CN116070083A - An Intelligent Identification Algorithm for Wire Rope Faults Based on Convolutional Neural Network - Google Patents

Abstract

The invention discloses a method and system for intelligent identification of steel wire rope faults based on a convolutional neural network, belonging to the field of mechanical fault diagnosis. The method includes: first collecting fault data, and collecting different types of original one-dimensional magnetic flux leakage signals through sensors; The damaged original one-dimensional magnetic flux leakage signal is transformed into two-dimensional time-frequency image by continuous wavelet transform to preserve the time-domain information and frequency-domain information of the signal; the two-dimensional time-frequency image is divided into 70% training set and 30% validation set The present invention builds a convolutional neural network model, after setting relevant model structures and parameters, loads the training set to carry out model training, constantly adjusts the model structure in the training process to obtain the optimal convolutional neural network model; finally The trained model is used to diagnose and identify wire rope faults, and the system generates fault categories and identification probabilities. This method can be used to identify various internal and external damages of steel wire ropes, reduce manual participation in traditional identification methods, and improve the accuracy of fault identification.

Description

An intelligent wire rope fault identification algorithm based on convolutional neural network

Technical Field

The invention relates to the technical field of fault diagnosis, and more specifically to an intelligent recognition algorithm for wire rope faults based on a convolutional neural network.

Background Art

Wire ropes are often used as the main load-bearing components in many work scenarios, such as mine ropeway transportation, cable-stayed bridges, elevator operations, etc. Harsh working environments will inevitably cause various types of damage to wire ropes, which is directly related to the safety of wire rope use. Nowadays, the demand for wire ropes in various industries is gradually increasing, and various departments are also paying more attention to the safety of wire rope use. Accurately evaluating the service status of wire ropes can not only ensure the safety of production, but also reduce the cost of wire rope usage and improve corporate benefits. At present, wire rope fault diagnosis requires manual participation with certain professional knowledge and the accuracy is not high, causing many companies to still use manual inspection and regular rope replacement methods to ensure the safety of wire rope use. However, manual inspection is affected by the status of the staff and it is difficult to detect internal damage, and regular replacement will cause huge waste. Therefore, it is of great significance to study an intelligent wire rope fault diagnosis method with high accuracy, high reliability and the ability to identify internal damage.

The Chinese patent with publication number CN 109019210 A discloses a tail rope health monitoring system and method for a hoisting system based on a convolutional neural network. Its technical content is: the image acquisition system is used to acquire the status of the tail rope, and transmit the status of the tail rope to the host computer through a mobile wireless sensor network. The host computer conducts in-depth mining of the image data, and performs fault analysis and early warning for the status of the tail rope. The technical problem is that the damage characteristics of the wire rope acquired by the image acquisition system are limited to the surface damage characteristics of the wire rope, and the deep damage characteristics of the wire rope cannot be acquired, so the internal damage characteristics of the wire rope cannot be detected; the image acquisition system can only acquire the surface damage characteristics of the wire rope facing the front of the camera, so it cannot detect the damage characteristics of the wire rope on the other side of the wire rope; the image acquisition system is greatly affected by the light source and environment, so the stability of acquiring the surface damage characteristics of the wire rope is poor, and the consistency of image acquisition cannot be guaranteed;

The Chinese patent with publication number CN 114275483 A discloses an intelligent online monitoring system for belt conveyors. Its technical content is: the sensor collects images of the conveyor belt and collects temperature data and vibration data of each target device at the same time; weak magnetic nondestructive testing is used to detect the damage of the inner core of the steel wire rope in the conveyor belt; the edge intelligent terminal is used to analyze the image, temperature data and vibration data and upload them to the cloud server; the cloud server uses the CNN convolutional neural network model to fuse and analyze various data to obtain the health status of the belt conveyor. The technical problems it has are: in this patent, the collected abnormal information of the steel wire rope inner core needs to be pre-processed by manual noise reduction. The traditional method of processing abnormal information of the steel wire rope inner core requires a lot of professional experience and manual participation, which has certain limitations and is not conducive to the generalization of its system; in addition, the image conversion process of the collected abnormal information of the steel wire rope inner core is not specified.

