CN116858531A - Fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt - Google Patents

Abstract

A fan gear box fault diagnosis method based on data enhancement and CSP-ResNeXt comprises the following steps: collecting fault vibration signal data collected by different fault types of the gearbox by using an acceleration sensor arranged on the gearbox; decomposing the collected fault vibration signals by adopting a wavelet packet decomposition method to obtain different wavelet packet coefficients, randomly selecting a group of wavelet coefficients, reducing the wavelet coefficients into time domain signals after distortion to expand the fault samples, and finishing the data enhancement operation of the fault samples; the RPM converts the time domain signal into a gray scale map through the relative position matrix Relative Position Matrix and inputs the gray scale map into the built CSP-ResNeXt network to train the diagnostic model. Finally, an intelligent fault diagnosis model of the fan gear box can be obtained, and vibration signals acquired in real time can be input by using the model to carry out fault diagnosis and identification. The invention uses wavelet packet distortion technology to enhance data, adopts an improved ResNeXt network, and realizes accurate fault identification of the fan gear box under serious unbalance of fault samples.

Description

Fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt

Technical Field

The invention relates to the technical field of health state monitoring of wind turbine generator gear boxes, in particular to a fan gear box fault diagnosis method based on data enhancement and CSP-ResNeXt.

Background

The wind turbine generator is operated in environments such as extreme temperature, heavy rain, snow storm, salt fog and the like for a long time, and as the operation time is increased, the fatigue strength, operation performance and the like of blades, main bearings, gearboxes, generators and other components are continuously reduced, so that abnormality and faults are caused, and the fan is abnormally operated and even stopped. The fan gear box is analyzed aiming at the fault components, and the proportion of the fan gear box in all the fault components is highest, so that the establishment of a fault diagnosis model for the fan gear box is very important.

In general, vibration signals are often adopted as signal sources for fault analysis aiming at rotating parts, strong noise signals are doped in the signal acquisition process, and the signals have the characteristics of nonlinearity and non-stationarity, so that the fault diagnosis is challenged. Meanwhile, vibration signals which are collected by the acceleration sensor and are in a normal state are collected by the wind turbine, and a few fault signal collection samples are used, so that the fault diagnosis model training is difficult.

Disclosure of Invention

The invention aims to solve the technical problems of providing a fan gear box fault diagnosis method based on data enhancement and CSP-ResNeXt, solving the problems that the energy of a fan gear box fault vibration signal is weak and is often polluted by strong noise, and the traditional fault diagnosis adopts the manual fault feature extraction of a time domain and frequency domain analysis method, so that the subjectivity is realized, and the diagnosis effect is poor.

In order to solve the technical problems, the invention adopts the following technical scheme:

a fan gear box fault diagnosis method based on data enhancement and CSP-ResNeXt comprises the following steps:

step1, collecting fault vibration signal data collected by different fault types of the gear box by using an acceleration sensor arranged on the gear box;

step2, expanding sample data by adopting wavelet packet distortion;

step3, converting the fault vibration signal into a gray level map through a relative position matrix;

step4, inputting the data after sample expansion into an improved ResNeXt network, and finally obtaining an intelligent fault diagnosis model of the fan gear box, and inputting vibration signals acquired in real time by using the model to perform fault diagnosis and identification.

The specific steps of Step2 are as follows:

step2.1, performing 2-layer decomposition on the original vibration signal by adopting wavelet packet transformation;

step2.2, randomly selecting a group of wavelet coefficients to carry out distortion operation;

step2.3, reducing the wavelet packet decomposition coefficient after distortion operation into a time domain signal.

In step2.1 described above, a 2-layer decomposition of the original vibration signal is achieved using the following function:

where i=0, (2 j-1), k and L are sequential indices of elements, (h, g) is a pair of finite impulse response filters of length L, let w be _j,i Ith group of wavelet coefficients representing jth decomposition level, w _0,0 Is the original signal.

The distortion function of the distortion operation in step2.2 described above is:

wherein d is a distortion coefficient randomly selected from a preset range, sign is a sign function, and in mathematical and computer operations, the function is to take a certain number of signs (positive or negative): when x >0, sign (x) =1; when x=0, sign (x) =0; when x <0, sign (x) = -1; abs is an absolute function, returning the absolute value of the w value.

