1. A plunger pump fault signal time-frequency graph noise reduction enhancing method is characterized by comprising the following steps:

step 1: acquiring vibration signals of the plunger pump under different fault levels by a triaxial vibration acceleration sensor arranged on a plunger pump shell;

step 2: slicing and segmenting the acquired vibration signal according to a preset size;

and step 3: carrying out time-frequency transformation on the sliced sample to obtain a time-frequency graph of the sample, and dividing all the time-frequency graphs into a training set and a testing set according to a preset proportion;

and 4, step 4: training a convolutional neural network classification model by using training set data, and performing fault classification on a test set;

and 5: acquiring a class activation view of a training set time-frequency graph through the trained convolutional neural network classification model, and integrating the class activation view to obtain a key area identification matrix;

step 6: enhancing all the time-frequency graphs according to the key area identification matrix;

and 7: and retraining the convolutional neural network classification model by using the enhanced training set to obtain a fault diagnosis model and predicting the fault degree of the sample.

2. The plunger pump fault signal time-frequency diagram noise reduction and enhancement method as claimed in claim 1, wherein the network structure of the convolutional neural network classification model is composed of three convolutional layers:

setting the size of a first layer of convolution kernels to be 3 x 3, setting the number of the first layer of convolution kernels to be 32, and setting the output size after convolution to be the same as the input size; the number of convolution kernels of the second layer and the third layer is set to 16;

and outputting classification results by utilizing two full-connection layers, and converting the forward propagation result of the convolutional neural network classification model into the probability distribution value of the classification label by adopting a SoftMax activation function in the last full-connection layer.

3. The plunger pump fault signal time-frequency diagram noise reduction enhancement method is characterized in that a cross entropy loss function is used for evaluating errors between a prediction result and a real label, an RMSProp optimization algorithm is adopted for iterative training to optimize a convolutional neural network classification model, and the formula of the loss function is as follows:

Loss(p,q)＝-∑_xp(x)q(x)

wherein p is the true value of the sample, q is the predicted label value, x is the serial number, p (x) is the probability distribution of the true value of the sample, and q (x) is the probability distribution of the predicted label value of the model.

4. The plunger pump fault signal time-frequency diagram noise reduction enhancement method according to claim 1, characterized in that the time-frequency diagrams of the training set samples are sequentially input into a preprocessing model to obtain an output feature diagram of a second layer of convolutional layer in the model, and each channel in the feature diagram is weighted by the gradient of the class relative to the channel to obtain a space diagram of the activation intensity of the input image to the class;

and accumulating the spatial maps of the activation intensity in sequence, only keeping data more than eighty bits in the spatial maps, obtaining key identification areas of different classes by adopting the same method for each class, and then overlapping the key identification areas of the different classes to obtain a key area identification matrix.

5. The plunger pump fault signal time-frequency diagram noise reduction enhancement method according to claim 1, wherein the enhancement processing comprises: and performing time-frequency transformation on the sample signal to obtain a two-dimensional array of the time-frequency transformation, performing point multiplication on the two-dimensional array and the key area identification matrix to obtain a processed two-dimensional array, and converting the processed two-dimensional array into a picture.

6. The utility model provides a plunger pump fault signal time-frequency diagram noise reduction reinforcing system which characterized in that includes:

module M1: acquiring vibration signals of the plunger pump under different fault levels by a triaxial vibration acceleration sensor arranged on a plunger pump shell;

module M2: slicing and segmenting the acquired vibration signal according to a preset size;

module M3: carrying out time-frequency transformation on the sliced sample to obtain a time-frequency graph of the sample, and dividing all the time-frequency graphs into a training set and a testing set according to a preset proportion;

module M4: training a convolutional neural network classification model by using training set data, and performing fault classification on a test set;

module M5: acquiring a class activation view of a training set time-frequency graph through the trained convolutional neural network classification model, and integrating the class activation view to obtain a key area identification matrix;

module M6: enhancing all the time-frequency graphs according to the key area identification matrix;

module M7: and retraining the convolutional neural network classification model by using the enhanced training set to obtain a fault diagnosis model and predicting the fault degree of the sample.

