domain, frequency domain and time domain. The time domain feature extraction method is directly performed on the basis of the collected signals (such as sound and vibration signals), and typically, the feature parameters are extracted by adopting a statistical method. The spectrum of a signal refers to a representation of the signal in the frequency domain, and can provide frequency information contained in the signal more intuitively than a time domain waveform. Frequency domain techniques are considered more efficient in terms of fault diagnosis because they have good ability to identify and isolate frequency components. The time-frequency analysis method considers the time domain and frequency domain information at the same time, clearly describes the change relation of the signal frequency along with the time course, and is often used for processing unsteady signals.

The essence of fault diagnosis of parts in turbine rotary mechanical equipment is to identify classification problems, a feature extraction method is utilized to provide key feature parameters in state information, and then a pattern recognition method is utilized to diagnose, wherein common pattern recognition methods comprise a support vector machine, a random forest method, a decision tree, a long-short-term memory neural network, a convolutional neural network and the like. However, large turbomachinery is often very complex, each unit is composed of a plurality of parts, and the parts are mutually connected and tightly coupled, so that extremely complicated relations are shown between fault reasons and fault symptoms, namely, one fault symptom can correspond to multiple fault reasons, and one fault reason can correspond to multiple fault symptoms. Moreover, large turbomachinery often presents some uncertainty, even for units of the same type, the characteristics and the linking characteristics of the various parts are different, and the operating time and the operating state are not completely determined. Therefore, when the conditions of the training data and the conditions of the diagnosis target data are different, diagnosis is very difficult.

In summary, the existing fault diagnosis method has the following main problems: first, early experimental simulations and finite element analysis are widely used for fault diagnosis of these turbine rotating components, which are time consuming and consume significant computational costs; secondly, the existing fault diagnosis method is generally based on monitoring data only, and failure mechanisms are not considered; thirdly, the running condition of the large-scale turbomachinery is complex and changeable, a plurality of uncertainties such as measurement errors, data acquisition deviations and the like exist in multivariate monitoring vibration data, and the uncertainty in the data is ignored in the existing fault diagnosis method based on the axis trajectory graph, so that the accuracy of a diagnosis result is not high; fourth, in the existing fault diagnosis method based on the axial trace diagram for the large-scale turbomachinery, manual feature extraction is needed, but the manual feature extraction brings additional uncertainty to fault detection, so that inaccuracy of a result is further caused.

Disclosure of Invention

The invention aims to overcome the defects in the prior art and provide a universal method and a universal system for automatic intelligent diagnosis of turbine mechanical faults.

The invention is realized by the following technical scheme:

a general method for automatically and intelligently diagnosing turbine mechanical faults comprises the following steps:

s1, collecting vibration signals in the x and y directions at the axis through a sensor arranged on the turbomachine and storing the vibration signals;

s2, carrying out data preprocessing and data cleaning on the vibration signals, and coupling in the x-direction and the y-direction to generate an axis track image;

s3, training to generate an OrbitNet model according to the extracted data of the vibration signals and verifying the model;

s4, performing fault identification on the large-scale turbine mechanical parts by using the OrbitNet model.

According to the above technical solution, in step S1, preferably, the vibration signal is stored in a local server or a cloud data platform on a monitoring site.

According to the above technical solution, preferably, the "data preprocessing" in step S2 includes: removing an abnormal value of the extracted vibration signal; and normalizing the data by adopting a minimum-maximum method, and marking and encoding the generated image by a binarization method to obtain multi-class output.

According to the above technical solution, preferably, the "data cleaning" in step S2 includes: cleaning the preprocessed vibration signals by adopting a Bayes wavelet denoising method; and evaluating the denoising effect on the vibration signal after data cleaning.

According to the above-described aspect, preferably, the "evaluating the denoising effect" includes: qualitatively evaluating the denoising effect through the time domain and frequency domain graphs; and quantitatively evaluating the denoising effect through the signal-to-noise ratio and the sample entropy ratio.

According to the above technical solution, preferably, in step S2, after the "generating the axial trace image", the axial trace image is compressed.

According to the above technical solution, preferably, step S3 includes: dividing the axis locus image into three groups of training samples, verification samples and test samples; utilizing a CNN model to automatically capture characteristics of a track image, and constructing an orbitNet model; adjusting the parameters of an OrbitNet model by adopting the training sample; evaluating the trained model precision by adopting the verification sample; and evaluating the recognition accuracy of the established model by adopting the test sample.

According to the above technical solution, preferably, in step S3, the training sample is subjected to image enhancement by a rotation, translation, and flipping method.