The specific steps of Step4 are as follows:

step4.1, the input data passes through a convolution kernel with a field of view of 7×7, a dimension number of 64, and a step length of 2;

the output data of Step4.2 and Step4.1 pass through a global maximum pooling layer with convolution kernel of 3 multiplied by 3 and step length of 2; the step4.3 and step4.2 output data enter a backbone network consisting of four different layers of ResNeXt residual blocks.

The backbone network composed of four ResNeXt residual blocks of different layers in the foregoing Step4.3 has the following composition structure:

one part of the three-layer joint structure consists of 3 layers of residual structures with the base number of 32 and convolution kernels of 1 multiplied by 1,3 multiplied by 3 and 1 multiplied by 1, and comprises a residual connection bypass, and finally, the channels are spliced and output;

the second part is composed of residual structures with 3 layers of convolution kernels of 1 multiplied by 1,3 multiplied by 3 and 1 multiplied by 1, the number of the base numbers of the layers is 32, and the residual structures comprise a residual connection bypass, and finally channels are spliced and output;

the third part is composed of residual structures with 3 layers of convolution kernels of 1 multiplied by 1,3 multiplied by 3 and 1 multiplied by 1, the number of the base numbers of the layers is 32, and the residual structures comprise a residual connection bypass, and finally channels are spliced and output;

the fourth part is composed of a residual structure with 3 layers of convolution kernels of 1×1,3×3 and 1×1 and with a base number of 32, and comprises a residual connection bypass, and finally channels are spliced and output.

The four main network residual blocks formed by ResNeXt residual blocks of different layers are stacked in a 3-4-6-3 way, CSP modules are inserted into each part of network structure, input channels are grouped when entering each part of stacked network, one half of the input channels enter the stacked network, and the other half of the input channels enter the CSP modules; then splicing the output channel of each part with the output channel of the CSP module to be used as the total output of each part of stacked network structure; and finally, summarizing the output of the fourth part, and entering a global average pooling layer with the convolution kernel of 1 multiplied by 1 and a full connection layer to realize fault classification.

The ResNeXt residual block activation function adopts a sigmoid function, and the calculation formula is as follows:

where x is the input and f (x) is the output.

The activation function used by each convolution kernel back and transition layer in the four ResNeXt residual block network parts adopts a Relu activation function, and the calculation formula is as follows:

f(x)＝max(0,x)

where x is the input and f (x) is the output.

The improved ResNeXt network, namely the CSP-ResNeXt network, adopts a cross entropy loss function as a full connection layer loss function, and the calculation formula is as follows:

wherein p (x) _i ) And q (x) _i ) Respectively representing a true probability distribution and a predicted probability distribution, and H (p, q) represents the difference between the predicted value and the true value;

the cross entropy loss function is matched with the softmax classifier to be used, the output result is processed at the full connection layer, the sum of the predictive values of a plurality of classifications is 1, and then the loss is calculated through the cross entropy, wherein the softmax function has the calculation formula:

wherein x is _i The output of the upper layer of the model is used as the input of a softmax classifier; the output calculation result softmax (x) may be regarded as the confidence that the predicted result is a true result.

The fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt provided by the invention has the following beneficial effects:

1. wavelet packet distortion is developed based on wavelet packet transforms to increase the number of false training samples. Here, the enhanced samples are similar to the original samples, but have different values. Thus, the balance among classes can be achieved, and the sample diversity of the training data set can be improved.

The CSP-ResNeXt introduces a grouping convolution idea on the basis of the original Resnet, so that the classification precision of a network structure is higher under the condition of the same complexity, and meanwhile, the generalization capability of a model is stronger due to the idea of adopting the stacked structure normalization of VGG in each residual block in different application scenes; after the CSP module is adopted, the calculation amount of the model is reduced, the gradient combination is richer, the repeated learning of the same gradient value in the forward gradient propagation of the model is avoided, and the convergence capacity of the model in training is enhanced.

Drawings

The invention is further illustrated by the following examples in conjunction with the accompanying drawings:

FIG. 1 is a schematic overall flow chart of an embodiment of the present invention;

FIG. 2 is a residual structure design;

FIG. 3 is a graph of the transformed graphical differential field signature in an embodiment;

FIG. 4 is a schematic diagram of a model training loss function curve in an embodiment;

FIG. 5 is a schematic diagram of an accuracy curve of model training in an embodiment.