7. The plunger pump fault signal time-frequency diagram noise reduction and enhancement system as claimed in claim 6, wherein the network structure of the convolutional neural network classification model is composed of three convolutional layers:

setting the size of a first layer of convolution kernels to be 3 x 3, setting the number of the first layer of convolution kernels to be 32, and setting the output size after convolution to be the same as the input size; the number of convolution kernels of the second layer and the third layer is set to 16;

and outputting classification results by utilizing two full-connection layers, and converting the forward propagation result of the convolutional neural network classification model into the probability distribution value of the classification label by adopting a SoftMax activation function in the last full-connection layer.

8. The plunger pump fault signal time-frequency diagram noise reduction enhancement system as claimed in claim 6, wherein a cross entropy loss function is used to estimate the error between the prediction result and the real label, and a RMSProp optimization algorithm is used to perform iterative training to optimize a convolutional neural network classification model, wherein the formula of the loss function is as follows:

Loss(p,q)＝-∑_xp(x)q(x)

wherein p is the true value of the sample, q is the predicted label value, x is the serial number, p (x) is the probability distribution of the true value of the sample, and q (x) is the probability distribution of the predicted label value of the model.

9. The plunger pump fault signal time-frequency diagram noise reduction and enhancement system as claimed in claim 6, wherein the time-frequency diagrams of the training set samples are sequentially input into the preprocessing model to obtain the output feature diagram of the second layer convolutional layer in the model, and each channel in the feature diagram is weighted by the gradient of the class relative to the channel to obtain a space diagram of the activation intensity of the input image to the class;

and accumulating the spatial maps of the activation intensity in sequence, only keeping data more than eighty bits in the spatial maps, obtaining key identification areas of different classes by adopting the same method for each class, and then overlapping the key identification areas of the different classes to obtain a key area identification matrix.

10. The plunger pump fault signal time-frequency diagram noise reduction enhancement system of claim 6, wherein the enhancement processing comprises: and performing time-frequency transformation on the sample signal to obtain a two-dimensional array of the time-frequency transformation, performing point multiplication on the two-dimensional array and the key area identification matrix to obtain a processed two-dimensional array, and converting the processed two-dimensional array into a picture.

Background

The aviation hydraulic pump is a key element of an airplane hydraulic system, and meanwhile, the airplane hydraulic system widely adopts the plunger pump due to the characteristics of compact structure, small rotational inertia, large flow, easiness in control and the like of the plunger pump. To further increase the power density of the plunger pump, increasing the rotational speed is an effective method. Common failure types of high-speed plunger pumps include cavitation, abrasion and the like, and the failures easily cause the shell to be damaged, abnormal vibration and other adverse consequences and even cause safety accidents. Therefore, the method has important significance in fault diagnosis of the aviation plunger pump.

The traditional fault diagnosis mainly compares the health state with the running state of the pump when a fault occurs, and comprises data acquisition, feature extraction and fault classification and identification. The method mainly utilizes the spectral analysis to extract relevant characteristics, and then combines relevant classification algorithms such as SVM, random forest and other models to identify faults. The following disadvantages currently exist: 1) the accuracy of diagnosis depends heavily on the extraction of features, the feature extraction needs manual design, and the time is long and the experience is depended on; 2) the manually extracted features cannot guarantee sufficient representation of the features when the fault occurs; 3) many current diagnostic methods do not perform well under noisy acquisition signals.

The deep learning technology has strong feature representation capability, can automatically extract features, and has wide application in speech recognition and image processing. At present, some scholars apply the deep learning method to fault diagnosis of various mechanical devices. However, the conventional fault analysis method can also provide a certain reference, and particularly, a time-frequency analysis method and a time-frequency analysis technology are mature. The invention provides a method for signal noise reduction enhancement in fault diagnosis by combining a time-frequency graph and a convolutional neural network, so that the reliability and accuracy of diagnosis are improved.

