CN112686181B - Hydraulic turbine fault diagnosis method based on interpolation axis track - Google Patents
Hydraulic turbine fault diagnosis method based on interpolation axis track Download PDFInfo
- Publication number
- CN112686181B CN112686181B CN202110003573.XA CN202110003573A CN112686181B CN 112686181 B CN112686181 B CN 112686181B CN 202110003573 A CN202110003573 A CN 202110003573A CN 112686181 B CN112686181 B CN 112686181B
- Authority
- CN
- China
- Prior art keywords
- layer
- interpolation
- axis track
- image
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/20—Hydro energy
Landscapes
- Image Analysis (AREA)
Abstract
The hydraulic turbine fault diagnosis method based on interpolation axis track includes the steps of firstly, performing change and interpolation on a unit shafting monitoring signal, and obtaining an interpolation signal according to extremum position distribution and sine function relation; an interpolation axis track is provided on the basis of the interpolation axis track; a corresponding convolutional neural network model is established to realize intelligent monitoring of the axial locus images; the method eliminates the huge differences among individuals of the traditional axis track, compresses signal data quantity, establishes a complete axis track standardization method, realizes the characteristic learning and modeling method of track information, can identify faults such as rotor misalignment, rotor imbalance, rotor bending and the like in the water turbine unit, and can be applied to the state monitoring and fault diagnosis tasks of various water turbines.
Description
Technical Field
The invention belongs to the technical field of hydroelectric generating set state monitoring and fault diagnosis, and particularly relates to a hydraulic turbine fault diagnosis method based on an interpolation axis track.
Background
The water turbine is used as core equipment of an energy conversion and power grid regulator, and the safety maintenance task is hard: on one hand, the water turbine is developed towards the complicated and large-scale direction, and the data acquisition capacity of the water turbine is greatly improved along with the continuous upgrading of a unit monitoring system, so that an advanced intelligent monitoring method is required; on the other hand, a large number of units with longer continuous service years are in a complex water conservancy electromagnetic coupling environment for a long time, so that equipment risks are continuously increased. Once a water turbine fails, serious consequences are caused, so that early abnormality of a unit is required to be discovered as soon as possible in daily operation and maintenance, and effective preventive measures are taken before the failure occurs.
In various monitoring methods, vibration and waviness are common monitoring information sources, and axle center tracks are common usage forms. The axis track is an image information synthesized by orthogonally mounting two displacement sensors on the section of a guide bearing of the hydroelectric generating set to acquire a set swing degree signal and further filtering and noise reduction. The axial locus can intuitively and comprehensively reflect the running condition of the unit, becomes one of important means for judging the equipment state by workers, can reflect the running state and fault condition of the main shaft of the water turbine, can also show abnormal hydraulic vibration when the unit generates the phenomena of low-frequency vortex strips, karman vortex strips, she Daoguo and the like of the tail water pipe, and has important application value in monitoring and maintaining the daily running of the water turbine generator.
In practical application, the axial locus is often required to be manually observed to judge the running state of the unit, and the mode mainly has the following defects: firstly, long-term continuous monitoring cannot be realized manually, so that the application range and the field of the axis track are greatly limited, and therefore, the judgment is still carried out on some sites by the amplitude of the original signal; secondly, the manual observation has strong subjective dependence, and people are required to have abundant experience knowledge; meanwhile, the track information has differences among units with different specifications, the standardization and normalization lack of corresponding methods, the universality of the axle center track is limited, and the local interferences such as various noises, overlarge amplitude values, abnormal data and the like are eliminated through reasonable signal reconstruction so as to keep the overall trend and main components of the signals; in addition, the track information contains rich equipment state content, has great blindness and subjectivity for the construction calculation and method selection of the information feature set, and lacks an intelligent decision model and a corresponding data processing method.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide a water turbine fault diagnosis method based on an interpolation axis track, which is mainly applied to a water turbine shafting monitoring signal and mainly aims to generate an interpolation axis track and realize water turbine fault diagnosis on the basis of the interpolation axis track. The technical scheme of the invention is as follows:
1) Interpolation is carried out on signals forming the axis track of the water turbine to obtain interpolation signals, the interpolation method takes the local extremum of the signals as a reference, interpolation is carried out according to a standard sine function, and the calculation method has universal adaptability and universality;
2) On the basis of interpolation signals, an interpolation axis track is established under a plane rectangular coordinate system, and various interpolation axis track samples under the normal running state and different fault states of the water turbine are required to be obtained;
3) Preprocessing an axis track image, converting the axis track image into a gray level image, performing gray level compression, and performing size adjustment on the gray level image to obtain a preprocessed image, so that the uniformity of data specifications during training of a convolutional neural network model is ensured;
4) Constructing a convolutional neural network model, comprising two convolutional layers, two pooling layers, a flat layer and a full-connection layer, and inputting preprocessed images of various hydraulic turbine fault modes and normal modes into the model for network training;
5) After training, the convolutional neural network model is used for intelligent monitoring and fault diagnosis of the water turbine, unknown monitoring signals are converted into preprocessed images, the trained convolutional neural network model is input for water turbine state identification, and an identification result is output.