The patent also discloses an automatic intelligent diagnosis system of turbomachinery trouble, include:

the sensor is used for collecting vibration signals in the x and y directions at the upper axis of the turbomachine;

the cloud data platform is respectively in communication connection with the sensor and the edge computing monitoring diagnosis system;

and the edge calculation monitoring and diagnosing system is used for carrying out data preprocessing and data cleaning on the vibration signals and carrying out fault identification on the large-scale turbomachinery parts by utilizing the OrbitNet model generated by training.

The beneficial effects of the invention are as follows:

the invention designs a new model orbit Net of a convolutional neural network based on an axis locus diagram aiming at large turbine rotating machinery, and utilizes the axis locus diagram generated by vibration data to automatically identify and diagnose faults of large turbine rotating parts through the seamless combination of advanced signal processing and a deep learning method of image processing. The invention is based on a convolutional neural network deep learning model, avoids the defect that the traditional fault diagnosis model requires to perform feature extraction on data or images in advance, realizes automatic extraction of fault features, effectively fuses monitoring data and failure mechanisms in fault diagnosis through an axis trajectory graph, improves diagnosis precision, and has good application and popularization values.

Drawings

FIG. 1 is a flow chart of a general method of the present invention for turbine mechanical fault identification.

FIG. 2 is an orbitNET fault diagnosis model based on an axial trace diagram of the present invention: a) Architecture, b) convolution layer.

FIG. 3 is a diagram of 4 different faults contained in three types of turbomachinery.

Fig. 4 is a failed centrifugal compressor wheel: a) Blade cracking, and b) shaft eccentricity.

Fig. 5 is SHBT generator data (containing 4 oil vortex events): a) x-direction time series data, b) x-direction data box diagrams, c) y-direction time series data, d) y-direction data box diagrams.

Fig. 6 is a comparison of x-direction vibration raw data versus denoising data in the XLA case: a) raw time series data, b) PSD indicators of raw data, c) histograms of raw data (shaded) and denoised data (bold line), d) PSD indicators of denoised data.

Fig. 7 is a diagram of vibration data axis trace of oil vortex fault of the SHBT generator: a) original vibration data in the x and y directions of a fault state, b) an original data axis trajectory graph, c) denoising vibration data, and d) an axis trajectory graph of denoising data.

FIG. 8 is a graph of vibration data axis traces for a rotor imbalance fault in an SXLA centrifugal compressor: a) normal denoising vibration data, b) a normal denoising data axis trace, c) fault denoising vibration data, d) a fault denoising data axis trace.

Fig. 9 is a diagram of the number of axial trace diagrams for OrbitNet model training, validation and testing.

Fig. 10 is a gray axis trace plot for four failure modes: a) rotor misalignment, b) oil vortex, c) friction, and d) rotor imbalance.

FIG. 11 is an OrbitNet model parameter tuning and cross-validation results.

Fig. 12 is a comparison of performance of three CNN models.

FIG. 13 is accuracy and error trend during the SquezeNet model parameter adjustment process.

FIG. 14 is a comparative analysis of fault identification for three OrbitNet and SqueezeNet models.

FIG. 15 is a device of the present invention employing an OrbitNet intelligent diagnostic system.

Detailed Description

The present invention will be described in further detail below with reference to the drawings and preferred embodiments, so that those skilled in the art can better understand the technical solutions of the present invention. All other embodiments, based on the embodiments of the invention, which would be apparent to one of ordinary skill in the art without making any inventive effort are intended to be within the scope of the invention.

Example 1 as shown, the present invention comprises the steps of:

s1, collecting and storing vibration signals in the x and y directions at the axis through a sensor arranged on the turbomachine, and storing the vibration signals in a local server or a cloud data platform on a monitoring site.

For subsequent establishment of an OrbitNet deep learning model, different types of turbomachine (e.g., turbine and centrifugal compressor) health and fault state data are extracted from various customer sites of the cloud data platform. In addition, the stored data typically includes alarms, trips, shutdown events, fault diagnostics comments, and turbine maintenance records, and typically also contains different types of faults such as misalignment, oil vortex, friction, imbalance, and blade cracking.

S2, carrying out data preprocessing and data cleaning on the vibration signals so as to improve the accuracy of the data driving model. And coupling in the x and y directions to generate an axial locus image, compressing the generated gray scale image into a size of 128x128 so as to improve the calculation efficiency, and simultaneously, keeping the key characteristics of each locus in the compressed image.

Wherein, "data preprocessing" includes: removing an abnormal value of the extracted vibration signal; and normalizing the data by adopting a minimum-maximum method, and marking and encoding the generated image by a binarization method to obtain multi-class output.