Therefore, the present invention proposes a fault diagnosis method based on continuous wavelet transform and convolutional neural network that can intelligently identify internal and external damage of wire ropes. This method is of great significance for solving the problems of wire rope fault diagnosis that require manual participation, low accuracy, and internal and external damage.

Summary of the invention

The technical problem to be solved by the present invention is to provide an intelligent wire rope fault recognition algorithm based on convolutional neural network, which can solve the problems that traditional recognition methods require manual participation and have low accuracy.

In order to solve the above technical problems, the present invention adopts the following technical means:

An intelligent wire rope fault identification algorithm based on convolutional neural network includes the following steps:

(1) Signal acquisition: The original one-dimensional magnetic flux leakage signal of different wire rope damage specimens is collected through the wire rope damage detection sensor;

(2) Signal preprocessing: After obtaining the original one-dimensional magnetic leakage signal of the damage, the magnetic leakage signal of each location is cut out from the original signal, and the original magnetic leakage signal is analyzed in time and frequency using continuous wavelet transform to generate the corresponding two-dimensional time and frequency image;

(3) The convolutional neural network training and learning is divided into a training set and a validation set: the present invention divides the two-dimensional time-frequency image generated in step (2) into: a training set of 70% and a validation set of 30%;

(4) Constructing a convolutional neural network model: The basic structure of the convolutional neural network model is a convolutional layer, a pooling layer, a fully connected layer, and a SoftMax layer. The convolutional neural network is trained and classified using a two-dimensional time-frequency image input. All model parameters are randomly initialized. The model structure and parameters are continuously adjusted through the training set to reduce the error between the predicted label and the actual label. Finally, the model parameters are determined to complete the construction of the wire rope fault intelligent identification model.

(5) The trained convolutional neural network model is used in the wire rope fault identification system, and the signal collected by the wire rope damage detection sensor is input and processed by the convolutional neural network model of step (4).

As a further improvement to this technical solution:

In the step (2), the original one-dimensional magnetic flux leakage signal is converted into a two-dimensional time-frequency image to retain the time domain information and frequency domain information of the signal. The continuous wavelet transform of the converted signal x(t) can be expressed as:

In the formula, x(t) represents the converted signal; ψ(t) is the mother wavelet, ψ ^* (t) is the complex conjugate of the function ψ(t); a is the scale factor, and b is the translation factor; after wavelet transform, the one-dimensional signal x(t) is decomposed into a series of wavelet coefficients related to the scale factor a and the translation factor b, and then converted into a two-dimensional time-frequency image.

In the step (2), the image resolution of the convolutional neural network input two-dimensional time-frequency image is a 3-channel image of 224*224, the batch size is 21, the initial learning rate is 0.0001, the optimization function uses adam, the number of training rounds is 12, and the activation function is ReLU.

In the step (3), all the two-dimensional time-frequency images generated by the step (2) are divided into 70% and 30%. 70% of the two-dimensional time-frequency images are used for convolutional neural network training and learning to form usable feature information, that is, 70% of the data is used as a training set. The remaining 30% of the two-dimensional time-frequency images, that is, 30% of the data, which have formed usable feature information to identify, are used as a verification set.

In the step (4), neurons in the same convolution layer can share their weights, and multiple weights can form a convolution kernel; the convolution layer has multiple convolution kernels that can simultaneously extract different feature information; the convolution kernel performs a convolution operation on the input image and obtains a final output value based on a nonlinear activation function; the convolution layer uses the back propagation in the derivation network to update the forward direction and the convolution kernels on the feature maps of the layers below it form a feature map, that is, the convolution kernel scans the entire input image to generate a feature map, and the convolution kernel can be regarded as a feature extractor, and different convolution kernels represent different feature extraction operations; the feature value Z _k of the kth feature map before the nonlinear transformation can be expressed as:

表示二维卷积。Where _Wk represents the kth convolution kernel, _bk represents the offset term, x is the input image of this convolution layer,

Represents a two-dimensional convolution.