A hydraulic turbine fault diagnosis method based on interpolation axis track comprises the following steps:
if two adjacent extrema are 1 and-1 in turn, the value of each non-extremum point is calculated according to equation (1):
if two adjacent extrema are-1 and 1 in turn, the value of each non-extremum point is calculated according to equation (2):
in the above formula, N represents the number of non-extremum points between two extremum points on the interpolated section, N represents the nth non-extremum point, and N is visible to be N at maximum, so as to obtain an interpolation signal;
step 4, obtaining interpolation axis tracks of the water turbine in normal running states and different fault states, wherein the fault modes comprise: rotor misalignment, rotor imbalance, rotor bending, excessive abnormal vibration of the rotor when draft tube low frequency vortex, karman vortex, she Daoguo phenomena occur;
step 5, segmenting the axial track image, setting an origin (0, 0) to be positioned at the right center of the axial track image by taking a plane rectangular coordinate system as a reference, and drawing four range straight lines as shown in (3):
y＝-1,y＝1,x＝-1,x＝1 (3)
according to the step 1 and the step 2, the axis track does not exceed the area surrounded by the four range straight lines, the four range straight lines are used as the boundary of the interpolation axis track, and then the interpolation axis track is preprocessed and converted into a preprocessed image;
wherein, represents convolution operation, M j Representing a set of preprocessed images, l representing a layer i network, k representing a convolution kernel parameter, i representing an ith picture, j representing a j-th convolution kernel,
represents the network offset of the jth convolution kernel at layer l, x j l Represents the j-th convolution kernel output at layer l, x j l-1 Indicating that the j-th convolution kernel is input at the layer l, and f (·) is an activation function;
the calculation method of the pooling layer neurons is shown as a formula (5):
x j l ＝f(β j l down(x j l-1 )+b j l ) (5)
where down (·) represents the downsampling function, l represents the layer i network, j represents the j-th convolution kernel,
representing the network bias of the jth convolution kernel at layer lPut (I) at>
Represents the network offset of the jth convolution kernel at layer l, x j l The j-th convolution kernel is represented to be output in the layer l, and the full connection layer and the output layer perform classification processing on the samples, as shown in the formula (4):
y k ＝f(w k x k-1 +b k ) (6)
wherein: k represents the sequence number of the network layer, y k Representing the output of the fully connected layer, x k-1 Is the characteristic vector, w k Is a weight coefficient, b k Representing a bias term;
step 7, after the convolutional neural network model is trained, the convolutional neural network model is applied to monitoring the state of the unit; when the convolutional neural network model is used, the unknown monitoring signals newly acquired by the unit are repeatedly subjected to the steps 1-5 to acquire corresponding interpolation axis tracks, the corresponding interpolation axis tracks are input into the convolutional neural network model, and the state of the unit is judged according to the identification result.
And (3) repeating the steps 1-7 for building the corresponding convolutional neural network model for the water turbines with different models and types.
When the convolutional neural network model is established, the training samples adopted comprise a preprocessed image of a fault sample to be detected and a preprocessed image of a normal running state, and the number of interpolation axis trajectories acquired from each state is more than or equal to 120.
The interpolation axis track described in the step 3 needs to be preprocessed before being used for training the convolutional neural network model, and is converted into an 8-bit gray level image, wherein the image size is 64×64.