Patent document CN106404386A (application number: CN201610757230.1) discloses a method for collecting, extracting and diagnosing early fault characteristic signals of a gearbox, in which an acoustic emission sensor is installed at a position to be monitored of gearbox equipment, and a bearing seat of the gearbox is generally selected to collect acoustic emission signals under the working state of the gearbox. And selecting noise-containing signals with different signal-to-noise ratios to calculate the singular spectrum slope under different decomposition layer numbers, gradually increasing the singular spectrum slope along with the increase of the decomposition layer numbers, and selecting the optimized decomposition layer number according to the optimization realization process of the decomposition layer number by utilizing the acquired acoustic emission signals. And according to the selected optimized decomposition layer number, analyzing and processing the acquired acoustic emission signals by utilizing the redundant lifting wavelet to obtain a time domain graph and a frequency domain graph of the signals. And judging the equipment fault condition through analyzing the time domain graph and the frequency domain graph.

Disclosure of Invention

Aiming at the defects in the prior art, the invention aims to provide a plunger pump fault signal time-frequency diagram noise reduction enhancement method and system.

The plunger pump fault signal time-frequency diagram noise reduction enhancement method provided by the invention comprises the following steps:

step 1: acquiring vibration signals of the plunger pump under different fault levels by a triaxial vibration acceleration sensor arranged on a plunger pump shell;

step 2: slicing and segmenting the acquired vibration signal according to a preset size;

and step 3: carrying out time-frequency transformation on the sliced sample to obtain a time-frequency graph of the sample, and dividing all the time-frequency graphs into a training set and a testing set according to a preset proportion;

and 4, step 4: training a convolutional neural network classification model by using training set data, and performing fault classification on a test set;

and 5: acquiring a class activation view of a training set time-frequency graph through the trained convolutional neural network classification model, and integrating the class activation view to obtain a key area identification matrix;

step 6: enhancing all the time-frequency graphs according to the key area identification matrix;

and 7: and retraining the convolutional neural network classification model by using the enhanced training set to obtain a fault diagnosis model and predicting the fault degree of the sample.

Preferably, the network structure of the convolutional neural network classification model is composed of three convolutional layers:

setting the size of a first layer of convolution kernels to be 3 x 3, setting the number of the first layer of convolution kernels to be 32, and setting the output size after convolution to be the same as the input size; the number of convolution kernels of the second layer and the third layer is set to 16;

and outputting classification results by utilizing two full-connection layers, and converting the forward propagation result of the convolutional neural network classification model into the probability distribution value of the classification label by adopting a SoftMax activation function in the last full-connection layer.

Preferably, a cross entropy loss function is used for evaluating errors between the prediction result and the real label, an RMSProp optimization algorithm is adopted for iterative training to optimize the convolutional neural network classification model, and the formula of the loss function is as follows:

Loss(p,q)＝-∑_xp(x)q(x)

wherein p is the true value of the sample, q is the predicted label value, x is the serial number, p (x) is the probability distribution of the true value of the sample, and q (x) is the probability distribution of the predicted label value of the model.

Preferably, the time-frequency graphs of the training set samples are sequentially input into the preprocessing model to obtain an output characteristic graph of the second layer of convolutional layer in the model, and each channel in the characteristic graph is weighted by the gradient of the class relative to the channel to obtain a space graph of the activation intensity of the input image to the class;

and accumulating the spatial maps of the activation intensity in sequence, only keeping data more than eighty bits in the spatial maps, obtaining key identification areas of different classes by adopting the same method for each class, and then overlapping the key identification areas of the different classes to obtain a key area identification matrix.

Preferably, the enhancement treatment comprises: and performing time-frequency transformation on the sample signal to obtain a two-dimensional array of the time-frequency transformation, performing point multiplication on the two-dimensional array and the key area identification matrix to obtain a processed two-dimensional array, and converting the processed two-dimensional array into a picture.