The pooling layer is usually placed after the convolution layer. It reduces the dimension of the feature map generated by the convolution by downsampling. The pooling layer can divide the feature map into multiple small areas to generate a new eigenvalue vector, which can be expressed as:

Where down(·) is the downsampling function; yi _,j,k represents the kth new feature map after pooling; Ri _,j is the area near the position (i,j), that is, the receptive field of the pooling; _xm,n,k represents the node in the receptive field; the pooling function used in this algorithm is the maximum pooling function, that is, the maximum value in the pooling domain is selected as the new feature value.

After the convolution layer and the pooling layer, there is a fully connected layer and a SoftMax layer, which is used to classify or logistically regress the results of the completed convolution layer and the pooling layer; the fully connected layer can be applied to a variety of classification models to divide T categories. The number of neurons in the fully connected layer is set to T. The output layer takes the concurrent feature map of the next layer as input, and uses f _v to represent the feature vector, then:

O＝f( _bo + _{wofv )} (4)

Where b _o is the bias vector and w _o represents the weight matrix.

b_o和w_o均为可学习参数，而梯度下降可用反向传播算法的卷积方式，从而使卷积神经网络自己不断训练学习，对于大批量的数据处理，底层优化算法的计算复杂度大大增加，而随机梯度下降法在分析大批量数据上具有高效率的优点，利用梯度下降可以使经验风险E_n(f_w)最小化，每次迭代都根据E_n(f_w)的梯度更新权重w，有：Convolutional neural network training and learning is based on gradient descent.

Both b _o and w _o are learnable parameters, and gradient descent can use the convolution method of the back-propagation algorithm, so that the convolutional neural network can continuously train and learn. For large-scale data processing, the computational complexity of the underlying optimization algorithm is greatly increased, and the stochastic gradient descent method has the advantage of high efficiency in analyzing large-scale data. The empirical risk _En (f _w ) can be minimized by using gradient descent. Each iteration updates the weight w according to the gradient of _En (f _w ), as follows:

Where γ represents the gain, and its value can be selected appropriately according to the specific situation. Under the assumption of sufficient regularity, when the initial estimate _w0 is close enough to the optimal value and when the gain γ is small enough, the algorithm reaches linear convergence.

A better optimization algorithm can be designed by replacing the scalar gain γ with a positive definite matrix Γ _t , which approximates the inverse of the Hessian matrix at the optimal point:

Second-order gradient descent is a well-known Newton algorithm. Under a sufficiently optimistic regularity assumption, when w ₀ is close enough to the optimal, second-order gradient descent reaches quadratic convergence.

Compared with the prior art, the present invention adopting the above technical solution has the following outstanding features:

(1) Optimization of signal preprocessing process

Traditional methods require manual preprocessing of signals such as denoising, which requires professional experience and consumes a lot of energy. The present invention introduces continuous wavelet transform to convert the one-dimensional original signal of wire rope broken wire damage into a two-dimensional time-frequency image, retaining the signal time-frequency information and optimizing the signal preprocessing process.

(2) Adaptive feature extraction method based on CNN

Traditional fault identification is based on manual feature extraction, which results in incomplete information extraction and low classification accuracy. The currently proposed recognition models are mainly shallow network models such as BP neural network and its improved network, support vector machine, and extreme learning machine. The learning ability of these recognition models is limited, and the recognition effect of wire rope damage depends on manually extracted features. The present invention proposes an adaptive feature extraction method based on convolutional neural network to mine deep features in wire rope damage signals. The convolutional neural network model is used to automatically extract features from the two-dimensional time-frequency image of the damage, and gradually fuses and optimizes them to form abstract features that are conducive to classification. This breaks through the limitations of traditional manual feature extraction methods and provides more discriminative features for wire rope damage identification.