Detailed Description

The technical scheme of the invention is described in detail below with reference to the accompanying drawings and examples.

1-2, a fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt includes the steps of:

step1, collecting fault vibration signal data collected by different fault types of the gear box by using an acceleration sensor arranged on the gear box;

step2, expanding sample data by adopting wavelet packet distortion;

step3, converting the fault vibration signal into a gray level map through a relative position matrix;

step4, inputting the data after sample expansion into an improved ResNeXt network, and finally obtaining an intelligent fault diagnosis model of the fan gear box, and inputting vibration signals acquired in real time by using the model to perform fault diagnosis and identification.

In Step1, in terms of vibration signal extraction and original data set construction, at least taking a normal vibration signal, a gear box bearing outer ring fault vibration signal, a gear box bearing inner ring fault vibration signal, a gear failure state fault vibration signal, a pitting state fault vibration signal and a wear state fault vibration signal to form a data set, and marking a data set classification label in a single-heat coding mode, so that a softmax classifier is conveniently adopted for carrying out fault diagnosis type classification in the later stage.

The specific steps of Step2 are as follows:

step2.1, decomposing the fault vibration signal by adopting wavelet packet transformation, wherein the decomposition order is 2, and obtaining a wavelet packet decomposition coefficient;

step2.2, randomly selecting a group of wavelet coefficients to carry out distortion operation;

step2.3, restoring the data into a time sequence signal through wavelet packet inverse transformation, and finishing data enhancement of a fault vibration signal;

also, in step2.3, data is input into a matrix of available relative positions (Relative Position Matrix, RPM) in the modified ResNeXt to convert the data into two-dimensional images to facilitate training of the diagnostic model.

The wavelet packet transformation is used as a classical time-frequency domain analysis method, and can decompose a signal into a group of wavelet coefficients with different frequency bands; the inverse transformation of the wavelet packet is opposite, and the original signal can be restored by reconstructing the wavelet packet decomposition coefficient; if the reconstructed signal is decomposed and reconstructed according to the same decomposition rule, the reconstructed signal is consistent with the original signal; the invention performs distortion processing on a set of randomly selected wavelet coefficients so that the reconstructed signal is similar to but slightly different from the original sample. The reconstructed signal may be used as an enhanced sample, especially for a minority class of vibration signal samples.

In step2.1 described above, a 2-layer decomposition of the original vibration signal is achieved using the following function:

where i=0, (2 j-1), k and L are sequential indices of elements, (h, g) is a pair of finite impulse response filters of length L, let w be _j,i Ith group of wavelet coefficients representing jth decomposition level, w _0,0 Is the original signal.

The distortion function of the distortion operation in step2.2 described above is:

wherein d is a distortion coefficient randomly selected from a preset range, sign is a sign function, and in mathematical and computer operations, the function is to take a certain number of signs (positive or negative): when x >0, sign (x) =1; when x=0, sign (x) =0; when x <0, sign (x) = -1; abs is an absolute function, returning the absolute value of the w value.

After distortion is carried out on part of wavelet coefficients, wavelet packet inverse transformation is carried out to obtain an output signal v, and the formula is expressed as follows:

wherein the method comprises the steps ofFor the post-distortion wavelet coefficients, v is the reconstructed wavelet coefficients, < >>I=0, (2 j-1), k and L are sequential indices of elements for a pair of finite impulse response filters of length L.

In Step3, the data is converted into two-dimensional images using the relative position matrix (Relative Position Matrix, RPM) available in the data input improvement ResNeXt to facilitate training of the diagnostic model, and the following steps are implemented:

the relative position matrix (Relative Position Matrix, RPM) contains the originalTime seriesThe redundant characteristics of the image are used for enabling the similarity information between classes and in the classes to be captured more easily in the converted image; for a time series x= (xt, t=1, 2, …, N), the RPM map can be obtained by:

1) For the original time series, a standard normal distribution Z is obtained by the following Z-score normalization method:

where μ represents the average value of X and σ represents the standard deviation of X.

2) 2) selecting a proper reduction factor k by adopting a Piecewise Aggregation Approximation (PAA) method to generate a new smooth time sequenceReducing dimension N to m:

by calculating the average value of the piecewise constants for dimension reduction, the approximate trend of the original time sequence can be kept, and a new smooth time sequence is finally obtainedIs m.