The plunger pump fault signal time-frequency diagram noise reduction enhancement system provided by the invention comprises:

module M1: acquiring vibration signals of the plunger pump under different fault levels by a triaxial vibration acceleration sensor arranged on a plunger pump shell;

module M2: slicing and segmenting the acquired vibration signal according to a preset size;

module M3: carrying out time-frequency transformation on the sliced sample to obtain a time-frequency graph of the sample, and dividing all the time-frequency graphs into a training set and a testing set according to a preset proportion;

module M4: training a convolutional neural network classification model by using training set data, and performing fault classification on a test set;

module M5: acquiring a class activation view of a training set time-frequency graph through the trained convolutional neural network classification model, and integrating the class activation view to obtain a key area identification matrix;

module M6: enhancing all the time-frequency graphs according to the key area identification matrix;

module M7: and retraining the convolutional neural network classification model by using the enhanced training set to obtain a fault diagnosis model and predicting the fault degree of the sample.

Preferably, the network structure of the convolutional neural network classification model is composed of three convolutional layers:

setting the size of a first layer of convolution kernels to be 3 x 3, setting the number of the first layer of convolution kernels to be 32, and setting the output size after convolution to be the same as the input size; the number of convolution kernels of the second layer and the third layer is set to 16;

and outputting classification results by utilizing two full-connection layers, and converting the forward propagation result of the convolutional neural network classification model into the probability distribution value of the classification label by adopting a SoftMax activation function in the last full-connection layer.

Preferably, a cross entropy loss function is used for evaluating errors between the prediction result and the real label, an RMSProp optimization algorithm is adopted for iterative training to optimize the convolutional neural network classification model, and the formula of the loss function is as follows:

Loss(p,q)＝-∑_xp(x)q(x)

wherein p is the true value of the sample, q is the predicted label value, x is the serial number, p (x) is the probability distribution of the true value of the sample, and q (x) is the probability distribution of the predicted label value of the model.

Preferably, the time-frequency graphs of the training set samples are sequentially input into the preprocessing model to obtain an output characteristic graph of the second layer of convolutional layer in the model, and each channel in the characteristic graph is weighted by the gradient of the class relative to the channel to obtain a space graph of the activation intensity of the input image to the class;

and accumulating the spatial maps of the activation intensity in sequence, only keeping data more than eighty bits in the spatial maps, obtaining key identification areas of different classes by adopting the same method for each class, and then overlapping the key identification areas of the different classes to obtain a key area identification matrix.

Preferably, the enhancement treatment comprises: and performing time-frequency transformation on the sample signal to obtain a two-dimensional array of the time-frequency transformation, performing point multiplication on the two-dimensional array and the key area identification matrix to obtain a processed two-dimensional array, and converting the processed two-dimensional array into a picture.

Compared with the prior art, the invention has the following beneficial effects:

according to the method, the time-frequency analysis is combined with the frequency analysis and the convolutional neural network, the time-frequency analysis is used for extracting the characteristics, the convolutional neural network is used for classifying the images, and the class activation intensity graph is used for enhancing the time-frequency graph, so that the model diagnosis performance is good and the accuracy is high under the noise condition.

Drawings

Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a diagram of a sample time-frequency transformation effect;

FIG. 3 is a flow chart of a time-frequency diagram enhancement process;

FIG. 4 is a diagram of a model network architecture;

FIG. 5 is a diagram of the original time-frequency transformation;

FIG. 6 is a diagram of time-frequency transform and enhancement processing;

FIG. 7 is a graph of accuracy and loss during training;

FIG. 8 is a graph of accuracy and loss for a test set at different noises.

Detailed Description

The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit of the invention. All falling within the scope of the present invention.

Example 1:

the invention aims to provide a fault diagnosis method for a high-speed aviation plunger pump, which can realize signal noise reduction enhancement, thereby realizing signal feature extraction and fault diagnosis and improving the diagnosis accuracy of a signal containing noise.