In addition, "data cleansing" includes: cleaning the preprocessed vibration signals by adopting a Bayes wavelet denoising method; and qualitatively evaluating the denoising effect on the vibration signal after data cleaning through time domain and frequency domain graphs, and quantitatively evaluating the denoising effect through a signal-to-noise ratio and a sample entropy ratio. Specifically, the present invention first measures the original signal that typically contains various noises for various reasons, such as initial spurious vibration signals, power instability, inaccurate readings, environmental disturbances, and transmission errors. The invention develops an advanced Bayesian wavelet denoising method to clean the original vibration signal, and furthest retains useful information to carry out fault diagnosis. After denoising, the denoising effect is qualitatively evaluated through graph comparison, and the effectiveness of the Bayesian wavelet packet denoising method is quantitatively evaluated through signal-to-noise ratio (SNR) and Sample Entropy Ratio (SER).

The discrete wavelet packet signal analysis method simultaneously decomposes the signal into approximation coefficients and detail coefficients, denoted a and D, respectively. The method comprises the steps of firstly decomposing an original signal into a first-stage approximation coefficient A and a detail coefficient D, then continuously decomposing a first coefficient into coefficients AA and AD, respectively representing the approximation coefficient and the detail coefficient of a second stage, and decomposing D into DA and DD until the first-stage approximation coefficient and the detail coefficient D are maintained to a preset level. The total number of coefficients of the J-th decomposition level is equal to 2J (j=1, 2, …, J), where J is a preset decomposition level. For example, given a time series of N data points, a total of 8 coefficient sequences are generated after three-layer decomposition, and the coefficients resulting from the decomposition can reconstruct the time series:

mathematically, this method decomposes a time series into the composition of scaling coefficients and wavelet coefficients, the series expressed as follows:

orthogonal wavelet functions with compact local support characteristics are used for discrete wavelet packet multi-layered analysis, which can effectively represent detailed portions in the original signal. The double sum symbol in the equation represents the sum of coefficients of the scaling domain and the conversion domain, and the symbol Z represents the set of integer fields. In the j-th scaling factor s _j (k) And wavelet coefficient w _j (k) In an iterative manner, it is derived from the following formula:

where m is equal to 2k+n, h (), and h ₁ (.) are filter functions of scaling coefficients and wavelet coefficients, respectively, which are required to satisfy 2 conditions, namely, their integral equals zero and their integral of square equals 1, respectively. Taking the original data point as a coefficient c _j Is set to be a constant value. And obtaining the expected decomposition coefficient by an iterative recursion method.

To facilitate integration, let us assume the noise reduction coefficient d _jk Is the jth level k decomposition coefficient, and has the expression:

wherein the noise is an independent distribution with a mean of 0 and a variance of 1. Thus, raw data denoising becomes a multivariate non-parametric estimation problem, with the aim of obtaining denoising coefficients.

The invention uses Bayes hypothesis test to judge whether the decomposition coefficient needs the noise removed. The above formula can be expressed as the original decomposition coefficient d _jk Is a conditional probability distribution over the denoised coefficients and standard deviation, namely:

generally we have no early information about noise, assuming that the no a priori assumptions of coefficients are:

wherein Y is _jk Is a random binary variable.

The posterior distribution of the decomposition coefficients can be obtained according to the bayesian principle:

wherein the decomposition coefficients consist of two parts, namely a normal distribution of non-zero coefficients and a point distribution of zero coefficients, the above formula can be converted into the following form:

thus, the decomposition coefficients may be thresholded by a null hypothesis test, applying an index function to obtain the denoising coefficients.

S3, training and generating an OrbitNet model according to the extracted data of the vibration signals, and verifying the model. Specifically:

s31, dividing the axis track image into three groups of training samples, verification samples and test samples, wherein the proportion of the training samples, the verification samples and the test samples is 50%, 10% and 40% respectively;

s32, training samples are enhanced through rotation, translation, overturning and other methods, the diversity of training images is improved, and the deep learning model can be trained by using enough samples, so that the characteristics of the images are better captured for fault identification of the rotating machinery.

S33, automatically capturing the characteristics of the track image by using the CNN model, and constructing an orbitNet model, wherein the model considers the characteristics of the compressed image, has enough generalization and robustness, and improves the efficiency and the precision of the model while reducing the overfitting.

S34, adjusting parameters of an OrbitNet model by using training samples. The model parameter estimation adopts a cross-validation strategy and a class entropy loss function, and the training sample is divided into two parts: parameter tuning (80%) and performance monitoring (20%). The verification image set is used for verifying the prediction precision and error of the training model, and once the verification effect exceeds the training effect, the optimal model is selected for subsequent verification and comparison research.

