CN116681941A - Fan main shaft multi-signal input fault diagnosis method based on Glow model - Google Patents

Abstract

The application discloses a fan main shaft multi-signal input fault diagnosis method based on a Glow model, which relates to the field of fan main shaft fault diagnosis and comprises the following steps: the method comprises the steps of data acquisition, glow model construction, fault diagnosis model training and fault diagnosis. The application aims at solving the problem of resource waste caused by long-time shutdown and maintenance due to fan faults, and is beneficial to improving the operation reliability of fan equipment.

Description

Multi-signal input fault diagnosis method for fan shaft based on Glow model

technical field

The invention relates to the field of fault diagnosis of a fan main shaft, in particular to a multi-signal input fault diagnosis method for a fan main shaft based on a Glow model.

Background technique

Bearings are a key component of the wind turbine transmission system, and the transmission system is a key component of any rotating machine. Some transmission system failures that are not related to bearings, such as gears and blades, may also be directly or indirectly caused by bearing failures. Planned maintenance and after-sales service have long been used in wind turbines. If the running condition of the bearing can be monitored in real time, it will be very important for the overall fault diagnosis of the transmission system and the operation and maintenance of the wind turbine.

The main bearing of the fan is mainly composed of outer ring, inner ring, ball body and cage. One end of the fan shaft is connected to the fan blades, and the other end is connected to the fan drive system. The inner ring of the main shaft is connected with the shaft, the outer ring is connected with the cage, and the ball body is the key part of the bearing rotation. Therefore, the inner ring, outer ring and ball of the main bearing of the wind turbine may fail. In order to ensure effective fault diagnosis of wind turbine main shaft, it is necessary to study the wind turbine main shaft fault theory first. During operation, the main bearing area is susceptible to failure due to external factors. Therefore, the fault principle of the wind turbine main bearing should be deeply analyzed to improve the fault diagnosis accuracy of the wind turbine main bearing.

Fault diagnosis technology originates from the sensory perception of temperature, sound, smell, etc. by operation and maintenance personnel to judge the existence of faults. Based on bearing vibration signal is the most widely used fault diagnosis method at present. Traditional fault diagnosis methods mainly involve signal acquisition and processing, and the establishment of fan fault sample feature library. The main solution is to extract fault attributes from input signals, and then diagnose by classifying fault attributes. In recent years, on the basis of observing the fault characteristics in the time-frequency domain by detecting the partial or overall vibration of the equipment, researchers have begun to conduct in-depth research on fault diagnosis combined with artificial intelligence algorithms, and use deep learning algorithms to realize fault diagnosis of fan bearings.

The emergence of artificial intelligence has brought a new research direction to the field of fan bearing fault diagnosis. Compared with traditional methods, the deep learning model trained by vibration signals can automatically learn and extract fault features to achieve "end-to-end" fault diagnosis. In the case of limited fault samples, fault diagnosis based on deep learning algorithm has greater advantages.

Therefore, it is an urgent problem for those skilled in the art to propose a multi-signal input fault diagnosis method for the fan shaft based on the Glow model to solve the difficulties existing in the prior art.

Contents of the invention

In view of this, the present invention provides a multi-signal input fault diagnosis method for fan main shaft based on the Glow model, which improves the problem of long-time downtime caused by fan failure and waste of resources caused by inspection and maintenance, and improves the reliability of fan equipment operation.

In order to achieve the above object, the present invention adopts the following technical solutions:

Obtaining data step: obtaining data sets through sensors;

Glow model construction steps: establish a Glow model through real sample data X and random variable Z;

Construction steps of the fault diagnosis model: Based on the convolutional neural network, a multi-image input fault diagnosis model for the fan shaft is proposed;

Fault diagnosis model training steps: Use the Glow model to supplement the generation of fan shaft fault samples in different scenarios, supplement the number of unbalanced fault category samples, and add different degrees of Gaussian white noise to the real samples to simulate the noise scene in the actual operation of the fan. The fault sample set is input into the fault diagnosis model through the multi-image of the fan main shaft, and according to the accuracy of the fault diagnosis, the trained multi-image input fault diagnosis model of the fan main shaft is obtained;

Fault diagnosis steps: Use the trained fan shaft multi-image input fault diagnosis model to diagnose the fan shaft to determine whether the fan shaft is faulty.