In order to achieve the aim, the invention provides a Grad-CAM-based plunger pump fault signal time-frequency diagram noise reduction enhancement method, which is characterized by comprising the following steps of:

s1: collecting vibration signal of shell when plunger pump is in fault

Collecting vibration signals of the plunger pump under different inlet pressures at sampling frequency fs

S2: data sample segmentation

The method adopted by the invention needs to perform time-frequency transformation on the original signal to obtain a time-frequency diagram of the sample, wherein the time-frequency diagram is the input of subsequent model training and testing. The original signal is divided into N sections, and the number of samples in each type is as follows: l_i＝n_iTotal number of samples/s:according to the above relationship, the vibration signal becomes an N-segment signal, each segment having a label. Is marked asWhere s is the length of each sample and the outermost subscript indicates the class.

m represents the number of sample categories; i denotes a subscript of a certain sample class;representing the first segment of the original signal after segmentation.

S3: segmented sample time-frequency transform

Obtaining a time-frequency diagram by adopting short-time Fourier transform; dividing data into training and testing sets, dividing samples into training set X in each group of time-frequency converted pictures of sample data according to a certain proportion_trainAnd test set X_test；

S4: pre-processing model training

And (3) constructing a CNN classification model for classifying and diagnosing the time-frequency graph, and learning the fault characteristics in the time-frequency graph mainly through three convolution layers and two full-connection layers. The details of the construction of the network structure are as follows:

convolutional layer conv2d _1 is the first layer to extract features of a picture and needs to be able to extract features sufficiently to provide an output containing sufficient useful information for subsequent layers. The number of convolution kernels (filters) is set to 32, the size is set to (3,3), and the size of the plane after convolution is set to be the same as the original (padding), so that the dimension of the subsequent feature map is consistent with the input conveniently. The Conv2d _1 layer weight initialization adopts RandomNormal; the activation function adopts a relu activation function, so that the convergence speed can be increased, and the condition of gradient disappearance is avoided. The volume normalization layer (batch _ normalization _1) and the random discard layer (dropout _1) are arranged after the convolution layer.

The parameter setting of convolution layer conv2d _2 is consistent with that of conv _1 except that the number of filters is different, and the number of convolution kernels is set to 16. Setting a batch normalization layer (batch _ normalization _2) and a random discard layer (dropout _2) after the convolutional layer

The parameter setting of the convolution layer conv2d _3 is consistent with conv2d _1, and the subsequent batch standardization layer batch _ normalization _3 and random discard layer (dropout _3) are also set

And finally, outputting a predicted classification result by the network, outputting the classification result by utilizing the two fully-connected layers, and converting the result of model forward propagation into a probability distribution value of a classification label by adopting SoftMax (software-based maximum max) as an activation function of the last fully-connected layer in order to obtain a multi-classification prediction result.

The error between the prediction result and the true label is evaluated using a cross entropy loss function. The optimization target is a minimum loss function, the model training optimization method adopts an RMSROP optimization algorithm, and the model training optimization method iterates for enough times to obtain a trained model. The formula for the loss function is:

where p (x) is the probability distribution of the true values of the samples, and q (x) is the probability distribution of the values of the model predictive labels.

S5: key region identification

After the pre-processing model is obtained, an effective diagnosis of the acquired signals can already be made. However, the time-frequency diagram with the noise signal is obviously different from that under the noise-free condition, so that the fault diagnosis accuracy of the noise-containing signal is obviously reduced. The invention provides a method for carrying out noise reduction enhancement processing on a time frequency graph by utilizing a class activation intensity graph of a model, which does not solve the problems and comprises the following steps:

inputting a model by a time-frequency diagram to obtain characteristic diagram output of a conv2d _2 layer in the model;

calculating the gradient of the model class output corresponding to the feature map by using a function provided by Keras;

averaging the gradient of the characteristic diagram channel, and taking the gradient average as the weight of the characteristic diagram;

multiplying the characteristic graph and the gradient mean value to obtain a gradient weighted characteristic graph, and then averaging the weighted characteristic graph channel by channel to obtain a class activation intensity graph of a sample;

superposing the class activation intensity maps of all samples, and only retaining data more than eighty bits to obtain a certain class of key identification areas;

and repeating the steps in different categories to finally obtain key identification areas with different fault grades, and overlapping the areas to obtain a signal enhancement processing matrix.