S35, evaluating the model precision after training by adopting a verification sample which is not subjected to image enhancement;

s36, evaluating the recognition accuracy of the established model by adopting a test sample which is not subjected to image enhancement, wherein the test result can verify the universality and stability of the established model for fault recognition under different application scenes.

S37, adopting the same training, verifying and testing image samples, and comparing the calculation performance and the recognition accuracy of the three models under different maximum pooling layers. Through comparative research, the performance and accuracy of the proposed axial locus network model in the fault identification of different turbomachinery are further described.

Convolutional neural network (Convolutional Neural Network, CNN) models have been a powerful deep learning tool for pattern recognition and image classification over the past few years. The CNN model automatically extracts complex features from an image through a multi-layer convolution operation, reduces the resulting additional errors due to manual feature extraction, and is typically composed of three key layers: convolution layer, downsampling layer (pooling layer) and full connection layer. The convolution layer is responsible for extracting features from the image, the downsampling layer reduces feature dimensions by a fixed algorithm to obtain the most important features in the image, and the full-connection layer rearranges the identified features into a flat form for subsequent prediction or identification tasks. Typically, in creating a CNN deep learning model, a combination of multiple convolution layers and downsampling layers are used to improve its accuracy. Compared with the traditional fully-connected hidden layer in the multi-layer neural network model, the convolution layer has the characteristics of few parameters, connection sparsity, capture translation invariance and the like, and the characteristics can greatly reduce the calculation cost of parameter optimization and the possibility of overfitting and provide the CNN model with enough calculation precision.

In the example, a Convolutional Neural Network (CNN) deep learning model based on an axial locus diagram is provided, which is called orbitNet, and the model is used for fault diagnosis of turbomachinery based on a track image through a special design model structure. The model integrates stride into the convolutional layer to improve image compression travel while improving learning ability. The model integrates the concept of multi-step length in the convolutional layer itself with (or without) the maximum pooling layer, in hopes of improving the image compression efficiency and the learning ability of the network model. The model consists of 7 two-dimensional convolution layers, the filter size increasing from f=16 to f=128 in units of the power of 2, with layers 1, 5 and last being simultaneously convolved and downsampled in the x and y directions (step size 2). Therefore, in the case of using the pooling layer conventionally for the next layer, the calculation amount can be reduced by 4 times, and the core size of the convolution layer of the first layer is 5×5, the convolution core sizes of the other layers are 3×3, so that the structure maintains good learning ability while reducing the network volume, and the main characteristics of the model are described as follows:

1) Convolution layer: in convolutional neural network models, convolutional layers are used to extract features from images. Definition { x } _i, y _i (i=1, 2,., N) is a series of picture training data, where x _i Is a feature map of size h×w×d, h×w is the length and width of a picture, and D is the number of channels per picture. For example, d=1 is a gray scale map, and d=3 is a color RGB (red, green, blue) map. Variable y _i E (i=1, 2.,. C) refers to the corresponding fault type in the study, where C is the number of fault types. The purpose of training the CNN model is to learn the weights and bias of the filter, minimizing the classification errors of the output layer. Let a convolution layer contain k filters, then the convolution layer is denoted as:

where the symbols represent the convolution operators,

ith Zhang Tezheng drawing X corresponding to m layers _i Parameter W ^(m) And parameter b ^(m) The weights and biases are learned by the M-th layer network, M is the number of layers of the CNN model, and f is a nonlinear activation function. In this CNN model we use a ReLU activation function that has been demonstrated to provide the network with a strong ability to model non-linearity problems, setting negative values in the mapping to zero, thus effectively improving computational efficiency while maintaining sufficient modeling accuracy and having some effect on solving gradient vanishing problems in parametric training.

The step size parameter is used to control the distance the filter moves along the x-axis and y-axis of the input image during the convolution process. A larger step value will result in a smaller spatial dimension for the next layer output, but an excessive step value may result in a loss of initial features, and typically we can use a zero-fill strategy to maintain the same output size. The output of the previous convolutional layer in the CNN model serves as the input of the next convolutional layer. Given the output length or height of the convolutional layer, d0 can be calculated by the following formula:

d ₀ ＝(d _i -F+2P)/S+1

d in _i Is the length or height of the input picture, and the other integer parameters F, P and S are the size of the convolution kernel, the size of the zero padding and the step size, respectively.