6.1 The size of the input image is 64 multiplied by 64, and the filling mode is no filling%valid) The step length is 1, the first layer convolution kernel size is 3 multiplied by 3, the channel number is 32, so that 32 characteristic images are obtained, and the characteristic image output size is (62, 62)
6.2 62 x 62 for the first pooled layer input image, filling in such a way that no filling (valid) is performed, maximum downsampling is performed for the non-overlapping region, so the step size is 3, the feature image output size is (21, 21),
6.3 The second convolution layer input image size is 21 x 21, the convolution kernel size is 3 x 3, the channel number is 64, so 64 feature images are obtained, the feature image sizes are (19, 19),
6.4 19×19 of the second-layer pooled layer input image, filling in such a manner that no filling (valid) is performed, and the non-overlapping region is downsampled maximally, so that the step size is 3, and the feature map output size is (6, 6)
6.5 Flat layer) to obtain one-dimensional output vector, 2304 neurons (6×6×64=2304).
The beneficial effects of the invention are as follows:
compared with the prior art, in order to fully play the role of the axis track of the water turbine in equipment fault diagnosis, the invention takes the axis track of the water turbine as an object, and mainly realizes the following purposes: the standardization and normalization of the axis tracks of the water turbines of different models are realized, so that the interference of factors among devices on track analysis is eliminated; the convolutional neural network is utilized to analyze the axis track information, so that the independent feature extraction of the axis track of the water turbine under different fault modes is realized, and the manual feature calculation is not required; on the basis of deep learning, an intelligent fault diagnosis method is realized, the defect of subjective experience of artificial observation is overcome, and the method has stronger reliability and higher efficiency.
The method has the specific advantages that:
1) The interpolation axis track provided by the invention can solve the problem of the difference between the numerical value and the range of the axis tracks of different water turbines, so that all the axis tracks can be compared in the same range and scale; 2) The method utilizes the strong data mining capability of the deep learning model to acquire the abstract features of the interpolation axis track, so that the problem of feature set construction in the traditional mode identification is solved; 3) The method provided by the invention is easy to use in the existing hydropower station, has the same signal measurement mode as the existing axle center track monitoring device, does not need to add a new sensor and a testing system, and has strong feasibility; 4) The method can be used for analyzing actual monitoring signals, and can also be used for monitoring and diagnosing simulation calculation, analog signals and the like of various water turbines.
Drawings
Fig. 1 is a flow chart of the present invention.
Fig. 2 is a calculation example of interpolation calculation of the present invention.
FIG. 3 is a schematic diagram of the convolutional neural network structure of the present invention.
Detailed Description
The invention is described in further detail below with reference to the drawings and examples.
Referring to fig. 1, a method for diagnosing a water turbine fault based on an interpolated axial track is characterized by comprising the following steps:
if two adjacent extrema are 1 and-1 in turn, the value of each non-extremum point is calculated according to equation (1):
if two adjacent extrema are-1 and 1 in turn, the value of each non-extremum point is calculated according to equation (2):
in the above formula, N represents the number of non-extremum points between two extremum points on the interpolated section, N represents the nth non-extremum point, and N is visible to be N at maximum, so as to obtain an interpolation signal;
step 4, obtaining interpolation axis tracks of the water turbine in normal running states and different fault states, wherein the fault modes comprise: rotor misalignment, rotor imbalance, rotor bending, excessive abnormal vibration of the rotor when draft tube low frequency vortex, karman vortex, she Daoguo phenomena occur;
step 5, segmenting the axial track image, setting an origin (0, 0) to be positioned at the right center of the axial track image by taking a plane rectangular coordinate system as a reference, and drawing four range straight lines as shown in (3):
y＝-1,y＝1,x＝-1,x＝1 (3)
according to the step 1 and the step 2, the axis track does not exceed the area surrounded by the four range straight lines, the four range straight lines are used as the boundary of the interpolation axis track, and then the interpolation axis track is preprocessed and converted into a preprocessed image;
wherein, represents convolution operation, M j Representing a set of preprocessed images, l representing a layer i network, k representing a network parameter, i representing an ith picture, j representing a jth convolution kernel,
represents the network offset of the jth convolution kernel at layer l, x j l Represents the j-th convolution kernel output at layer l, x j l-1 Indicating that the j-th convolution kernel is input at the layer l, and f (·) is an activation function;
the calculation method of the pooling layer neurons is shown as a formula (5):
x j l ＝f(β j l down(x j l-1 )+b j l ) (5)
where down (·) represents the downsampling function, l represents the layer i network, j represents the j-th convolution kernel,
represents the network offset of the jth convolution kernel at layer l,>
represents the jth convolutionNetwork bias of core at layer l, x j l The j-th convolution kernel is represented to be output in the layer l, and the full connection layer and the output layer perform classification processing on the samples, as shown in the formula (4):
y k ＝f(w k x k-1 +b k ) (6)
wherein: k represents the sequence number of the network layer, y k Representing the output of the fully connected layer, x k-1 Is the characteristic vector, w k Is a weight coefficient, b k Representing a bias term;
step 7, after the convolutional neural network model is trained, the convolutional neural network model is applied to monitoring the state of the unit; when the convolutional neural network model is used, the unknown monitoring signals newly acquired by the unit are repeatedly subjected to the steps 1-5 to acquire corresponding interpolation axis tracks, the corresponding interpolation axis tracks are input into the convolutional neural network model, and the state of the unit is judged according to the identification result.