S6: data reprocessing

After the time-frequency transformation is carried out on the sample signals, the element corresponding multiplication is carried out on the sample signals and the signal enhancement processing matrix obtained in S5, and then the processed time-frequency transformation array is converted into a time-frequency graph. The process is shown in the attached drawings

S7: model retraining

And (4) retraining the model by using the training set after the re-processing, so that the obtained model can more accurately identify the key areas.

S8: fault diagnosis

Test set X_testInputting the trained CNN model, and predicting the fault degree of the sample.

Example 2:

example 2 is a preferred example of example 1.

Referring to fig. 1, which is a flow chart of the method for noise reduction and enhancement of the time-frequency diagram of the fault signal of the plunger pump based on the Grad-CAM of the present invention, the method for diagnosing the cavitation fault of the high-speed aviation plunger pump comprises the following steps:

s1: a vibration sensor is arranged on a shell of the plunger pump, and is connected with a collecting device to collect vibration signals when the pump generates cavitation of different degrees under different inlet pressures, wherein the sampling frequency is 10240 Hz.

S2: the original vibration signal is divided into N segments, each segment being a sample. Each sample was assigned a cavitation rating according to the flow loss. The implementation case is divided into four grades, namely severe cavitation, medium cavitation, slight cavitation and no cavitation, and each section of sample length in the implementation case adopts 256 data points.

The severity of cavitation is measured by flow loss, and is expressed as:

wherein: q. q.s_tIs the theoretical flow rate, q_inIs the actual inlet flow.

Dividing a training set and a test set, dividing the converted picture into the training set and the test set according to the test set proportion of 0.2, namely taking 80% of samples as a training set X_trainAnd the rest as test set X_test。

S3: sample time-frequency transformation

The single rotating speed working condition of the operation working condition of the embodiment has more periodic components of the acquired vibration signal, so that the time-frequency analysis by adopting short-time Fourier transform is more suitable. The effect is shown in fig. 2 by selecting appropriate short-time fourier transform parameters.

S4: establishing a preprocessing model

Model structure as shown in fig. 4, the input is the time-frequency plot after the time-frequency transform of the samples, the first layer is convolution layer conv2d _1, the convolution kernel size settings (3,3), and the number is 32. Then through the batch normalization layer and the random discard layer. And then connected to the next winding layer conv2d _ 2. The convolution kernel size of the second convolutional layer is set to (3,3), the number of convolution kernels is set to 16, and then the second convolutional layer is connected to a third convolutional layer conv2d _3 through a batch normalization layer and a random discard layer, and the third convolutional layer setting is consistent with the second convolutional layer. And expanding the data through a batch normalization layer and a random discarding layer. After expansion, a full connection layer with the neuron number of 32 is arranged, and finally, the output layer is connected through batch standardization and random abandon layers. The output layer outputs the category of cavitation failures.

And establishing a loss function, and selecting an optimization algorithm to train the model. The Loss function uses Cross Entropy (Cross entry Loss) as the Loss function. For the four types of fault classes in this example, assuming that the true value of the sample is q (x) ═ 0, 1, 0, 0 and the predicted value of the model is p (x) ═ a, b, c, d, then:

Loss(p,q)＝-0·loga-1·logb-0·logc-0·logd

wherein a, b, c and d represent the inverse of the probability of the model respectively estimating the four fault classes.

And training and testing the pretreatment model to obtain the pretreatment model with good performance on the test set.