In the above method, optionally, in the data acquisition step, an acceleration sensor is used to collect data sets, the data sets include but not limited to wind turbine bearing normal state signals, ball fault vibration signals, inner ring fault vibration signals and outer ring fault signals.

In the above method, optionally, in the Glow model construction step, the real sample data X and the random variable Z, where Z obeys a known simple prior distribution π(Z), and the sample data X obeys a complex distribution p(X), there is a The transformation function f satisfies the establishment of a mapping from Z to X

f:Z→X (1)

So that for every sampling point in π(Z), there can be a new sample point corresponding to it in p(X), so as to obtain the generated samples. In the standardized flow model, the random variable Z prior distribution π(Z) Usually a Gaussian distribution is chosen, and the standardized flow transforms a simple distribution into a complex distribution by applying a series of reversible transformation functions, repeatedly replacing new variables according to the variable substitution theorem, and finally obtaining the probability distribution of the final target variable.

In the above method, optionally, the generation process of the standardized flow model can be defined by the following formula:

Z～π(Z) (2)

X＝g _θ (Z) (3)

In the formula, Z is expressed as a hidden variable; p _θ (Z) is the sample distribution of hidden variable Z; g _θ is an invertible function, and hidden variable Z can be expressed as Among them, f _θ consists of a series of transformed functions: /> θ is the parameter of the generated model, and the relationship between the sample X and the latent variable Z ₀ can be written as:

By outputting x until the initial distribution z is traced back, the probability density function of the model given a sample data x can be expressed as:

The training loss function for flow-based generative models is the negative log-likelihood on the training dataset:

In the above method, optionally, in the Glow model construction step, the Glow model consists of a series of repeated layers named scales, each scale includes a squeeze function and a flow step, and the flow step is followed by a segmentation function; the segmentation function is in The input is divided into two equal parts in the channel dimension; half of which enters the subsequent layer and the other half enters the loss function; the split is to reduce the influence of gradient disappearance, which will appear when the model is trained in an end-to-end manner; flow The steps include three parts: activation constant layer, 1x1 reversible convolution layer and affine coupling layer.

In the above method, optionally, the activation constant layer is used for activation normalization, which uses the scale and deviation parameters of each channel to perform affine transformation on the activation, similar to batch normalization, and initializes these parameters so that in the given In the case of a small batch of initial data, the subsequent behavior of each channel has zero mean and unit variance. After initialization, the scale and bias are regarded as regular trainable parameters independent of the data;

The 1x1 reversible convolutional layer is used to reverse the ordering of channels, where the weight matrix is initialized as a random rotation matrix, and the number of input and output channels of the convolutional layer is the same;

Affine coupling layers build bijective functions by superimposing a series of simple bijections, in each simple bijection a part of the input vector is updated using a function that is simply reversed, but which depends in complex ways on The remainder of the input vector, the affine coupling layer can be divided into three parts: zero initialization, split and concatenation, permutation.

In the above method, optionally, in the fault diagnosis model building step, the multi-image input fault diagnosis model of the fan main shaft realizes the fault diagnosis function by learning the fault characteristics of multiple image inputs: the input is two grayscale images with a size of 28*28 , each input is convolved through multiple convolutional layers, the output is connected to the fully connected layer, and the multiplication layer is used to multiply the input from the two fully connected layers; the output is the classification output layer, and the activation function is softmax.

In the above method, optionally, in the fault diagnosis model training step, the fault samples are divided into 19 categories according to the fault category, fault location and damage diameter.

In the above method, optionally, the vibration signal in the vibration data set of the main shaft of the fan is a one-dimensional time-series signal, and the one-dimensional time-series signal from the fan end and the drive end of the fan main shaft are respectively converted into a two-dimensional image signal, and the two-dimensional convolutional network is used to perform Fault sample generation and fault diagnosis; the original 1D time-domain signal needs to be normalized before sample construction.