3) Calculating the relative position between the two time stamps, and sequencing the preprocessed time seriesConversion into a two-dimensional matrix M:

as indicated above, the matrix characterizes the relative positional relationship between every two time stamps in the time series; each row and each column of the sequence is referenced by a certain time stamp, and the information of the whole sequence is further characterized.

4) Finally, M is converted into a gray value matrix by using minimum-maximum normalization, and a relative displacement matrix F is finally obtained:

finally, the time series vibration signals are converted into gray level images through a relative displacement matrix, and the gray level images can be input into a diagnosis classification model for training.

In the construction of CSP-ResNeXt network, the following method should be used for construction: the ResNeXt network structure adopts the VGG stacking idea and the acceptance dividing-converting-merging idea at the same time, and a grouping convolution method is adopted in a basic residual block of the ResNeXt network structure, so that the consistency of the topological structure of each aggregation is kept, the expansibility of a model is enhanced, and meanwhile, the structural design difficulty of the network for a specific data set is also reduced; the CSP module is introduced on the basis, so that the learning capacity of the neural network is enhanced, the calculation bottleneck is eliminated to a certain extent, the memory cost is reduced, and the time consumed by reasoning calculation can be reduced by introducing the CSP module into the ResNeXt network; the CSP module achieves the reduction of the amount of inference computation by cross-phase, achieves this function by optimizing the repeated gradient information in the network, enriches and increases various combinations of gradients by focusing on variability of gradients from the beginning and ending of the network phase integrating feature mapping. The method of introducing the grouping convolution on the basis of the battleneck layer of the ResNet changes the original 1×1,3×3 and 1×1 convolution structure sharing convolution kernel parameters into grouping sharing convolution kernel parameters in respective groups, ensures that the calculated amount is not greatly different, and enhances the learning capacity of the network; meanwhile, in order to improve the speed of reasoning calculation, a CSP local cross-stage method is introduced, the input of a butteleneck layer is divided into two parts, one part is directly spliced with the output obtained by improving the input of the Resnet block through improving the Resnet block, and the multiplexing of redundancy gradients is reduced; the gradient flows are transmitted through different network paths by dividing the gradient flows, and the transmitted gradient information can be greatly different by switching the cascade and conversion steps, so that the multiplexing of redundant gradients is reduced, and the design idea of a residual structure is shown in figure 2.

The specific steps of Step4 are as follows:

step4.1, the input data passes through a convolution kernel with a field of view of 7×7, a dimension number of 64, and a step length of 2;

the output data of Step4.2 and Step4.1 pass through a global maximum pooling layer with convolution kernel of 3 multiplied by 3 and step length of 2; the step4.3 and step4.2 output data enter a backbone network consisting of four different layers of ResNeXt residual blocks.

The backbone network composed of four ResNeXt residual blocks of different layers in the foregoing Step4.3 has the following composition structure:

one part of the three-layer joint structure consists of 3 layers of residual structures with the base number of 32 and convolution kernels of 1 multiplied by 1,3 multiplied by 3 and 1 multiplied by 1, and comprises a residual connection bypass, and finally, the channels are spliced and output;

the second part is composed of residual structures with 3 layers of convolution kernels of 1 multiplied by 1,3 multiplied by 3 and 1 multiplied by 1, the number of the base numbers of the layers is 32, and the residual structures comprise a residual connection bypass, and finally channels are spliced and output;

the third part is composed of residual structures with 3 layers of convolution kernels of 1 multiplied by 1,3 multiplied by 3 and 1 multiplied by 1, the number of the base numbers of the layers is 32, and the residual structures comprise a residual connection bypass, and finally channels are spliced and output;

the fourth part is composed of a residual structure with 3 layers of convolution kernels of 1×1,3×3 and 1×1 and with a base number of 32, and comprises a residual connection bypass, and finally channels are spliced and output.

The four-layer convolution performs feature extraction, and the effect is better by performing feature extraction through four-layer convolution layers with the same base number and convolution kernel.