2) Pooling layer: the largest pooling layer (or downsampling) is 2 in size on the x-axis and y-axis and can be used after convolution layers with convolution kernel sizes f=32 and f=64. The max pooling layer achieves feature downsampling by taking the maximum value in a certain region. The output of the convolution layer can be subjected to batch normalization processing, so that the average value of the output of the activation function is close to zero, the standard deviation is close to 1, the model training can be ensured to be stable, and the super-parameters can be adjusted more easily.

Furthermore, two dropout layers with a probability of p=0.25 are used in the model, and one dropout layer with a probability of p=0.5 is used at the end of the model. The dropout layer deletes partial branches by randomly setting partial weights in the model to zero, so that more generalized characteristics can be learned, the learning capacity of the network is enhanced, the possibility of overfitting is reduced, and the training time of each iteration step is shortened. Finally, we add L2 (sum of squares of all model weight parameters) regularization term to the objective function of all 7 convolutional layers, increasing the generalization of the model and reducing the likelihood of overfitting.

3) Full tie layer: extracting deeper, more abstract features at the end of the network by reshaping the feature map into an n-dimensional vector, defined as: a (a)

Wherein X is _c 、

W _c And b _c Respectively input, output, weight and deviation of the full-connection layer, LC is the number of the full-connection … … layers, g (-) is the activation function of the layers, and the study adopts a softmax function at the full-connection layer for downstream image classification tasks.

4) Class cross entropy: the present study uses multi-class cross entropy as a loss function to obtain optimal model parameters. Furthermore, we additionally consider L2 regularization in the loss function, which model is trained by minimizing the loss function, specifically expressed as:

wherein θ= { W ⁽¹⁾ ，...，W ^(M) ，b ⁽¹⁾ ，...，b ^(M) The parameter which is required to be trained in the OrbitNet CNN model is y _i,j Is X _i The corresponding j-th real label is g _j (θ,X _i ) For sample X _i λ is the penalty factor of the L2 regularization term.

5) Overfitting was avoided: a well-known disadvantage of deep learning models is that over-fitting tends to occur during training, resulting in a model that performs well in practice with lower accuracy. In addition to the dropout layer and regularization techniques described above, data enhancement and early-stop strategies are employed to reduce the possible overfitting. The data enhancement technology comprises rotation, turnover and translation of the track image, and the methods can enhance the diversity of the training data set, so that the generalization and the robustness of the model are improved. In the early-stop cross-validation strategy, training data is divided into two groups: parameter tuning and performance verification. In the training process, the model verification error is continuously monitored, and when the model verification error continuously becomes large, the training is stopped.

S4, fault identification is carried out on the large-scale turbine mechanical parts by using an OrbitNet model.

The method comprises the steps of firstly cleaning the multidimensional vibration signals by a Bayesian wavelet packet drying method, reducing uncertainty in original data, then constructing an axis track image by using cleaning signals, and improving the representativeness of the image. And a convolutional neural network CNN diagnosis model orbitNet special for the axis trajectory graph is designed, and the trained orbitNet model is utilized to perform fault identification on large-scale turbomachine components. The method carries out image enhancement on the obtained axis track, and further improves the representativeness of the image, thereby improving the general type and the robustness of the built model. The axis trace diagram effectively represents a failure mechanism of a typical failure of a turbomachine component, so that the method provided by the invention effectively combines data and the failure mechanism, and improves the accuracy and the universality of the method.

Embodiment 2. This patent also discloses an automatic intelligent diagnosis system for turbine mechanical faults, comprising: the sensor is used for collecting vibration signals in the x and y directions at the upper axis of the turbomachine; the cloud data platform is respectively in communication connection with the sensor and the edge computing monitoring diagnosis system; and the edge calculation monitoring and diagnosing system is used for carrying out data preprocessing and data cleaning on the vibration signals and carrying out fault identification on the large-scale turbomachinery parts by utilizing the OrbitNet model generated by training.

Example 3 through six actual turbomachine operating data and fault information collected from different customers, the validity and feasibility of the general fault diagnosis method and implementation flow of the chapter study was verified. As shown in fig. 3, there are 4 common types of failure for bearing components: oil film whirl, imbalance, rotor misalignment, and friction. In order to verify the universality of the method, fault data of three different types of turbomachinery such as a steam turbine, a centrifugal compressor and a generator are collected, and the turbomachinery is applied to different industrial fields such as coal chemical processing, coal manufacturing, chemical plants and the like. Taking SXLA centrifugal compressor as an example, fig. 4 shows the phenomenon of impeller cracking (fig. 4 a)) observed on a plurality of blades and the resulting shaft eccentricity (fig. 4 b). The multiple damaged impellers cause the shaft to shift from the original center (O-mark in fig. 4 b) to a new center (O-mark in fig. 4 b), thereby creating an imbalance in the compressor.