S5: key region identification

And (4) obtaining a feature map of a conv2d _2 layer in the model and a gradient between the model output class and the feature map by utilizing the preprocessing model. And obtaining a key point identification area of the preprocessing model according to the specification S5 by using the feature map and the gradient.

S6: data reprocessing

And performing point multiplication on the data after the sample time-frequency transformation and the key identification area matrix, converting the data into a time-frequency graph, and performing reprocessing on the training set and the test set by referring to the graph 3.

S7: diagnostic model training

And establishing and training a model structure and a training method as well as a preprocessing model.

S8: and inputting the test set data into the diagnosis model, and predicting the idle call degree of the test set.

More specifically, the invention utilizes a plunger pump fault simulation experiment table to acquire vibration signals under different inlet pressures through a triaxial vibration acceleration sensor arranged on a plunger pump shell. The laboratory bench can measure import export flow, divides the cavitation degree through flow loss degree, as shown in the following table:

inlet pressure (Mpa)	0.25	0.15	0.10	0
					Loss of flow	1.0％	2.0％	8.0％	76.0％
Severity of cavitation	Without empty stomach	Slight cavitation	Medium-sized air speech	Severe cavitation

The acquisition frequency of the vibration signal is 10240Hz, and the acquired original signals are divided into four types. Segmenting each type of data and converting the data into a time-frequency graph, and dividing a training set and a test set according to a proportion, wherein the conditions of the data set are as follows:

categories	Training set	Test set
			Severe cavitation	[email protected](128,128,3)	[email protected](128,128,3)
Moderate cavitation	[email protected](128,128,3)	[email protected](128,128,3))
			Slight cavitation	[email protected](128,128,3)	[email protected](128,128,3)
Without cavitation	[email protected](128,128,3)	[email protected](128,128,3)
			Total of	768	192

Results of the experiment

(1) Sample time-frequency graph conversion result and enhancement processing

According to the method, the samples are subjected to time-frequency transformation and stored into pictures, and meanwhile, the noise reduction enhancement method provided by the invention is utilized for processing, and the processing results are shown in fig. 5 and fig. 6.

(2) Model prediction accuracy

Models are built and trained by utilizing python and TensorFlow, the performance of the models is verified on a test set, the accuracy of the trained models in the verification set can reach 99.5%, and the accuracy and the loss curve in the training process are shown in figure 7.

(3) Model noise immunity

In order to better simulate the complex working conditions of real fault monitoring, white noise with different signal-to-noise ratios (SNR) is added into the original signals of the test set. The model is trained on data without white noise, and due to the data processing enhancement mode of the invention, the method has good anti-noise performance, such as figure 8, which is a graph of accuracy and loss of a test set under different noises.

The calculation formula of the vibration signal-to-noise ratio is as follows:

wherein: p represents the power level of the signal, P_signalIndicating the magnitude of the signal power, P, containing no noise_nosieRepresenting the power level of the noise.

The anti-noise capability under the condition that SNR is-4-10 dB is verified through experiments, and the result is as follows:

SNR	0	2	4	6	8	10
							rate of accuracy	63.54％	78.12％	90.10％	95.31％	97.91％	98.43％
Loss of power	0.90	0.60	0.37	0.22	0.13	0.08

The result shows that the cavitation fault diagnosis method provided by the invention has good anti-noise capability, the accuracy of the test set is higher than 80% under the condition that the signal-to-noise ratio is higher than 2dB, and the corresponding loss function value is smaller, thereby showing that the diagnosis reliability is high and the robustness is strong.

Those skilled in the art will appreciate that, in addition to implementing the systems, apparatus, and various modules thereof provided by the present invention in purely computer readable program code, the same procedures can be implemented entirely by logically programming method steps such that the systems, apparatus, and various modules thereof are provided in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Therefore, the system, the device and the modules thereof provided by the present invention can be considered as a hardware component, and the modules included in the system, the device and the modules thereof for implementing various programs can also be considered as structures in the hardware component; modules for performing various functions may also be considered to be both software programs for performing the methods and structures within hardware components.

The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.