The four main network residual blocks formed by ResNeXt residual blocks of different layers are stacked in a 3-4-6-3 way, CSP modules are inserted into each part of network structure, input channels are grouped when entering each part of stacked network, one half of the input channels enter the stacked network, and the other half of the input channels enter the CSP modules; then splicing the output channel of each part with the output channel of the CSP module to be used as the total output of each part of stacked network structure; and finally, summarizing the output of the fourth part, and entering a global average pooling layer with the convolution kernel of 1 multiplied by 1 and a full connection layer to realize fault classification.

In Step4, inputting the training set into the built CSP-ResNeXt network for training; setting the size of the super parameter batch to be 128, and adopting an adam random optimization algorithm as a gradient descent optimization algorithm: the learning rate alpha is set to 0.001, and the first moment estimates the attenuation coefficient beta ₁ Set to 0.9, the second moment estimated attenuation coefficient beta ₂ Set to 0.999, the smoothing term ε is set to 1e ^-8 The specific iteration formula is as follows:

m _t ＝β ₁ m _t-1 +(1-β ₁ )g _t

wherein t is the training times; m is m _t And v _t The first and second moment estimates of the gradient, respectively, can be considered as the desired Eg _t ]、E[g _t ² ]Is an approximation of (a); g _t Gradient values calculated for each parameter θ;and->Is to m _t And v _t Such that an unbiased estimate of the desire can be approximated; to prevent the denominator from being 0, a smoothing term ε is set.

The ResNeXt residual block activation function adopts a sigmoid function, and the calculation formula is as follows:

where x is the input and f (x) is the output.

The activation function used by each convolution kernel back and transition layer in the four ResNeXt residual block network parts adopts a Relu activation function, and the calculation formula is as follows:

f(x)＝max(0,x)

where x is the input and f (x) is the output.

The improved ResNeXt network, namely the CSP-ResNeXt network, adopts a cross entropy loss function as a full connection layer loss function, and the calculation formula is as follows:

wherein p (x) _i ) And q (x) _i ) Respectively representing a true probability distribution and a predicted probability distribution, and H (p, q) represents the difference between the predicted value and the true value;

the cross entropy loss function is matched with the softmax classifier to be used, the output result is processed at the full connection layer, the sum of the predictive values of a plurality of classifications is 1, and then the loss is calculated through the cross entropy, wherein the softmax function has the calculation formula:

wherein x is _i The output of the upper layer of the model is used as the input of a softmax classifier; the output calculation result softmax (x) may be regarded as the confidence that the predicted result is a true result.

Examples:

to verify the feasibility of the invention, the method of the invention was implemented using the university of eastern public gearbox dataset.

The experimental algorithm processing platform equipment adopts a high-performance computer model to train and test a software environment of Windows11, selects a Python3.10 programming language, adopts a GPU acceleration library of CUDA11.8, and adopts a deep learning framework of Pytorch2.0.1.

The experimental data set is divided into 5 types, namely health, tooth breakage, gear surface abrasion, gear foot crack and gear root crack, the number of samples in each type is 500, the total number of samples is 2500, and the data set is based on a training set and a test set 8:2, the training set sample number is 2000, and the verification set sample number is 500.

The model network code is:

an example of a graphical differential field signature transformed according to the method of the present invention is shown in figure 3 below.

In the aspect of model training, the number of model iterations is set to 500, the batch size is set to 8, the thread is set to 2, and the model loss function and the accuracy are as shown in the following fig. 4 and 5, it can be seen that when the model is iterated 500 times, the model basically converges, and the accuracy can reach more than 95%.

Claims (10)

1. A fan gear box fault diagnosis method based on data enhancement and CSP-ResNeXt is characterized by comprising the following steps:

step1, collecting fault vibration signal data collected by different fault types of the gear box by using an acceleration sensor arranged on the gear box;

step2, expanding sample data by adopting wavelet packet distortion;

step3, converting the fault vibration signal into a gray level map through a relative position matrix;

step4, inputting the data after sample expansion into an improved ResNeXt network, and finally obtaining an intelligent fault diagnosis model of the fan gear box, and inputting vibration signals acquired in real time by using the model to perform fault diagnosis and identification.

2. The fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt according to claim 1, wherein the specific steps of Step2 are as follows:

step2.1, performing 2-layer decomposition on the original vibration signal by adopting wavelet packet transformation;

step2.2, randomly selecting a group of wavelet coefficients to carry out distortion operation;

step2.3, reducing the wavelet packet decomposition coefficient after distortion operation into a time domain signal.