(1) Data preprocessing

As shown in fig. 5, the axial trace image used in this example is created from vibration time-series data acquired by four eddy current sensors. In the SHBT example, the time history data with oil vortex faults in the x-direction (fig. 5 a)) and y-direction (fig. 5 c)) are shown in fig. 5, which illustrates the normal operating conditions of the machine and the conditions when four oil vortex fault events occur. Fig. 5 also shows a box diagram of normal operating conditions and four fault conditions in the x-direction (fig. 5 b)) and the y-direction (fig. 5 d)). The box plot is a descriptive statistical method that contains five quantiles of the data, namely the maximum (upper quartile) and the minimum (lower quartile) (excluding any outliers), the first quartile (25 quartiles or Q1), the median (50 quartiles or Q2) and the third quartile (75 quartiles or Q3), with the distance between the upper and lower quartiles being the quartile range (Interquartile Range, IQR). As shown in FIGS. 5 b) and 5 c), points that exceed 1.5 XIQR are generally identified as outliers and marked with an asterisk in the figure.

Turbomachinery typically operates continuously at high speeds, e.g., SHBT speeds up to 11900r/min, with some data points shown in FIG. 5 for ease of illustration. Three conclusions were drawn by observing fig. 5: first, the time series in the x-direction (fig. 5 a)) has a similar trend as the time series in the y-direction (fig. 5 c); second, the amplitude in the x-direction (fig. 5 a)) is greater than the amplitude in the y-direction (fig. 5 c)); third, the median value of data in the oil vortex fault state is lower than that in the normal state.

(2) Data cleansing

Removing noise in each original time sequence by adopting multi-resolution Bayesian wavelet packet analysis, selecting Daubechies as a wavelet basis function during analysis, and carrying out discrete wavelet packet multi-resolution analysis. The wavelet function is selected because it has the characteristics of orthogonal and local compact amplification, can effectively characterize local details in the vibration signal, acts as a mathematical microscope, and identifies possible noise from the decomposed wavelet coefficients by bayesian hypothesis testing. On the basis of the given original signal, three layers of discrete wavelet packet transformation (j=3) is performed by using a DB8 function with 8 fluctuations, and the original signal is subjected to denoising processing, so as to generate 8 decomposition coefficient sequences, which respectively represent the details and approximations of the original signal. Then, a Bayesian factor is calculated, and noise reduction evaluation is performed on the decomposed 8 coefficients. Reconstructing the processed coefficient into a denoised signal, and drawing an axis locus diagram.

The raw data and the denoising data are further visually compared in the time domain and the frequency domain. Fig. 6 compares raw and denoised vibration data in the x-direction in the case where SXLA fails. Fig. 6 a) shows only the original time series and cannot intuitively distinguish the images before and after denoising, however, the normalized histogram (fig. 6 c)) shows a significant difference between the original (shaded) and denoised (solid line) time series in the two vibration ranges of-5 μm, -3 μm and 3.5 μm,5 μm. There is also a significant difference in the Welch functional spectral density PSD index of the denoised data (fig. 6 d)) versus the original data (fig. 6 b)). Therefore, the difference between the denoising signal and the original signal shows that the Bayesian wavelet denoising method adopted in the chapter provides an effective tool for denoising sensor data.

(3) Axis trace image rendering and processing

Each axial trace plot is plotted from 1204 data points (32 cycles of rotation) collected by sensors mounted in the x and y directions of the turbomachine. Under normal and each fault condition, approximately 150 axis trajectories are generated from the denoised data points, respectively. Fig. 7 shows a plot of the axial trace generated from a set of vibration time sequences acquired by the SHBT generator, which contains oil vortex failure. For comparison, two axis trajectories (fig. 7 b) and fig. 7 d)) are obtained from the original time sequence (fig. 7 a)) and the denoised (fig. 7 c)), respectively, and it can be seen from the figure that 2 improvement cases of the axis trajectories after denoised: in the time series shown in fig. 7 a) and 7 c), the difference between the noise reduced data and the original data cannot be intuitively observed; the improvement of the denoised data on the axial trace is significant. This again illustrates that the bayesian wavelet denoising method provides an effective tool for smoothing the axis trajectories and reducing noise in the raw data.

Further, fig. 8 shows axial trace diagrams of SXLA centrifugal compressor in normal state (fig. 8 b)) and unbalanced state (fig. 8 d)), respectively. The vibration data after denoising in the normal state (fig. 8 a)) and in the unbalanced state (fig. 8 c)) are used in the drawing, and it is apparent that the axial trace in the normal state is nearly circular (fig. 8 b)), and the axial trace in the unbalanced state is nearly elliptical (fig. 8 d)).