(3) Wire rope fault intelligent identification system based on CNN

Aiming at the low accuracy and generalization of the existing broken wire quantitative identification model, a quantitative identification method for broken wire damage in wire rope based on convolutional neural network is proposed. The two-dimensional time-frequency image converted from the leakage magnetic signal of broken wire damage is used as the input of convolutional neural network, and the convolutional neural network with different structural parameters is tested. The optimal convolutional neural network recognition model is established, and high-precision identification of different wire rope damage is achieved.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is the typical structure of a convolutional neural network.

FIG. 2 is an overall flow chart of the present invention.

Figure 3 is an intelligent wire rope fault identification system based on convolutional neural network.

Figure 4 shows the locations of broken wires in the wire rope, wherein Figure 4a shows the locations of broken wires outside the wire rope, which are 1, 2, 3, 4 and 5 broken wires respectively; Figure 4b shows the locations of broken wires inside the wire rope, which are 1, 2 and 3 broken wires respectively.

FIG5 is a flow chart of signal acquisition.

Figure 6 is a two-dimensional time-frequency image of various types of broken wires in a 24mm diameter steel wire rope.

Figure 7 shows the classification results of various types of broken wires in a 24mm diameter wire rope using the selected model.

Figure 8 is the visualization result of the adaptive feature extraction of the convolutional neural network.

Figure 9 is the confusion matrix of broken wire classification of 24mm diameter CNN network.

Figure 10 is the confusion matrix for the classification of broken wires in the BP network with a diameter of 24 mm.

DETAILED DESCRIPTION

Time-frequency imaging is a technique that converts signal frequency and time series data into two-dimensional time-frequency images. Time-frequency conversion can provide insight into deeper features of the original signal, which is conducive to fault classification. Wavelet transform is a commonly used signal processing method in the field of fault diagnosis. This method is very suitable for processing leakage magnetic mutation signals caused by wire rope damage. There are many methods for time-frequency conversion of one-dimensional signals, among which wavelet transform is very suitable for processing leakage magnetic mutation signals caused by wire rope damage. Therefore, continuous wavelet transform is an effective method for converting one-dimensional signals of wire rope damage into two-dimensional time-frequency images. Since its proposal, deep learning has attracted attention from various fields, among which convolutional neural networks have achieved many successful applications in the field of fault diagnosis. As a deep learning model, the convolutional neural network model has a typical structure shown in Figure 1, and its effect on image feature classification is remarkable. The network can automatically extract the feature information of the input image. After multi-level network processing, redundant feature information is removed and abstract features that are easy to classify are formed at a higher level, thereby achieving high-precision classification of image categories.

FIG. 2 shows the overall process of the present invention.

As shown in FIG3 , a wire rope fault intelligent identification system based on a convolutional neural network is shown. Taking a wire rope with a diameter of 24 mm as an example, an intelligent wire rope fault identification algorithm based on a convolutional neural network in this embodiment is as follows: (1) Signal acquisition: The original one-dimensional magnetic leakage signal of different wire rope damaged specimens is collected by a wire rope damage detection sensor; (2) Signal preprocessing: After obtaining the original one-dimensional magnetic leakage signal of the damage, the magnetic leakage signal of each location is cut out from the original signal, and the original magnetic leakage signal is subjected to time-frequency analysis using continuous wavelet transform to generate a corresponding two-dimensional time-frequency image; (3) Convolutional neural network training and learning are divided into a training set and a validation set: The present invention divides the two-dimensional time-frequency image generated in step (2) into: 70% for the training set and 30% for the validation set; (4) Convolutional neural network model is constructed: The basic structure of the convolutional neural network model is a convolution layer, a pooling layer, a fully connected layer and a SoftMax layer. The two-dimensional time-frequency image input by the convolutional neural network is trained and classified. All model parameters are randomly initialized. The model structure and parameters are continuously adjusted through the training set to reduce the error between the predicted label and the actual label. Finally, the parameters of the model are determined to complete the construction of the intelligent recognition model of wire rope faults; (5) The trained convolutional neural network model is used in the wire rope fault recognition system, and the signal collected by the wire rope damage detection sensor is input and processed by the convolutional neural network model of step (4). It includes the following steps:

Step 1: Signal acquisition: The original one-dimensional magnetic leakage signals of different wire rope damage specimens are collected through the wire rope damage detection sensor. The main research object of the present invention is the intelligent fault identification of 24mm diameter wire rope. In order to ensure the diversity and richness of the experimental data collection, 5 external broken wires and 3 internal broken wires are set to experimentally verify the rationality of the invention principle. The specific position of the broken wire is shown in Figure 4. The signal acquisition process is shown in Figure 5. The original one-dimensional magnetic leakage signals of 8 types of internal and external broken wires are mainly collected by sensors, where 100 samples are collected for each broken wire signal, and 8 types of broken wire signals are 800 samples in total. Labels are named for various wire rope faults, and the fault descriptions of 8 types of labels are shown in Table 1.0.

Table 1.0. Fault label naming

Step 2: Signal preprocessing: After obtaining the original one-dimensional leakage magnetic flux signal of the damage, the leakage magnetic flux signal of each location is cut out from the original signal, and the original leakage magnetic flux signal is analyzed in time and frequency using continuous wavelet transform to generate the corresponding two-dimensional time-frequency image. In order to extract the features of the original one-dimensional leakage magnetic flux signal of the 24mm diameter steel wire rope with high accuracy, the leakage magnetic flux signal of each broken wire is separated from the original one-dimensional leakage magnetic flux signal through the 800 original damage signal samples collected, and the separated leakage magnetic flux signal is used as a new sample. The leakage magnetic flux signal of each broken wire is divided into multiple data blocks, each of which contains 1024 data points, skipping the traditional method of artificial signal denoising preprocessing, and converting each data block into a two-dimensional time-frequency image based on the continuous wavelet transform method as the input of the convolutional neural network. The advantage of continuous wavelet transform is that it can retain both the time domain information and the frequency domain information of the signal, and will not lose the characteristic information of the original signal. In order to give full play to the ability of the convolutional neural network model to extract the features of two-dimensional time-frequency images, the experimental data collected are the original one-dimensional magnetic leakage signal, and the time-frequency analysis technology is used to convert the original one-dimensional magnetic leakage signal into a two-dimensional time-frequency image.

Step 3: Divide the training set and the validation set: The convolutional neural network training and learning is divided into a training set and a validation set: The present invention divides the two-dimensional time-frequency image generated in step (2) into: 70% for the training set and 30% for the validation set. The present invention divides the original damaged signal sample into two parts: a training set and a validation set. The divided training set is used in the model training process, and each epoch will output a training set accuracy and loss rate, so as to appropriately adjust the structure and parameters of the model according to the accuracy and loss rate of the training model, thereby gradually improving the model's ability to extract deep features. The validation set is used to comprehensively evaluate the model that has been trained in the training set, and will also output the validation set accuracy and loss rate, and evaluate the rationality of the training model structure and parameters based on the validation set results. The present invention divides the training set and the validation set into a ratio of 70% and 30% respectively.

Step 4: Construct a convolutional neural network model: The basic structure of the convolutional neural network model is a convolutional layer, a pooling layer, a fully connected layer, and a SoftMax layer. The two-dimensional time-frequency image input by the convolutional neural network is used for training and classification. All model parameters are randomly initialized. The model structure and parameters are continuously adjusted through the training set to reduce the error between the predicted label and the actual label. Finally, the parameters of the model are determined to complete the construction of the intelligent recognition model of wire rope faults. Using MATLAB language and Simulink library, a deep learning training model based on CNN is built. The main structure of the convolutional neural network model is: convolutional layer, pooling layer, fully connected layer, and SoftMax layer. The divided two-dimensional time-frequency images of the training set and the validation set are used as the input of the convolutional neural network for training and classification. All model parameters are randomly initialized.