It can be seen from the above-mentioned technical solutions that, compared with the prior art, the present invention provides a multi-signal input fault diagnosis method for the main shaft of a fan based on the Glow model. Early symptoms, determine the type and degree of failure, so that the wind farm can adopt an effective maintenance plan, avoid the occurrence of major safety accidents and the waste of resources caused by inspection and maintenance, optimize the equipment inspection and maintenance work plan, and improve the reliability of wind turbine equipment operation; can effectively avoid The long downtime caused by fan failure can reduce the loss of power generation, reduce the failure rate and maintenance cost of the gearbox by at least 80%, increase the utilization rate of fan lubricating oil by about 10%, and save about 9% of the operation and maintenance cost.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only It is an embodiment of the present invention, and those skilled in the art can also obtain other drawings according to the provided drawings without creative work.

Fig. 1 is the flow chart of the multi-signal input fault diagnosis method for fan main shaft based on the Glow model provided by the present invention;

Fig. 2 is the structure diagram of the Glow model provided by the present invention;

Fig. 3 is a flow step structure diagram in the Glow model provided by the present invention;

Fig. 4 is the fault sample construction process figure provided by the present invention;

Fig. 5 is a multi-signal input fault diagnosis model diagram provided by the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In this application, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any relationship between these entities or operations. any actual relationship or order, the terms "comprises," "comprises," or any other variation thereof are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that includes a series of elements includes not only those elements, but also Including other elements not expressly listed, or also including elements inherent in such process, method, article or apparatus. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

Referring to Fig. 1, the present invention discloses a multi-signal input fault diagnosis method for fan main shaft based on the Glow model, comprising the following steps:

The multi-signal input fault diagnosis method of fan main shaft based on Glow model is characterized in that, comprising the following steps:

Obtaining data step: obtaining data sets through sensors;

Glow model construction steps: establish a Glow model through real sample data X and random variable Z;

Construction steps of the fault diagnosis model: Based on the convolutional neural network, a multi-image input fault diagnosis model for the fan shaft is proposed;

Fault diagnosis model training steps: Use the Glow model to supplement the generation of fan shaft fault samples in different scenarios, supplement the number of unbalanced fault category samples, and add different degrees of Gaussian white noise to the real samples to simulate the noise scene in the actual operation of the fan. The fault sample set is input into the fault diagnosis model through the multi-image of the fan main shaft, and according to the accuracy of the fault diagnosis, the trained multi-image input fault diagnosis model of the fan main shaft is obtained;

Fault diagnosis steps: Use the trained fan shaft multi-image input fault diagnosis model to diagnose the fan shaft to determine whether the fan shaft is faulty.

Further, in the step of acquiring data, the data set is collected by the acceleration sensor, and the data set includes but not limited to the normal state signal of the wind turbine bearing, the vibration signal of the ball body fault, the vibration signal of the inner ring fault and the outer ring fault signal.

Further, as shown in Figure 2, the real sample data X and random variable Z in the Glow model construction step, where Z obeys a known simple prior distribution π(Z), and the sample data X obeys a complex distribution p(X), There exists a transformation function f that satisfies the establishment of a mapping from Z to X

f:Z→X (1)

So that for every sampling point in π(Z), there can be a new sample point corresponding to it in p(X), so as to obtain the generated samples. In the standardized flow model, the random variable Z prior distribution π(Z) Usually a Gaussian distribution is chosen, and the standardized flow transforms a simple distribution into a complex distribution by applying a series of reversible transformation functions, repeatedly replacing new variables according to the variable substitution theorem, and finally obtaining the probability distribution of the final target variable.

Furthermore, the generation process of the standardized flow model can be defined by the following formula:

Z～π(Z) (2)

X＝g _θ (Z) (3)

In the formula, Z is expressed as a hidden variable; p _θ (Z) is the sample distribution of hidden variable Z; g _θ is an invertible function, and hidden variable Z can be expressed as Among them, f _θ consists of a series of transformed functions: /> θ is the parameter of the generated model, and the relationship between the sample X and the latent variable Z ₀ can be written as:

By outputting x until traced back to the initial distribution _z , the probability density function of the model given a sample data x can be expressed as:

The training loss function for flow-based generative models is the negative log-likelihood on the training dataset:

Further, as shown in Figure 3, in the Glow model construction step, the Glow model consists of a series of repeated layers named scales, each scale includes a squeeze function and a flow step, and the flow step is followed by a segmentation function; the segmentation function is in The input is divided into two equal parts in the channel dimension; half of which enters the subsequent layer and the other half enters the loss function; the split is to reduce the influence of gradient disappearance, which will appear when the model is trained in an end-to-end manner; flow The steps include three parts: activation constant layer, 1x1 reversible convolution layer and affine coupling layer.