(4) Image grouping

And dividing the axle center track image generated by the denoised vibration signal into three groups of training, verification and test. The training and validation set was randomly selected from all 4 failure modes contained in the first 4 scenarios, while the test data was mainly from the remaining two scenarios, which also contained the unused axial trace pictures in the training and validation. Fig. 9 shows the number of axial trace patterns created from the denoised signal, which will then be used for OrbitNet modeling, where each bar contains the number of axial trace patterns for model training (bottom), validation (middle) and testing (top). Ideally we should provide almost the same number of images for all scenes. After data preprocessing, we generate 80 images from the normal running state of 6 customer scenes, 55 images from the fault state, and 100 images from the other 3 fault scenes, so that the built model can consider the common data unbalance phenomenon in practical application.

(5) Image preprocessing

The number of the existing images is still small, and the existing images are insufficient for building a general deep learning model. In order to improve the accuracy, effectiveness and efficiency of the model, gray processing is carried out on all generated images, the images are adjusted to be 128×128 (original 640×480) small size by using Lanczos resampling filter [153], then image enhancement processing such as centering, normalization, rotation, translation, overturning and the like is carried out on the training axis track images, and only centering and normalization processing is carried out on the verification and test images. Fig. 10 shows a gray scale axial trace plot of 4 failure modes that will be used for subsequent modeling and failure recognition.

Each set of axial trace images was centered around zero and normalized by its standard deviation. By rotating 217 training images by 20 degrees, both width and height are moved by 0.1, and horizontally and vertically flipped, a new training image is generated, resulting in an additional 1085 images (1302 samples total). It should be noted that unlike the training image samples, the images used for verification and testing are not enhanced for assessing the flexibility of the model in fault recognition.

(6) Construction of OrbitNet model

The convolution layer stride of the OrbitNet diagnostic model constructed by using the training sample is 2×2. The model does not use a maximum pooling layer, and the special design can control and reduce the mapping size from one convolution layer to the next, and the seven-layer CNN model in the example contains 361445 trainable parameters to be optimized.

In the embodiment, a cross-validation strategy is adopted, 1302 training images are utilized to train the OrbitNet deep learning model, a training image set is divided into two parts, 976 images (75%) are used for training model parameters, and 326 images (25%) are used for monitoring the prediction accuracy of the model. In this way, the resulting diagnostic model can avoid possible overfitting and have acceptable accuracy. Meanwhile, in parameter adjustment, the sample batch size (batch size) is set to 50 to improve training efficiency, and model parameters are updated 20 times in each iteration step, and when the detection precision is greater than the training precision, the model is selected to be saved. In this example, the iteration step is preset to 60 by a trial-and-error method to show the variation trend of the training performance. The trend of model training and checking accuracy and errors thereof are shown in fig. 11, from which it can be seen that model parameters reach a steady state at iteration 39, with a training accuracy of 99.67% and a checking accuracy of 99.69%.

(7) Model verification test and comparison

In this example, 43 axial trace images were used to evaluate the model that had been constructed and other 175 images were used to test the model. Fig. 12 shows the results of the OrbitNet model training, validation and testing, and it can be seen from fig. 12 that the fault recognition accuracy of the OrbitNet model is 100.0% in both of these scenarios. The verification and test results show that the proposed OrbitNet model has high recognition accuracy for common fault modes of different turbomachines.

To further verify the performance and accuracy of the proposed OrbitNet model, two additional structurally similar models were added for comparison with the SqueezeNet model using the same image samples. Unlike the OrbitNet model, which uses step-wise downsampling, the OrbitNet1 model is designed with one max_working_2d (MP 2D) layer for downsampling, whereas the OrbitNet2 model has two MP2D layers, resulting in OrbitNet and OrbitNet1 models containing 361445 parameters, slightly more than 360005 parameters of OrbitNet 2. The SquezeNet model used in the comparative study has a compressed structure with accuracy similar to AlexNet. The same image set and model construction scheme was used to train, verify and test the 4 models, and the comparison results are shown in fig. 12.