And (3) repeating the steps 1-7 for building the corresponding convolutional neural network model for the water turbines with different models and types.
When the convolutional neural network model is established, the training samples adopted comprise a preprocessed image of a fault sample to be detected and a preprocessed image of a normal running state, and the number of interpolation axis trajectories acquired from each state is more than or equal to 120.
The interpolation axis track described in the step 3 needs to be preprocessed before being used for training the convolutional neural network model, and is converted into an 8-bit gray level image, wherein the image size is 64×64.
6.1 The size of the input image is 64 multiplied by 64, and the filling mode is no filling%valid) The step length is 1, the first layer convolution kernel size is 3 multiplied by 3, the channel number is 32, so that 32 characteristic images are obtained, and the characteristic image output size is (62, 62)
6.2 62 x 62 for the first pooled layer input image, filling in such a way that no filling (valid) is performed, maximum downsampling is performed for the non-overlapping region, so the step size is 3, the feature image output size is (21, 21),
6.3 The second convolution layer input image size is 21 x 21, the convolution kernel size is 3 x 3, the channel number is 64, so 64 feature images are obtained, the feature image sizes are (19, 19),
6.4 19×19 of the second-layer pooled layer input image, filling in such a manner that no filling (valid) is performed, and the non-overlapping region is downsampled maximally, so that the step size is 3, and the feature map output size is (6, 6)
6.5 Flat layer) to obtain one-dimensional output vector, 2304 neurons (6×6×64=2304).
Referring to fig. 1, a known sample signal 1 is collected, a point on a main shaft of a water turbine is selected as a test point, a pair of vibration signals or swing degree signals of the point are obtained, the directions of the obtained pair of signals belong to the radial plane on the shaft, the pair of signals are mutually perpendicular, and the test point preferentially selects the upper guide, the lower guide and the water guide positions;
referring to fig. 1, a known sample signal 1f (t) is fourier transformed into a frequency domain by using formula (7), and a butterworth low-pass filtering method (formula (8)) is used to filter out high-frequency parts thereof and remove noise, so that the signal is smoothed; then the signal is converted into a space domain through inverse Fourier transform (formula (9)), and noise reduction of the signal is completed:
G(u)＝H(u)*F(u) (8)
wherein t represents a time continuous variable, u represents a frequency variable, j represents a virtual unit
F (u) is the Fourier spectrum of the signal with noise, G (u) is the Fourier spectrum of the signal after smoothing, H (u) is the transfer function of the filter, and the specific expression is formula (10):
D 0 is the cut-off frequency, n is the filter order, D (u) is the magnitude of the frequency u, used to control the shape of the transfer curve, and the known sample signal is filtered and then referred to as the filtered signal;
referring to fig. 1 and 2, the part of the known sample signal before the first extreme point and the part after the last extreme point are truncated, and then the two sets of signal extreme values are replaced, all maxima are replaced with 1, and all minima are replaced with-1; keeping the point values with the values of-1 and 1 unchanged, interpolating and resampling the values of all other signal points, and carrying out piecewise interpolation on non-extremum points between two adjacent extremum values, wherein for each pair of adjacent extremum values, the interpolation rule is as follows: if two adjacent extrema are 1 and-1 in turn, each point value is calculated according to equation (1)
If two adjacent extrema are-1 and 1 in turn, each point value is calculated according to equation (2)
In the above formula, N represents the number of non-extremum points between two extremum points on the interpolated section, N represents the nth non-extremum point, and N is visible to be N at maximum, so as to obtain an interpolation signal; if cosine function interpolation is adopted, the phases in the formula (7) and the formula (8) can be adjusted, and the reference of 1 and-1 is still needed; fig. 2 illustrates an interpolation calculation process of two sets of signals, in which the 1 st line and the 3 rd line are filtered signals in the X and Y directions, and the 2 nd line and the 4 th line are interpolation signals corresponding to each other.