Specifically, the image resolution of the two-dimensional time-frequency image input to the convolutional neural network is a 3-channel image of 224*224, the batch is 21, the initial learning rate is set to 0.0001, the optimization function uses adam, the number of training rounds is 12, and the activation function is ReLU. Because the RGB image input to the two-dimensional convolutional neural network has certain requirements on the size, it is necessary to perform image preprocessing on the two-dimensional time-frequency image converted by the continuous wavelet transform. Since it is characterized by a single-channel grayscale image, the grayscale image is copied to 3 channels and a basis is set in each channel, so that the two-dimensional time-frequency image is converted into a two-dimensional image with 3 channels. The conversion results of each two-dimensional time-frequency image are shown in Figure 6.

Step (4a): The following uses the broken wire data of 24mm diameter steel wire rope to train and verify the convolutional neural network model, and uses the two-dimensional time-frequency images converted from the leakage magnetic signals of 8 types of internal and external broken wires as the input of the convolutional neural network for damage identification, and selects 70% of them for model training and 30% for model verification. In order to achieve the optimal configuration of the deep learning model, convolutional neural network models with various structural parameters are designed, as shown in Table 1.1, which lists 15 models with different structures and parameters.

Table 1.1. Classification results of convolutional neural networks with different structures and parameters

Note: (The format “S _i ,N _i ” represents the “convolution kernel size, number of convolution kernels” of the convolutional layer of the convolutional neural network; “S _Si ” is the size of the pooling window in the pooling layer, and “i” represents the i-th convolutional layer or pooling layer.)

As can be seen from Table 1.1, when the structure and parameters of the convolutional neural network model are different, the accuracy of detecting the broken wire leakage signal of the wire rope is also different, but the accuracy of the number 15 reached 99.58%, so the CNN model consisting of three convolutional layers, three maximum pooling layers and one fully connected layer was selected as the intelligent recognition model for wire rope faults. At this time, the various structures and parameters of the model are saved, and the detailed parameters are shown in Table 1.2. That is, a convolutional neural network quantitative recognition model consisting of three convolutional layers, three maximum pooling layers and one fully connected layer is built in MATLAB language based on the Simulink library.

Table 1.2. Specific parameters of selected convolutional neural networks

As shown in the table, model 15 uses a three-layer two-dimensional convolution structure with a convolution kernel size of 11. The number of convolution kernels in the three layers is 8, 12, and 12 respectively, and the pool window size is 2. The pooling layer performs dimensionality reduction processing on the feature information extracted by the convolution layer after the convolution layer. The maximum pooling function is used, the pooling size is 2×2, and the step size is 2. Finally, the fully connected layer outputs 8 variables, which correspond to 8 types of broken wires in wire ropes. The training curves and error curves of 8 types of broken wires collected by the model 15 are shown in Figure 7. It can be concluded that the training curves and error curves tend to be stable after 12 rounds of model training, with an accuracy rate of up to 99.58% and an error rate of only 0.44%.

Step (4b): The structure and parameter settings of the selected model No. 15 are all completed. After the two-dimensional time-frequency images of the eight types of wire rope faults are processed by the multi-layer convolutional neural network model, the characteristic information is visualized using the t-SNE algorithm, as shown in Figure 8. The result forms eight obvious and easily classified damage features. This shows that the intelligent recognition of wire rope faults based on convolutional neural networks can accurately classify wire rope damage features.

Step (4c): In order to further verify the classification results of wire rope faults, the two-dimensional time-frequency images of 8 types of wire rope faults are visualized again based on the error matrix, and the analysis results are shown in Figure 9. The rows in the error matrix represent the predicted data (output class), the columns represent the actual data (target class), the diagonal units are correct values, and the non-diagonal units represent the misclassified data. The number and percentage of the classification results are displayed in each unit of the error matrix. It can be seen from the figure that the convolutional neural network model significantly divides the fault types of 24mm diameter wire ropes. Only the first fault type has an incorrect label, but it has no effect on the classification results overall. The classification accuracy of the other seven fault types has reached 100%. It further illustrates that the intelligent identification of wire rope faults based on convolutional neural networks can accurately classify the two-dimensional time-frequency images converted from various wire rope broken wire signals.