Furthermore, the activation constant layer is used for activation normalization, which uses the scale and deviation parameters of each channel to perform affine transformation on the activation, similar to batch normalization, and initializes these parameters so that the given initial data is small In the batch case, the post-behavior actions for each channel have zero mean and unit variance, and after initialization, scale and bias are treated as regular trainable parameters independent of the data;

The 1x1 reversible convolutional layer is used to reverse the ordering of channels, where the weight matrix is initialized as a random rotation matrix, and the number of input and output channels of the convolutional layer is the same;

Affine coupling layers build bijective functions by superimposing a series of simple bijections, in each simple bijection a part of the input vector is updated using a function that is simply reversed, but which depends in complex ways on The remainder of the input vector, the affine coupling layer can be divided into three parts: zero initialization, split and concatenation, permutation.

Further, as shown in Figure 4, in the fault diagnosis model construction step, the multi-image input fault diagnosis model of the fan main shaft realizes the fault diagnosis function by learning the fault characteristics of multiple image inputs: the input is two grayscale images with a size of 28*28 Image, each input is convolved through multiple convolutional layers, the output is connected to the fully connected layer, and the multiplication layer is used to multiply the input from the two fully connected layers; the output is the classification output layer, and the activation function is softmax.

Further, in the fault diagnosis model training step, the fault samples are divided into 19 categories according to the fault category, fault location and damage diameter. The fault classification of the fan shaft is shown in Table 1:

Table 1 Fault classification of fan shaft

Furthermore, the vibration signal in the fan shaft vibration data set is a one-dimensional time-series signal, and the one-dimensional time-series signal from the fan end and the drive end of the fan shaft is converted into a two-dimensional image signal, and the fault samples are generated through a two-dimensional convolutional network and fault diagnosis; the original 1D time-domain signal needs to be normalized before sample construction.

Further, in the fault diagnosis model training step, in order to verify the performance of the generated model, we use image quality indicators such as maximum mean difference (MMD), peak signal-to-noise ratio (PSNR), and feature similarity (FSIM) to evaluate the quality of the generated fault samples. authenticity. The maximum mean difference can use the size of the sum of each image projection to judge the distribution difference of the two images. The smaller the value, the smaller the image distribution difference. The peak signal-to-noise ratio is based on the relative value of the mean square error of the original image and the generated image and the square of the maximum possible signal value of the image. The larger the value of PSNR, the higher the quality of the generated sample. Feature similarity can be used to evaluate image quality, and the higher the value, the better the similarity.

Further, referring to FIG. 5 , in the fault diagnosis step, the multi-signal input fault diagnosis model judges the fault category by image quality.

In a specific embodiment, the model disclosed in the present invention is simulated,

1. Table 2 is the quality evaluation of fault samples generated under different noise intensities. According to the size of the three indicators, the image quality of the original fault samples generated by supplementary Glow model is better than that of fault samples generated by supplementary noise. Compared with the fan shaft fan end fault sample, the image quality of the generated drive end fault sample is better. In the comparison of different image quality metrics, we found that the quality of the faulty samples generated by the model does not decrease significantly in the case of noise interference. The implementation results show that the proposed model has a good ability to generate fan shaft fault samples.