Referring to fig. 1, an interpolation axis locus 4 is plotted from a pair of interpolation signals 3, an origin (0, 0) is set as a center position with reference to a planar rectangular coordinate system, and 4 range straight lines are plotted: y= -1, y= 1, x= -1, x=1, the 4 straight lines are used as boundaries of interpolation axis tracks and are converted into 8bit gray scale images, the image size is 64×64, which is called a preprocessing image, and the preprocessing image is an information input source of a convolutional neural network model.
Referring to fig. 1, a failure sample and a normal sample are collected to train an identification model, and the failure mode mainly includes: rotor misalignment, rotor imbalance, rotor bending, excessive abnormal vibration of the rotor when phenomena such as draft tube low frequency vortex band, karman vortex train, she Daoguo and the like occur; the selection of the fault sample is determined according to the actual demand of the unit, the fault sample can be one or more of the above ranges, the convolutional neural network model after training can monitor unknown monitoring signals and judge the running state of the unit, and the running state of the unit comprises a fault state and a normal state.
Referring to fig. 1 and 3, the convolutional neural network model is composed of a convolutional layer, a pooling layer, a flattening layer, a full connection layer and an output layer, wherein the convolutional layer and the pooling layer are two; the convolution kernel of the convolution layer is convolved with the input image to extract fault characteristics, and the output sizes of the characteristic diagrams of the two layers are 62 multiplied by 62 and 19 multiplied by 19 respectively; the pooling layer performs scaling mapping on the extracted features, and the output sizes of the feature images are 21×21 and 6×6 respectively; the flattening layer is arranged after the last pooling layer, and is mainly used for unidimensionally unifying multi-dimensional input and comprises 2304 neurons, and is commonly used for transition from a convolution layer to a full-connection layer; 256 neurons are arranged on the full-connection layer and are used for integrating all characteristic information of the convolution layer and the pooling layer; the output layer classification category is determined according to the actual situation, and the number of the fault categories is added with the normal state.
Referring to fig. 1 and 3, in the convolutional neural network, the output of the upper layer is the input of the lower layer, and the implementation process is as follows: the input image size is 64×64, the filling mode is no filling (valid), the step size is 1, the first layer convolution kernel size is 3×3, the channel number is 32, so that 32 characteristic images are obtained, and the characteristic image output size is (62, 62)
The first layer of the pooled layer input image is 62 x 62, the filling mode is valid (no filling), the maximum downsampling is carried out on the non-overlapped area, so the step size is 3, the characteristic image output size is (21, 21),
the second convolution layer input image has a size of 21×21, a convolution kernel has a size of 3×3, and a channel number of 64, so that 64 feature images are obtained, the feature images have sizes (19, 19),
the second layer of pooling layer input image is 19×19, the filling mode is valid (no filling), and the non-overlapping area is subjected to maximum downsampling, so that the step size is 3, and the characteristic image output size is (6, 6)
The flattened layer obtains a one-dimensional output vector, obtaining 2304 neurons (6×6×64=2304); after the model establishment is completed, the unknown monitoring signals are converted into preprocessed images, and then the preprocessed images are input into a convolutional neural network model to judge the running state of the unit.