3. The fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt according to claim 2, wherein in step2.1, the following function is used to implement 2-layer decomposition of the original vibration signal:

where i=0, (2 j-1), k and L are sequential indices of elements, (h, g) is a pair of finite impulse response filters of length L, let w be _j,i Ith group of wavelet coefficients representing jth decomposition level, w _0,0 Is the original signal.

4. The fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt according to claim 2, wherein the distortion function of the distortion operation in the step2.2 is:

wherein d is a distortion coefficient randomly selected from a preset range, sign is a sign function, and in mathematical and computer operations, the function is to take a certain number of signs (positive or negative): when x >0, sign (x) =1; when x=0, sign (x) =0; when x <0, sign (x) = -1; abs is an absolute function, returning the absolute value of the w value.

5. The fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt according to claim 1, wherein the Step4 specifically comprises the following steps:

step4.1, the input data passes through a convolution kernel with a field of view of 7×7, a dimension number of 64, and a step length of 2;

the output data of Step4.2 and Step4.1 pass through a global maximum pooling layer with convolution kernel of 3 multiplied by 3 and step length of 2;

the step4.3 and step4.2 output data enter a backbone network consisting of four different layers of ResNeXt residual blocks.

6. The fan gear box fault diagnosis method based on data enhancement and CSP-ResNeXt according to claim 5, wherein the four main networks consisting of ResNeXt residual blocks of different layers in the step4.3 have the following composition structure:

one part of the three-layer joint structure consists of 3 layers of residual structures with the base number of 32 and convolution kernels of 1 multiplied by 1,3 multiplied by 3 and 1 multiplied by 1, and comprises a residual connection bypass, and finally, the channels are spliced and output;

the second part is composed of residual structures with 3 layers of convolution kernels of 1 multiplied by 1,3 multiplied by 3 and 1 multiplied by 1, the number of the base numbers of the layers is 32, and the residual structures comprise a residual connection bypass, and finally channels are spliced and output;

the third part is composed of residual structures with 3 layers of convolution kernels of 1 multiplied by 1,3 multiplied by 3 and 1 multiplied by 1, the number of the base numbers of the layers is 32, and the residual structures comprise a residual connection bypass, and finally channels are spliced and output;

the fourth part is composed of a residual structure with 3 layers of convolution kernels of 1×1,3×3 and 1×1 and with a base number of 32, and comprises a residual connection bypass, and finally channels are spliced and output.

7. The fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt according to claim 6, wherein the four main network residual blocks formed by ResNeXt residual blocks of different layers are stacked in a structure of 3-4-6-3, CSP modules are inserted into each part of the network structure, input channels are grouped when entering each part of the stacked network, half of the input channels enter the stacked network, and the other half of the input channels enter the CSP modules; then splicing the output channel of each part with the output channel of the CSP module to be used as the total output of each part of stacked network structure; and finally, summarizing the output of the fourth part, and entering a global average pooling layer with the convolution kernel of 1 multiplied by 1 and a full connection layer to realize fault classification.

8. The fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt according to claim 7, wherein the ResNeXt residual error block activation function adopts a sigmoid function, and the calculation formula is as follows:

where x is the input and f (x) is the output.

9. The fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt according to claim 8, wherein the activation function used by each convolution kernel and transition layer in the four ResNeXt residual block network parts adopts a Relu activation function, and the calculation formula is as follows:

f(x)＝max(0,x)

where x is the input and f (x) is the output.

10. The fan gearbox fault diagnosis method based on data enhancement and CSP-ResNeXt according to claim 9, wherein the improved ResNeXt network, namely the full connection layer loss function of the CSP-ResNeXt network, adopts a cross entropy loss function, and the calculation formula is as follows:

wherein p (x) _i ) And q (x) _i ) Respectively representing a true probability distribution and a predicted probability distribution, and H (p, q) represents the difference between the predicted value and the true value;

the cross entropy loss function is matched with the softmax classifier to be used, the output result is processed at the full connection layer, the sum of the predictive values of a plurality of classifications is 1, and then the loss is calculated through the cross entropy, wherein the softmax function has the calculation formula:

wherein x is _i The output of the upper layer of the model is used as the input of a softmax classifier; the output calculation result softmax (x) may be regarded as the confidence that the predicted result is a true result.