As can be seen from fig. 12, the training accuracy of the OrbitNet2 model is highest, 99.9%, and the loss is smallest, 0.008. However, compared with the other three models, the model needs the longest training time of 3.5 hours, and the training needs 52 iteration steps to obtain the optimal model, but the test precision is the lowest in all models and is 86.3%, which shows that the OrbitNet2 model with two MP2D layers can have fitting training. Overall, the first two stride models are superior to the OrbitNet2 model in both training efficiency and recognition accuracy. In addition, the SquezeNet model is most efficient on-line, with the shortest training time, on the order of 10 minutes, because its compressed architecture is such that only 1/9 of the model parameters need to be trained, while its training and validation accuracy is similar to that of OrbitNet1, however, it can be seen from FIG. 13 that its training process is unstable. Therefore, we recommend using the OrbitNet model for fault diagnosis analysis, because the training time of this model is reasonable, 3664 seconds, and the highest test accuracy of 100.0% is achieved.

FIG. 14 is a comparison of test accuracy results of 4 trained models for predicting different states of a turbomachine. It can be observed from fig. 14 that, first, the two models, orbitNet and OrbitNet1, are better able to identify all machine operating conditions than OrbitNet2, except for imbalance faults. Under the state that the machine rubs, the OrbitNet2 model can only keep 60% of recognition accuracy, but the performance of the OrbitNet2 model is slightly better than that of the OrbitNet1 model for rotor imbalance faults. And secondly, the identification precision of the OrbitNet model in all five modes is obviously superior to that of other models. Therefore, in this example, it is recommended to use the OrbitNet model based on the vibration data axis trace map for turbine mechanical state evaluation and fault diagnosis.

The foregoing is merely a preferred embodiment of the present invention and it should be noted that modifications and adaptations to those skilled in the art may be made without departing from the principles of the present invention, which are intended to be comprehended within the scope of the present invention.

Claims (10)

1. The universal method for automatically and intelligently diagnosing the turbine mechanical faults is characterized by comprising the following steps of:

s1, collecting vibration signals in the x and y directions at the axis through a sensor arranged on the turbomachine and storing the vibration signals;

s2, carrying out data preprocessing and data cleaning on the vibration signals, and coupling in the x-direction and the y-direction to generate an axis track image;

s3, training to generate an OrbitNet model according to the extracted data of the vibration signals and verifying the model;

s4, performing fault identification on the large-scale turbine mechanical parts by using the OrbitNet model.

2. The universal method for automated intelligent diagnosis of turbomachine fault according to claim 1, wherein in step S1, the vibration signal is stored in a local server or cloud data platform at a monitoring site.

3. The universal method for automated intelligent diagnosis of turbomachine fault according to claim 1, wherein "data preprocessing" in step S2 comprises: removing an abnormal value of the extracted vibration signal; and normalizing the data by adopting a minimum-maximum method, and marking and encoding the generated image by a binarization method to obtain multi-class output.

4. A universal method for automated intelligent diagnosis of turbomachine faults according to claim 1 or 3, wherein "data cleaning" in step S2 comprises: and cleaning the preprocessed vibration signals by adopting a Bayes wavelet denoising method.

5. The universal method for automated intelligent diagnosis of a turbomachine fault, according to claim 4, wherein the step S2 of "data cleaning" further comprises: and evaluating the denoising effect on the vibration signal after data cleaning.

6. The universal method for automated intelligent diagnosis of turbomachine fault, according to claim 5, wherein said evaluating denoising effect comprises: qualitatively evaluating the denoising effect through the time domain and frequency domain graphs; and quantitatively evaluating the denoising effect through the signal-to-noise ratio and the sample entropy ratio.

7. The universal method for automatically and intelligently diagnosing a turbomachine fault according to claim 1, wherein the axial locus image is compressed after the axial locus image is generated in the step S2.

8. The universal method for automated intelligent diagnosis of turbomachine fault in accordance with claim 1, wherein step S3 comprises: dividing the axis locus image into three groups of training samples, verification samples and test samples; utilizing a CNN model to automatically capture characteristics of a track image, and constructing an orbitNet model; adjusting the parameters of an OrbitNet model by adopting the training sample; evaluating the trained model precision by adopting the verification sample; and evaluating the recognition accuracy of the established model by adopting the test sample.

9. The universal method for automatic intelligent diagnosis of turbomachinery faults according to claim 8, wherein in step S3, the training samples are subjected to image enhancement through rotation, translation and overturning methods.

10. An automatic intelligent diagnosis system for turbine mechanical faults, comprising:

the sensor is used for collecting vibration signals in the x and y directions at the upper axis of the turbomachine;

the cloud data platform is respectively in communication connection with the sensor and the edge computing monitoring diagnosis system;

and the edge calculation monitoring and diagnosing system is used for carrying out data preprocessing and data cleaning on the vibration signals and carrying out fault identification on the large-scale turbomachinery parts by utilizing the OrbitNet model generated by training.

Images