Claims (5)
1. A hydraulic turbine fault diagnosis method based on interpolation axis track is characterized by comprising the following steps:
step 1, acquiring a pair of mutually perpendicular swing degree or vibration signals at the section of a shaft system of a hydroelectric generating set, wherein the section positions are near the upper guide, the lower guide and the water guide positions, and filtering and denoising by using a Butterworth low-pass filter after acquiring the mutually perpendicular swing degree signals of a pair of test points to obtain a filtering signal; the extremum of the two groups of filtering signals is replaced, all the maximum values are replaced by 1, and all the minimum values are replaced by-1;
step 2, keeping the points with the values of-1 and 1 unchanged, interpolating and resampling the values of all the rest swing degree signal points, and carrying out piecewise interpolation on the non-extremum points between two adjacent extremums, wherein for each pair of adjacent extremums, the interpolation rule is as follows:
if two adjacent extrema are 1 and-1 in turn, the value of each non-extremum point is calculated according to equation (1):
if two adjacent extrema are-1 and 1 in turn, the value of each non-extremum point is calculated according to equation (2):
in the above formula, N represents the number of non-extremum points between two extremum points on the interpolated section, N represents the nth non-extremum point, and N is visible to be N at maximum, so as to obtain an interpolation signal;
step 3, carrying out axis track drawing on the pair of interpolation signals which finish the step 2, sequentially selecting points of the interpolation signals in two directions to form an abscissa, and carrying out in a plane rectangular coordinate system, wherein the axis track abscissa drawn is between-1 and 1, and obtaining an axis track image of the interpolation signals, which is called an interpolation axis track;
step 4, obtaining interpolation axis tracks of the water turbine in normal running states and different fault states, wherein the fault modes comprise: rotor misalignment, rotor imbalance, rotor bending, excessive abnormal vibration of the rotor when draft tube low frequency vortex, karman vortex, she Daoguo phenomena occur;
step 5, segmenting the axial track image, setting an origin (0, 0) to be positioned at the right center of the axial track image by taking a plane rectangular coordinate system as a reference, and drawing four range straight lines as shown in (3):
y＝-1,y＝1,x＝-1,x＝1 (3)
according to the step 1 and the step 2, the axis track does not exceed the area surrounded by the four range straight lines, the four range straight lines are used as the boundary of the interpolation axis track, and then preprocessing is carried out to convert the interpolation axis track into a preprocessed image;
step 6, a convolutional neural network model is established by utilizing a preprocessed image of the interpolation axis track, wherein the convolutional neural network model comprises 2 convolutional layers, 2 pooling layers and 1 full-connection layer, and the convolutional layer model is shown as (4):
wherein, represents convolution operation, M j Represents the set of preprocessed images, l represents the first layer network, k represents the convolution kernel weight, i represents the ith picture, j represents the jth convolution kernel, b j l Represents the network offset of the jth convolution kernel at layer l, x j l Represents the j-th convolution kernel output at layer l, x j l-1 Representing the jth rollThe product is input in the layer I, and f (·) is an activation function;
the calculation method of the pooling layer neurons is shown as a formula (5):
x j l ＝f(β j l down(x j l-1 )+b j l ) (5)
where Down (·) represents the downsampling function, l represents the first layer network, j represents the j-th convolution kernel, β j l Representing the multiplicative offset of the jth convolution kernel at layer l, b j l Represents the network offset of the jth convolution kernel at layer l, x j l The j-th convolution kernel is represented to be output in the layer l, and the full connection layer and the output layer perform classification processing on the samples, as shown in a formula (6):
y k ＝f(w k x k-1 +b k ) (6)
wherein: k represents the sequence number of the network layer, y k Representing the output of the fully connected layer, x k-1 Is the characteristic vector, w k Is a weight coefficient, b k Representing a bias term;
step 7, after the convolutional neural network model is trained, the convolutional neural network model is applied to monitoring the state of the unit; when the convolutional neural network model is used, the unknown monitoring signals newly acquired by the unit are repeatedly subjected to the steps 1-5 to acquire corresponding interpolation axis tracks, the corresponding interpolation axis tracks are input into the convolutional neural network model, and the state of the unit is judged according to the identification result.
2. The method for diagnosing a water turbine fault based on an interpolated axial center trajectory according to claim 1, wherein the steps 1 to 7 are repeated for different types of water turbines to build the corresponding convolutional neural network model.
3. The method for diagnosing a water turbine fault based on an interpolated axial trajectory according to claim 1, wherein the training samples used in the construction of the convolutional neural network model include a preprocessed image of the fault sample to be detected and a preprocessed image of a normal operation state, and the number of interpolated axial trajectories obtained from each state is greater than or equal to 120.