Table 2 Quality evaluation of fault samples generated

2. The Glow model is used to generate new samples based on the vibration data of the two sets of fan shafts, and the number of unbalanced fault samples is supplemented. The accuracy of fault diagnosis is obtained by using the multi-input fault diagnosis model of the fan shaft for the fault sample set with a balanced number. The sample set construction under the sample imbalance scenario is shown in Table 3, where the number of samples for the balanced category is 2000 for each category, and the number of samples for the unbalanced category is 1000 for each category. Table 4 shows the fault diagnosis accuracy of different sample sets under the sample imbalance scenario. When the sample imbalance degree is relatively low, the model can achieve 100% accurate diagnosis of fan shaft failure. As the degree of sample imbalance increases, the fault diagnosis accuracy of the sample set decreases. However, the model can still effectively diagnose the existence of fan shaft faults, which shows that the proposed model has a good fault diagnosis ability in the sample imbalance scenario.

Table 3 Sample set construction under sample imbalance scenario

Table 4 Fault diagnosis results of different sample sets under sample imbalance scenario

3. In order to verify the fault diagnosis performance of the proposed method in complex scenarios, the sample sets were constructed based on different numbers of training samples, and the number of samples was supplemented through the Glow model until the number of samples in the sample set was the same, and the proposed method was tested on the same test set. diagnostic performance of the method. The sample set construction in the sample insufficient scenario is shown in Table 5.

Table 6 shows the fault diagnosis accuracy of the original sample set and the supplementary sample set under the under-sample scenario. As shown in the figure, insufficient samples will lead to a decline in the accuracy of fault diagnosis in the sample set. By generating a sample set that complements the fault samples for the model, the accuracy of fault diagnosis will be higher. Although the fault diagnosis effect of the sample set supplemented by the deep learning generation model is not good enough, it has also been verified that the proposed fault diagnosis model has a certain fault diagnosis performance when the samples are insufficient, and the fan shaft fault samples generated by Glow can effectively improve Accuracy of fault diagnosis.

Table 5 Sample set construction under sample insufficient scenario

Table 6 Fault diagnosis results of the original sample set and the supplementary sample set in the case of insufficient samples

4. Consider that the collection of vibration data in the actual operation of the fan may be disturbed by noise. In order to verify the effectiveness of the proposed fault diagnosis model in practical applications, the noise scene in the actual operation of the fan is simulated by adding different degrees of Gaussian white noise to the real sample data.

Table 7 shows the fault diagnosis accuracy of different sample sets under noise interference. When the fault samples are balanced and sufficient, the fault diagnosis accuracy of the sample set will be significantly higher than other sample sets. When the samples are insufficient and unbalanced, the accuracy of fault diagnosis of the sample set decreases. In special scenarios, the accuracy of fault diagnosis of each sample set is relatively low, but the accuracy of fault diagnosis of each sample set is above 97.5%. The research shows that the proposed model has a good performance in the fault diagnosis of the fan shaft in the noise scene.

Table 7 Fault diagnosis results of different sample sets under different noise intensities

Each embodiment in this specification is described in a progressive manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for related parts, please refer to the part of the description of the method embodiment. The systems and system embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is It can be located in one place, or it can be distributed to multiple network elements. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. It can be understood and implemented by those skilled in the art without creative effort.

The above description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (9)