Step (4d): Compare the method proposed in the present invention with the existing traditional wire rope fault classification and identification method BP neural network. The 8 types of internal and external damage signals of the wire rope with a diameter of 24mm collected in step 1 are used as the input signals of the BP neural network, and the training results of the BP network model are expressed in the error matrix, as shown in Figure 10. It can be seen that the final classification accuracy of the BP neural network is only 84.2%, and the overall accuracy of the convolutional neural network classification is 99.6%. BP is much lower than the accuracy obtained by CNN. Among them, the classification results of the first and fourth fault types have a large loss rate, which is 48.6% and 48.1% respectively. It can also be seen from other units of the error matrix that the first and second fault types are confused during classification, and the fourth fault type is also confused with the third type. Comparing the error matrix Figure 9 with Figure 10, the intelligent identification of wire rope faults based on convolutional neural networks used in the present invention has significant advantages in both accuracy and loss rate.

Step 5: Use the trained convolutional neural network model in the wire rope fault identification system, input the signal collected by the wire rope damage detection sensor, and process it through the convolutional neural network model of step (4).

The above is only a preferred implementation case of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims (8)

1. A wire rope fault intelligent identification algorithm based on convolutional neural network, characterized by comprising the following steps: (1) Signal acquisition: The original one-dimensional magnetic flux leakage signal of different wire rope damage specimens is collected through the wire rope damage detection sensor; (2) Signal preprocessing: After obtaining the original one-dimensional magnetic leakage signal of the damage, the magnetic leakage signal of each location is cut out from the original signal, and the original magnetic leakage signal is analyzed in time and frequency using continuous wavelet transform to generate the corresponding two-dimensional time and frequency image; (3) The convolutional neural network training and learning is divided into a training set and a validation set: the present invention divides the two-dimensional time-frequency image generated in step (2) into: a training set of 70% and a validation set of 30%; (4) Constructing a convolutional neural network model: The basic structure of the convolutional neural network model is a convolutional layer, a pooling layer, a fully connected layer, and a SoftMax layer. The convolutional neural network is trained and classified using a two-dimensional time-frequency image input. All model parameters are randomly initialized. The model structure and parameters are continuously adjusted through the training set to reduce the error between the predicted label and the actual label. Finally, the model parameters are determined to complete the construction of the wire rope fault intelligent identification model. (5) The trained convolutional neural network model is used in the wire rope fault identification system, and the signal collected by the wire rope damage detection sensor is input and processed by the convolutional neural network model of step (4). 2. According to the convolutional neural network-based intelligent wire rope fault identification algorithm of claim 1, it is characterized in that in the step (2), the original one-dimensional magnetic leakage signal is converted into a two-dimensional time-frequency image to retain the time domain information and frequency domain information of the signal, and the continuous wavelet transform of the converted signal x(t) can be expressed as:

In the formula, x(t) represents the converted signal; ψ(t) is the mother wavelet, ψ ^* (t) is the complex conjugate of the function ψ(t); a is the scale factor, and b is the translation factor; after wavelet transform, the one-dimensional signal x(t) is decomposed into a series of wavelet coefficients related to the scale factor a and the translation factor b, and then converted into a two-dimensional time-frequency image. 3. According to the convolutional neural network-based intelligent wire rope fault identification algorithm of claim 1, it is characterized in that, in the step (2), the image resolution of the convolutional neural network input two-dimensional time-frequency image is a 3-channel image of 224*224, the batch is 21, the initial learning rate is 0.0001, the optimization function uses adam, the number of training rounds is 12, and the activation function is ReLU. 4. According to claim 1, the wire rope fault intelligent identification algorithm based on convolutional neural network is characterized in that, in the step (3), the two-dimensional time-frequency images generated by step (2) are divided into 70% and 30%, and 70% of the two-dimensional time-frequency images are used for convolutional neural network training and learning to form usable feature information, that is, 70% of the data is used as a training set, and the usable feature information has been formed to identify the remaining 30% of the two-dimensional time-frequency images, that is, 30% of the data is used as a verification set. 5. According to claim 1, the wire rope fault intelligent identification algorithm based on convolutional neural network is characterized in that, in the step (4), neurons in the same convolution layer can share their weights, and multiple weights can form a convolution kernel; the convolution layer has multiple convolution kernels that can simultaneously extract different feature information; the convolution kernel performs a convolution operation on the input image and obtains a final output value based on a nonlinear activation function; the convolution layer uses the back propagation in the derivation network to update the forward direction and the convolution kernel on the feature map of each layer below it to form a feature map, that is, the convolution kernel scans the entire input image to generate a feature map, the convolution kernel can be regarded as a feature extractor, and different convolution kernels represent different feature extraction operations; the feature value Z _k of the kth feature map before the nonlinear transformation can be expressed as:

Where _Wk represents the kth convolution kernel, _bk represents the offset term, x is the input image of this convolution layer,

Represents a two-dimensional convolution. 6. According to the convolutional neural network-based intelligent wire rope fault identification algorithm of claim 5, it is characterized in that the pooling layer is generally after the convolution layer, and the feature map generated by the convolution is reduced in dimension by downsampling, and the pooling layer can divide the feature map into multiple small areas to generate a new eigenvalue vector, which can be expressed as:

Where down(·) is the downsampling function; yi _,j,k represents the kth new feature map after pooling; Ri _,j is the area near the position (i,j), that is, the receptive field of the pooling; _xm,n,k represents the node in the receptive field; the pooling function used in this algorithm is the maximum pooling function, that is, the maximum value in the pooling domain is selected as the new feature value. 7. A wire rope fault intelligent identification algorithm based on convolutional neural network according to claim 6, characterized in that after the convolution layer and the pooling layer, there are a fully connected layer and a SoftMax layer, which are used to classify or logistically regress the results generated by the completed convolution layer and the pooling layer; the fully connected layer can be applied to a variety of classification models to divide T categories, the number of neurons in the fully connected layer is set to T, the output layer takes the concurrent feature map of its next layer as input, and the feature vector is represented by f _v , then: O＝f( _bo + _{wofv )} (4) Where b _o is the bias vector and w _o represents the weight matrix. 8. According to the convolutional neural network-based intelligent wire rope fault identification algorithm of claim 1, it is characterized in that, as in step (4), the convolutional neural network training and learning is completed based on gradient descent, and the convolutional neural network model

Both b _o and w _o are learnable parameters, and gradient descent can use the convolution method of the back-propagation algorithm, so that the convolutional neural network can continuously train and learn. For large-scale data processing, the computational complexity of the underlying optimization algorithm is greatly increased, and the stochastic gradient descent method has the advantage of high efficiency in analyzing large-scale data. The empirical risk _En (f _w ) can be minimized by using gradient descent. Each iteration updates the weight w according to the gradient of _En (f _w ), as follows:

In the formula, γ represents the gain, and its value can be selected appropriately according to the specific situation. Under the assumption of sufficient regularity, when the initial estimate _w0 is close enough to the optimal value and when the gain γ is small enough, the algorithm reaches linear convergence; A better optimization algorithm can be designed by replacing the scalar gain γ with a positive definite matrix Γ _t , which approximates the inverse of the Hessian matrix at the optimal point:

Second-order gradient descent is a well-known Newton algorithm. Under a sufficiently optimistic regularity assumption, when w ₀ is close enough to the optimal, second-order gradient descent reaches quadratic convergence.

Images