4. The method for diagnosing a water turbine fault based on an interpolated axial trajectory according to claim 1, wherein the interpolated axial trajectory in step 3 is subjected to preprocessing before being used for training a convolutional neural network model, and is converted into an 8-bit gray-scale image with an image size of 64×64.
5. The method for diagnosing a water turbine fault based on an interpolated axial trajectory according to claim 1, wherein in the step 6, a convolution kernel of a convolution layer is convolved with an input image to extract fault features, and the output sizes of feature maps of two layers are 62×62 and 19×19 respectively; the pooling layer performs scaling mapping on the extracted features, and the output sizes of the feature images are 21×21 and 6×6 respectively; the flattening layer is arranged after the last pooling layer, and is mainly used for unidimensionally unifying multi-dimensional input and comprises 2304 neurons, and is commonly used for transition from a convolution layer to a full-connection layer; 256 neurons are arranged on the full-connection layer and are used for integrating all characteristic information of the convolution layer and the pooling layer, and the classification types of the output layer are determined according to actual conditions; in the convolutional neural network, the output of the upper layer is the input of the lower layer, and the specific implementation process is as follows:
6.1 The input image size is 64 multiplied by 64, the filling mode is no filling, the step length is 1, the first layer convolution kernel size is 3 multiplied by 3, the channel number is 32, so that 32 characteristic images are obtained, the output size of the characteristic images is (62, 62), and the specific calculation formula is that
6.2 62×62 for the first pooled layer input image, filling in a non-filled manner, and maximum downsampling for the non-overlapping region, so that the step size is 3, the feature image output size is (21, 21), and the specific calculation formula is
6.3 The second convolution layer input image has a size of 21×21, a convolution kernel has a size of 3×3, and a channel number of 64, so that 64 feature images are obtained, the feature images have a size of (19, 19), and a specific calculation formula is that
6.4 19×19 of the second-layer pooled layer input image, filling in a non-filling manner, and performing maximum downsampling on the non-overlapping region, so that the step size is 3, the output size of the feature map is (6, 6), and the specific calculation formula is
6.5 Flat layer) to obtain one-dimensional output vector, and 2304 neurons are obtained, wherein the specific formula of the number of the neurons is 6×6×64=2304.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110003573.XA CN112686181B (en) | 2021-01-04 | 2021-01-04 | Hydraulic turbine fault diagnosis method based on interpolation axis track |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110003573.XA CN112686181B (en) | 2021-01-04 | 2021-01-04 | Hydraulic turbine fault diagnosis method based on interpolation axis track |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112686181A CN112686181A (en) | 2021-04-20 |
| CN112686181B true CN112686181B (en) | 2023-06-23 |
Family
ID=75457101
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110003573.XA Active CN112686181B (en) | 2021-01-04 | 2021-01-04 | Hydraulic turbine fault diagnosis method based on interpolation axis track |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112686181B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113565666B (en) * | 2021-07-05 | 2024-05-14 | 江西洪屏抽水蓄能有限公司 | Method for judging running state of unit based on vibration data characteristics of water turbine generator unit |
| CN113657156B (en) * | 2021-07-13 | 2024-03-08 | 佛山市富丽图数字印刷有限公司 | Double-axis-track-based roll-to-roll printing unit fault diagnosis method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101667307A (en) * | 2009-09-10 | 2010-03-10 | 南京大学 | Short-time shaft centerline orbit array |
| CN102663261A (en) * | 2012-04-27 | 2012-09-12 | 长安大学 | Method for extracting rotating machinery rotor shaft centerline orbit by using time-frequency slice technology |
| CN104155542A (en) * | 2014-07-15 | 2014-11-19 | 中国船舶重工集团公司第七一九研究所 | Detection method suitable for flickering generated by high-frequency inter-harmonics |
| CN104165686A (en) * | 2014-06-17 | 2014-11-26 | 中州大学 | Rotor axis track purification method based on binary empirical mode decomposition |