1. the multi-signal input fault diagnosis method of fan main shaft based on Glow model, it is characterized in that, comprises the following steps: Obtaining data step: obtaining data sets through sensors; Glow model construction steps: establish a Glow model through real sample data X and random variable Z; Construction steps of the fault diagnosis model: Based on the convolutional neural network, a multi-image input fault diagnosis model for the fan shaft is proposed; Fault diagnosis model training steps: Use the Glow model to supplement the generation of fan shaft fault samples in different scenarios, supplement the number of unbalanced fault category samples, and add different degrees of Gaussian white noise to the real samples to simulate the noise scene in the actual operation of the fan. The fault sample set is input into the fault diagnosis model through the multi-image of the fan main shaft, and according to the accuracy of the fault diagnosis, the trained multi-image input fault diagnosis model of the fan main shaft is obtained; Fault diagnosis steps: Use the trained fan shaft multi-image input fault diagnosis model to diagnose the fan shaft to determine whether the fan shaft is faulty. 2. the fan shaft multi-signal input fault diagnosis method based on Glow model according to claim 1, is characterized in that, In the data acquisition step, the data set is collected by the acceleration sensor, and the data set includes but not limited to the normal state signal of the wind turbine bearing, the vibration signal of the ball body fault, the vibration signal of the inner ring fault and the outer ring fault signal. 3. the fan shaft multi-signal input fault diagnosis method based on Glow model according to claim 1, is characterized in that, In the Glow model construction step, the real sample data X and the random variable Z, where Z obeys the known simple prior distribution π(Z), and the sample data X obeys the complex distribution p(X), there is a transformation function f that satisfies the establishment from Z to X mapping f:Z→X (1) So that for every sampling point in π(Z), there can be a new sample point corresponding to it in p(X), so as to obtain the generated samples. In the standardized flow model, the random variable Z prior distribution π(Z) Usually a Gaussian distribution is chosen, and the standardized flow transforms a simple distribution into a complex distribution by applying a series of reversible transformation functions, repeatedly replacing new variables according to the variable substitution theorem, and finally obtaining the probability distribution of the final target variable. 4. the fan shaft multi-signal input fault diagnosis method based on Glow model according to claim 3, is characterized in that, The generation process of the standardized flow model is defined by the following formula: Z～π(Z) (2) X＝g _θ (Z) (3) In the formula, Z is expressed as a hidden variable; p _θ (Z) is the sample distribution of hidden variable Z; g _θ is an invertible function, and hidden variable Z can be expressed as Among them, f _θ consists of a series of transformed functions: /> θ is the parameter of the generated model, and the relationship between the sample X and the latent variable Z ₀ is: By outputting x until the initial distribution z is traced back, the probability density function of the model given a sample data x can be expressed as: The training loss function for flow-based generative models is the negative log-likelihood on the training dataset: 5. the fan shaft multi-signal input fault diagnosis method based on Glow model according to claim 1, is characterized in that, In the Glow model construction step, the Glow model consists of a series of repeated layers named scales, each scale includes a squeeze function and a flow step, and the flow step is followed by a split function; the split function divides the input into two in the channel dimension Equal parts; half of them go to the following layer, and the other half goes to the loss function; the flow step consists of three parts: activation constant layer, 1x1 reversible convolution layer and affine coupling layer. 6. the multi-signal input fault diagnosis method of fan main shaft based on Glow model according to claim 5, is characterized in that, The activation constant layer is used for activation normalization, which uses the scale and bias parameters of each channel to affine transform the activation, similar to batch normalization, and initializes these parameters so that given a small batch of initial data , the post-behavioral actions of each channel have zero mean and unit variance, and after initialization, the scale and bias are treated as regular trainable parameters independent of the data; The 1x1 reversible convolutional layer is used to reverse the ordering of channels, where the weight matrix is initialized as a random rotation matrix, and the number of input and output channels of the convolutional layer is the same; Affine coupling layers build bijective functions by superimposing a series of simple bijections, in each simple bijection a part of the input vector is updated using a function that is simply reversed, but which depends in complex ways on The remainder of the input vector, the affine coupling layer can be divided into three parts: zero initialization, split and concatenation, permutation. 7. the multi-signal input fault diagnosis method of fan main shaft based on Glow model according to claim 1, is characterized in that, In the fault diagnosis model construction step, the fan shaft multi-image input fault diagnosis model realizes the fault diagnosis function by learning the fault characteristics of multiple image inputs: the input is two grayscale images with a size of 28*28, and each input is passed through multiple The convolutional layer performs convolution, the output is connected to the fully connected layer, and the multiplication layer is used to multiply the input from the two fully connected layers; the output is the classification output layer, and the activation function is softmax. 8. The multi-signal input fault diagnosis method of fan main shaft based on Glow model according to claim 1, characterized in that, In the fault diagnosis model training step, the fault samples are divided into 19 categories according to the fault category, fault location and damage diameter. 9. The multi-signal input fault diagnosis method of fan main shaft based on Glow model according to claim 8, characterized in that, The vibration signal in the fan shaft vibration data set is a one-dimensional time-series signal. The one-dimensional time-series signal from the fan end and the drive end of the fan shaft is converted into a two-dimensional image signal, and the generation of fault samples and fault diagnosis are performed through a two-dimensional convolutional network; The original 1D time-domain signal needs to be normalized before sample construction.