| CN104400560A (en) * | 2014-11-07 | 2015-03-11 | 西安交通大学 | On-line measurement method for axis orbit of main shaft under cutting condition of numerical control lathe |
| CN108827454A (en) * | 2018-04-25 | 2018-11-16 | 哈尔滨电气股份有限公司 | A kind of turbine shafting vibrating data collection and processing method |
-
2021
- 2021-01-04 CN CN202110003573.XA patent/CN112686181B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101667307A (en) * | 2009-09-10 | 2010-03-10 | 南京大学 | Short-time shaft centerline orbit array |
| CN102663261A (en) * | 2012-04-27 | 2012-09-12 | 长安大学 | Method for extracting rotating machinery rotor shaft centerline orbit by using time-frequency slice technology |
| CN104165686A (en) * | 2014-06-17 | 2014-11-26 | 中州大学 | Rotor axis track purification method based on binary empirical mode decomposition |
| CN104155542A (en) * | 2014-07-15 | 2014-11-19 | 中国船舶重工集团公司第七一九研究所 | Detection method suitable for flickering generated by high-frequency inter-harmonics |
| CN104400560A (en) * | 2014-11-07 | 2015-03-11 | 西安交通大学 | On-line measurement method for axis orbit of main shaft under cutting condition of numerical control lathe |
| CN108827454A (en) * | 2018-04-25 | 2018-11-16 | 哈尔滨电气股份有限公司 | A kind of turbine shafting vibrating data collection and processing method |
Non-Patent Citations (3)
| Title |
|---|
| Towards optimal system scheduling with synthetic inertia provision from wind turbines;Zhongda Chu;《 IEEE Transactions on Power Systems》;20200414;4056-4066页 * |
| 基于EEMD的水轮机摆度信号提纯及特征提取分析;汪泉;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20190415;C037-126 * |
| 基于变分模态分解和多尺度排列墒的滚动轴承故障诊断;张建财;《噪声与振动控制》;20191231;第39卷(第6期);181-186页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112686181A (en) | 2021-04-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112686181B (en) | Hydraulic turbine fault diagnosis method based on interpolation axis track | |
| CN111947928B (en) | Multi-source information fusion bearing fault prediction system and method | |
| CN111914883B (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
| CN108647786B (en) | Rotary machine on-line fault monitoring method based on deep convolution countermeasure neural network | |
| CN109782603A (en) | The detection method and monitoring system of rotating machinery coupling fault | |
| Chen et al. | Fault identification of gearbox degradation with optimized wavelet neural network | |
| CN112665852B (en) | Variable working condition planetary gearbox fault diagnosis method and device based on deep learning | |
| CN110991295A (en) | Self-adaptive fault diagnosis method based on one-dimensional convolutional neural network | |
| CN107506710A (en) | A kind of rolling bearing combined failure extracting method | |
| CN112305388B (en) | On-line monitoring and diagnosing method for insulation partial discharge faults of generator stator winding | |
| CN110056640B (en) | Speed reducer wireless fault diagnosis method based on acceleration signal and edge calculation | |
| CN114755017B (en) | Variable-speed bearing fault diagnosis method of cross-domain data driving unsupervised field shared network | |
| CN112324745A (en) | Portable hydraulic system leakage signal acquisition device and acquisition method thereof | |
| CN114034481A (en) | Fault diagnosis system and method for rolling mill gearbox | |
| CN108181059B (en) | Multi-phase flow pipeline leakage sound wave signal identification method based on wavelet signals | |
| CN112733612A (en) | Cross-domain rotating machinery fault diagnosis model establishing method and application thereof | |
| CN212363649U (en) | System for realizing fault diagnosis method of gearbox | |
| CN113408578A (en) | Harmonic reducer fault diagnosis method and system based on generative countermeasure network | |
| CN112801033A (en) | AlexNet network-based construction disturbance and leakage identification method along long oil and gas pipeline | |
| CN116738314A (en) | Air compressor fault detection method based on neural network | |
| CN112257616A (en) | Phase modulator fault diagnosis method, device and system based on vibration signals | |
| CN112734001A (en) | Wind power transmission chain intelligent fault diagnosis method based on order spectrum migration | |
| CN113486599B (en) | Method for calculating effective stroke of oil pumping unit | |
| CN109916624A (en) | A kind of ball screw assembly, fatigue failure diagnostic method based on Hilbert Huang | |
| CN113447267B (en) | Vibration signal analysis-based complete gearbox state evaluation method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |