Steel wire rope damage width identification method based on residual network multi-channel fusion

Abstract

The invention relates to the technical field of steel wire rope detection, and discloses a steel wire rope damage width identification method based on residual network multi-channel fusion, and the method comprises an offline training step and an online detection step, and the offline training step comprises the steps: S10, taking steel wire ropes with the same damage depth and different damage widths as training samples, and carrying out online detection; acquiring magnetic flux leakage quantity data of the training sample by using a Hall sensor array; and S20, performing continuous wavelet transform on the magnetic flux leakage quantity data to obtain each wavelet component, and then performing moving average filtering denoising on each wavelet component to obtain denoised data. The method has the beneficial effects that the continuous wavelet transform can provide frequency information on different scales, can provide related information of signals in a frequency domain for a classification model, and can describe the time-frequency change of the signals more comprehensively. And meanwhile, classification model input is performed by using the time spectrogram, so that the advantage of extracting features by using a deep learning model can be better utilized, and the model can capture the dynamic characteristics of data.

Description

Technical Field

[0001] The present invention relates to the technical field of wire rope detection, and in particular to a wire rope damage width identification method based on residual network multi-channel fusion. Background Art

[0002] As one of the most commonly used load-bearing tools, wire ropes are often used in various engineering applications such as elevators, transportation, and lifting. However, over time, the in-service wire ropes are subjected to repeated friction and complex stress, which can cause wire breakage, wear, or corrosion. Wire rope defects are divided into two categories. One is the local damage type (LF), which mainly refers to discontinuities in the wire rope, such as broken wires and pits on the surface of the wire. Its characteristic is that the metal cross-sectional area of ​​the wire rope suddenly decreases. The other is the loss of metal cross-sectional area type (LMA), which refers to the sudden decrease in the metal cross-sectional area of ​​the wire rope, which is measured by comparing the detection point with the reference point on the wire rope that symbolizes a larger metal cross-sectional area.

[0003] There are many methods for non-destructive testing of wire ropes in the prior art, such as ultrasonic testing, X-ray testing, electromagnetic testing, machine vision testing, etc. Although the method of target detection by machine vision has gradually become popular, its quantitative detection of internal defects and surface defects of wire ropes is still limited. Two-dimensional image processing based on computer vision can only detect surface defects, which brings great challenges to wire rope defect detection. Therefore, wire rope defect detection by electromagnetic testing is still the mainstream.

[0004] Although in the existing technology, the electromagnetic detection method has good sensitivity and reliability for detecting surface defects of wire ropes, due to the complex testing environment and monitoring conditions of the wire rope, the signals measured by the electromagnetic detection method have various interferences and noises. Feature identification of wire rope defects from the collected signals is still a hot topic in current research.

[0005] In the patent literature, such as the authorization announcement number CN110231395B, the name is a wire rope broken wire damage identification method and system, which performs continuous wavelet transform on the extracted detected leakage magnetic signal, records the peak-to-peak value of the transformed wavelet coefficient; then, imports the recorded peak value and characteristic values ​​such as wire rope diameter into the trained neural network; finally, the parameters output by the neural network are obtained to complete the quantitative damage identification of the number of broken wires in the wire rope. Although this technical solution can perform quantitative damage identification on wire rope broken wire data, this technical solution uses the original features in the one-dimensional data as input, does not consider the frequency changes of the data in different time periods, and cannot capture the dynamic characteristics of the data.

[0006] Therefore, the present invention proposes a wire rope damage width identification method based on residual network multi-channel fusion, and performs LMA defect detection on the wire rope through continuous wavelet transform and residual network. Summary of the invention

[0007] In order to overcome the defects that the leakage magnetic signals measured by the existing electromagnetic detection method in the prior art have various interferences and noises, and the characteristics of wire rope defects cannot be identified from the collected signals, and the defects of wire rope damage cannot be accurately measured, the detailed technical solution is as follows:

[0008] A wire rope damage width recognition method based on residual network multi-channel fusion includes an offline training step and an online detection step. The offline training step includes: S10, using wire ropes with the same damage depth and different damage widths as training samples, and using a Hall sensor array to obtain magnetic leakage data of the training samples; S20, using continuous wavelet transform to obtain each wavelet component of the magnetic leakage data, and then using sliding average filtering to denoise each wavelet component to obtain denoised data; S30, converting the denoised data into a time spectrum diagram, using the time spectrum diagram as input and the damage width of the wire rope as output, training an image classification model based on a residual network to obtain a trained image classification model based on a residual network; the online detection step includes: S40, using a Hall sensor array to obtain the magnetic leakage data of the wire rope to be detected, performing continuous wavelet transform and sliding average filtering denoising on the magnetic leakage data, and obtaining denoised data; S50, inputting the denoised data into an image classification model based on a residual network to obtain the damage width of the corresponding wire rope output by the image classification model based on the residual network.

[0009] Furthermore, in S10, a multi-channel Hall sensor array is used to detect the magnetic leakage data of the steel wire rope, and the magnetic leakage data is converted into a voltage signal to obtain a curve graph of time and voltage.

[0010] Further, in S20, the voltage signal in the time-voltage curve graph is transformed using wavelet functions of different scales and positions to obtain local features of the signal, so as to achieve simultaneous analysis in the time domain and the frequency domain;

[0011] Assume that the function ψ(t)∈L 2 (R) satisfies the following three conditions (a), (b), and (c):

[0012]

[0013] In formula (a), ψ(t) is a function that alternates between positive and negative, and its function value has equal energy in the positive and negative parts; formula (b) indicates that ψ(t) is a wavelet, and its oscillation waveform rapidly decays to zero as the independent variable increases; formula (c) is to enable the inverse transformation of wavelet transform;

[0014] Based on the above expression, for any signal f(t)∈L 2 (R) defines the expression of continuous wavelet transform as shown in formula 2:

[0015]

[0016] Where a>0, b is the position parameter, is the complex conjugate of ψ(t), and CWT stands for continuous wavelet transform.

[0017] Further, in S20, the parts of the time-voltage curve graph containing the damaged and undamaged parts of the wire rope are respectively selected, and the center frequency α of the damaged part and the center frequency β of the undamaged part are respectively obtained by Fourier transform. The relative frequency factor k is calculated according to formula (3), and k is used to measure the relationship between the high-frequency characteristic frequency and the low-frequency characteristic frequency. The wavelet scale is measured by k;

[0018] In continuous wavelet transform, it is necessary to select a suitable wavelet mother function ψ(t) according to different signals, which is usually an oscillation function with zero mean and fast decay. Combined with the data collected from the leakage magnetic signal of wire rope defects, the characteristics of the Cmor wavelet function are used to extract the features of wire rope defects.

[0019]

[0020] The expression of Cmor wavelet mother function is shown in formula (4):

[0021]

[0022] Among them, ω0 is the frequency parameter used to control the frequency of the wavelet function, i is the imaginary unit, π is the circumference of a circle, and the constant is a complex exponential, is a Gaussian function; the Cmor wavelet mother function can decompose the signal at different time or space scales, provide the characteristics of the signal at different scales, and capture the local details and global advantages of the signal when processing one-dimensional data.

[0023] Furthermore, in S20, the sliding average filtering denoising is to regard the sampling values ​​obtained in a sampling period as a queue with a fixed length, put the new data at the end of the queue and discard the data at the head of the queue, and then perform arithmetic averaging on the data in the queue to obtain denoised data.

[0024] Furthermore, after the sliding average filtering denoising and before the time spectrum conversion, the data of multiple channels are subjected to wavelet decomposition at the same transformation scale, and the decomposed wavelets are added to obtain the average value to obtain the denoised data of multi-channel fusion.

[0025] Furthermore, in S30, the denoised data after multi-channel fusion is evenly cut with a period of 1 s, and the cut data is converted into a time spectrum diagram.

[0026] Furthermore, the image classification model is a ResNet50 model, which includes 5 stages; the first stage consists of a convolutional layer of size 7×7 and a maximum pooling layer; the subsequent 4 stages use the first bottleneck layer and the second bottleneck layer.

[0027] Furthermore, the number of input and output channels of the second bottleneck layer is the same, and the second bottleneck layer directly adds the input to the input after performing a convolution operation and uses it as the output, so that the number of input channels matches the number of output channels; the number of input and output channels of the first bottleneck layer is different, and the input of the first bottleneck layer is divided into two paths, one path is sampled through a convolution operation, and the other path is used to make the input and output dimensions different through a convolution operation.

[0028] A wire rope damage width identification system based on residual network multi-channel fusion is used to implement the above method.

[0029] According to any of the above embodiments, the present invention has at least the following beneficial effects: the present invention converts the magnetic flux leakage of the wire rope into voltage by using a Hall sensor array to obtain a time-voltage curve; uses continuous wavelet transform to obtain each wavelet component, then uses a sliding average filter to denoise each wavelet component, and converts it into a time spectrum diagram, and then uses a residual network Resnet50 to perform image classification, thereby realizing wire rope defect width detection. Compared with the prior art, the present invention has the following advantages: (1) Continuous wavelet transform can provide frequency information at different scales, which helps to analyze the change trend and characteristics of the signal more finely, can provide the classification model with relevant information of the signal in the frequency domain, and can more comprehensively describe the time-frequency change of the signal. (2) One-dimensional data is converted into a time spectrum diagram, and the image classification model is used for classification, which can better utilize the advantages of deep learning model to extract features. The time spectrum diagram is more discriminative than the original features in the one-dimensional data. It can show the frequency change of the data in different time periods, which helps the model capture the dynamic characteristics of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the process of the present invention.

[0031] Figure 2(a) is the working state diagram of the Hall sensor measuring the wire rope.

[0032] Figure 2 (b) is a schematic diagram of the structure of wire rope width damage.

[0033] Figure 2 (c) is a schematic diagram of the structure of wire ropes with different width damage.

[0034] Figure 3 This is a schematic diagram of 8-channel wire rope defect data.

[0035] Figure 4 This is the data waveform before 8-channel denoising.

[0036] Figure 5 This is the waveform of the data after 8-channel denoising.

[0037] Figure 6 It is an architectural diagram of the classification model of the present invention.

[0038] Figure 7 It is a partial time spectrum diagram during the experimental process of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] A wire rope damage width identification method based on residual network multi-channel fusion, such as Figure 1As shown, it includes an offline training step and an online detection step. The offline training step includes: S10, using steel wire ropes with the same damage depth and different damage widths as training samples, and using the Hall sensor array to obtain the magnetic leakage data of the training samples; S20, using continuous wavelet transform to obtain each wavelet component of the magnetic leakage data, and then using sliding average filtering to denoise each wavelet component to obtain denoised data; S30, converting the denoised data into a time spectrum diagram, using the time spectrum diagram as input and the damage width of the steel wire rope as output, training the image classification model based on the residual network, and obtaining a trained image classification model based on the residual network; the online detection step includes: S40, using the Hall sensor array to obtain the magnetic leakage data of the steel wire rope to be detected, performing continuous wavelet transform and sliding average filtering denoising on the magnetic leakage data, and obtaining denoised data; S50, inputting the denoised data into the image classification model based on the residual network, and obtaining the corresponding damage width of the steel wire rope output by the image classification model based on the residual network.

[0041] S10, wire rope damage width detection based on Hall sensor array.

[0042] In the present invention, a plurality of steel wire ropes with the same damage depth and different damage widths are prepared in advance as training samples, and then the magnetic leakage of the steel wire ropes is detected by using a Hall sensor array. During the detection process, the magnetic leakage data is converted into a voltage signal, which is transmitted to a host computer using a data transmission module to obtain a time-voltage diagram corresponding to the magnetic leakage data.

[0043] In order to ensure that the sensors in the Hall sensor array are at a suitable working voltage, a reference circuit is adapted for the Hall sensor array according to the working principle of the voltage comparator circuit during the test of the present invention. The appropriate working voltage is selected for the Hall sensor through the onboard potentiometer. The working state diagram is shown in FIG. Figure 2 (a) As shown. When the wire rope has defects such as broken wires, the magnetic flux will suddenly change. Figure 2 (b) and Figure 2 In (c), defect samples with different damage widths are shown, which represent different wire rope defect conditions. Figure 2 In (c), from left to right, the steel wire ropes have a damage depth of 2 mm and damage widths of 2, 4, 6, 8, and 10 mm. Figure 3 The following shows some data that can typically indicate wire rope defects when the damage width is 2mm and the depth is 2mm. Figure 3It can be seen that when the detector detects a wire rope defect, an obvious fluctuation will appear on the time-voltage curve, but the time when the fluctuation occurs in the 8 channels is not at the same moment, that is, the 8 Hall sensors do not sense the wire rope defect at the same time. The time shift of the waveform indicates that the magnetic flux data collected by the multi-channel sensor cannot be analyzed at the same time. For this reason, the present invention uses continuous wavelet transform to perform multi-channel data fusion to reduce the impact.

[0044] S20. Magnetic flux leakage signal denoising and feature extraction.

[0045] The data set acquired by the detector is usually mixed with a certain amount of noise. The noise sources include changes in the angle and position of the detector during measurement, uneven material of the wire rope, or surrounding electromagnetic interference noise. In order to remove noise and obtain a reliable and effective data set, the present invention adopts continuous wavelet transformation. The traditional Fourier transform integrates the signal over the entire time axis, so its result lacks time local information. This makes it impossible to perform joint time-frequency analysis and determine the time location of the frequency component. Because the detector acquisition time is real-time, the defective part of the wire rope cannot be feature extracted through the traditional Fourier transform. Although the short-time Fourier transform has a certain time localization capability, its fixed time window makes it impossible to change the resolution of time and frequency.

[0046] The present invention uses continuous wavelet transform to extract the features of wire rope defect detection signals on the data set. Continuous wavelet transform transforms the signal by using wavelet functions of different scales and positions to obtain the local features of the signal, so as to achieve the goal of simultaneous analysis in the time domain and frequency domain. Assuming that the function ψ(t)∈L 2 (R) satisfies the three conditions (a), (b), and (c) in Expression 1.

[0047]

[0048] Formula a indicates that ψ(t) is an alternating positive and negative function, and its function value is equal in a certain energy in the positive and negative parts. Formula b indicates that ψ(t) is a wavelet, and its oscillation waveform rapidly decays to zero as the independent variable increases. Formula c is to enable the inverse transformation of wavelet transform. Based on the above expressions, for any signal f(t)∈L 2 (R) defines the expression of continuous wavelet transform as shown in formula (2).

[0049]

[0050] Where a>0, b is the position parameter, is the complex conjugate of ψ(t). The present invention explores the selection of wavelet scale length γ and the ability of different wavelet functions in extracting wire rope defects for the wire rope defect data set collected by the detector. The width of the wavelet function is determined by the scale parameter, which affects the frequency resolution. When the scale is smaller, the frequency is higher and the wavelet function decays faster. The larger the scale, the lower the frequency and the slower the wavelet function decays. The present invention selects the damaged and undamaged parts of the wire rope in the data set respectively, and uses Fourier transform to obtain the center frequency α of the damaged part and the center frequency β of the undamaged part respectively, and calculates the relative frequency factor k according to formula 3. k is used to measure the relationship between the high-frequency characteristic frequency and the low-frequency characteristic frequency, and the wavelet scale is measured by k. In addition, the present invention comprehensively considers the frequency resolution and the signal duration, and selects the wavelet scale length γ as 2. In addition, in the continuous wavelet transform, it is necessary to select a suitable wavelet mother function ψ(t) according to different signals, which is usually an oscillation function with zero mean and fast decay. Combined with the data collected from the magnetic leakage signal of the wire rope defect, the present invention uses the characteristics of the Continuous Morlet (Cmor) wavelet function to extract the features of the wire rope defect.

[0051]

[0052] The expression of Cmor wavelet mother function is shown in formula 4:

[0053]

[0054] Among them, ω0 is the frequency parameter used to control the frequency of the wavelet function, i is the imaginary unit, and π is the circumference of a circle. The purpose of is to normalize the amplitude of the wavelet function to ensure that the wavelet functions at different scales have similar energy levels. With Gaussian function Multiplying makes the Cmor wavelet function have the effect of frequency modulation and amplitude modulation. It means that the frequency of the sine wave is modulated so that the frequency of the sine wave changes with time. This shows that the Cmor wavelet function can adapt to the situation where the detector output signal changes with time, proving that it has good adaptability to non-static signals. Gaussian function It is used to form a Gaussian window, which can localize the signal on the time axis, so that the wavelet analysis can focus on a specific time period to analyze the signal characteristics output by the detector. The Cmor wavelet mother function can decompose the signal at different time or space scales, providing the characteristics of the signal at different scales, enabling the Cmor wavelet mother function to capture the local details and global advantages of the signal when processing the one-dimensional data obtained by the detector.

[0055] By using the sliding average filtering method for each wavelet component, the sampling values ​​obtained in a sampling cycle are regarded as a queue of fixed length, and the new data is put into the end of the queue while the data at the head of the queue is discarded, and then the data in the queue is arithmetic averaged to obtain a new filtering result. By obtaining each wavelet component through continuous wavelet transform and then processing the wavelet component, periodic interference can be suppressed, the signal-to-noise ratio of the signal can be improved, and the one-dimensional sequence obtained by the detector can be made smoother while better maintaining the original characteristics of the signal. Figure 3 The wavelet components after sliding average filtering are shown. It can be seen from the figure that for the one-dimensional linear data obtained by the detector of the present invention, the sliding average filtering can remove the errors caused by instrument jitter during measurement.

[0056] In the present invention, the MSE (mean-square error), SNR (Peak Signal-to-Noise Ratio), and NCC (Normalized Correlation Cofficient) evaluation indicators are used to measure the noise reduction effect. The expression of MSE is shown in (5), where A signal refers to "pure signal", A denoised It refers to the signal after filtering. The smaller the MSE, the better the noise reduction effect.

[0057] The expression of SNR is shown in (6), where A signal Refers to the useful signal, A noise Refers to pure noise. The larger the SNR, the less noise contained in the signal and the better the noise reduction effect. The expression of NCC is shown in (7), where A s Refers to the pure signal, A d It refers to the filtered signal, reflecting the overall similarity of the signal waveform before and after denoising. The larger the NCC, the higher the correlation between the denoised signal and the original signal.

[0058]

[0059] The denoising results are shown in Table 1. It can be seen from Table 1 that the denoising method of sliding average filtering of wavelet components after continuous wavelet transform is better than the existing mainstream filtering method, and its MSE and SNR indicators are higher than %, respectively. The waveform results before denoising are as follows Figure 4 As shown, the waveform after denoising is as follows Figure 5 As shown in the figure, the denoised waveform is relatively smooth and still retains relevant features such as inflection points compared to the original waveform.

[0060]

[0061] Table 1. Denoising results of mainstream methods

[0062] 2.3 Wire rope damage identification model based on Resnet50

[0063] In order to identify the damage width of the wire rope from a large amount of wire rope data, the present invention uses a detector composed of a Hall sensor array to collect data on the same wire rope, the same damage depth, and different damage widths, and then uses continuous wavelet transform to obtain wavelet components for the original one-dimensional data, and uses a sliding average filter method to reduce the noise in the original data for each wavelet component. And under the same transformation scale, the data of 8 channels are decomposed by wavelet, and the decomposed wavelets are added to obtain the average value to obtain the processed signal. In order to improve the accuracy of wire rope damage width classification, the data after filtering and denoising was evenly cut with a period of 1s, and the cut data was converted into a time-spectrogram. The ResNet50 model was used to classify the time-spectrogram. The wire rope damage width detection result was obtained by classifying the time-spectrogram with ResNet50. Compared with the traditional one-dimensional signal data classification, the performance is better. The time-spectrogram provides a two-dimensional representation of the signal, which can capture time and frequency information at the same time. Traditional one-dimensional signal analysis may ignore important contextual information. In addition, the architecture of ResNet50 has deep layers and residual connections, which can learn more complex patterns, while the traditional one-dimensional signal classification method is based on a fixed recognition method and does not have strong robustness.

[0064] The ResNet structure is as follows Figure 6 As shown, for any input time spectrum There are 5 stages. For stage 1, F stage1 , which consists of a convolutional layer of size 7×7 and a maximum pooling layer. Two types of BTNK (Bottleneck) are used in the subsequent four stages. The difference is that the number of input and output channels of BTNK2 is the same, and the number of input and output channels of BTNK1 is different. BTNK2 directly adds the input to the input after convolution operation and uses it as output, so that the number of input channels matches the number of output channels. BTNK1 divides the input into two paths, one is sampled through convolution operation, and the other is the effect of input and output dimension difference through convolution operation. Combined with the complexity of the time spectrum graph, the present invention uses 4 groups of residual network structures. The residual network ResNet50 can solve the problems of gradient disappearance and gradient explosion during training, allowing the network to learn more complex functions. The present invention uses 1s as the time period, extracts the data output by 8 Hall sensors, converts the data output after wavelet decomposition and denoising fusion into a time spectrum graph, and obtains different types of time spectrum graphs by separately and independently measuring the wire ropes under different damage widths. After processing, there are 2680 one-dimensional detector data. After artificially removing the error data, a total of 270 time spectrograms are generated. Some of the time spectrograms are as follows: Figure 7 As shown. It is necessary to ensure that the number of time spectra of each damage width type used for model training is consistent. All the converted time spectra are mixed, and 80% of them are randomly selected as training sets, and 20% are used as test sets. The optimizer used for training is Adam, and the loss function is the cross entropy loss function. A total of 30 rounds of training are performed. For more parameters and parameter configurations for training, see Table 2.

[0065]

[0066] Table 2 Other relevant training parameters

[0067] The accuracy, recall, and F1-score indicators are used to evaluate the difference between the wire rope defect recognition proposed in the present invention and the mainstream classification model. Among them, TP refers to the number of positive classes predicted as positive classes, and FN refers to the number of positive classes predicted as negative classes. FP is the number of negative classes predicted as positive classes, and TN is the number of negative classes predicted as negative classes. The positive class refers to samples with wire rope defects. The negative class refers to samples with undamaged wire ropes. The calculation formula of the accuracy is shown in (8). The higher the accuracy, the better the model prediction effect, and the greater the probability of identifying the damaged part of the wire rope from the sample.

[0068]

[0069] The expression of recall rate is shown in formula (9). The larger the recall rate, the more accurate the model prediction effect.

[0070]

[0071] The calculation expression of F1-score is shown in formula (10), which is the harmonic mean of precision and recall. F1-score takes both precision and recall into consideration and is used to describe the probability that the model can predict the damaged part of the wire rope and its prediction accuracy.

[0072]

[0073] Using the same data set and training parameters, the present invention tests the mainstream one-dimensional signal recognition and classification model in the wire rope defect detection task, and the results are shown in Table 3. Table 3 shows that the ResNet50 used in the model can better classify the time spectrum under the same training parameters and data set, and can better detect the width of the wire rope damage. In addition, ResNet50 has the best classification effect on the time spectrum with a damage width of 8mm, and the classification effect on the time spectrum with a damage width of 6mm needs to be improved, but the overall detection accuracy of ResNet50 for wire ropes with different damage widths meets the requirements.

[0074] Model Accuracy Recall F1-score GoogleNet 0.25 0.17 0.20 AlexNet 0.41 0.26 0.32 VGG16 0.87 0.82 0.8 ResNet50 0.97 0.86 0.91

[0075] Table 3 Classification results of mainstream classification models

[0076] 4. Conclusion

[0077] The present invention converts the magnetic flux leakage of the wire rope into voltage by using a Hall sensor array, obtains a time-voltage curve, uses continuous wavelet transform to obtain each wavelet component, then uses a sliding average filter to denoise each wavelet component, and converts it into a time spectrum graph, and then uses a residual network Resnet50 for image classification, thereby realizing wire rope defect width detection. Compared with the method of directly pre-processing one-dimensional data and then classifying it, the present invention has the following advantages:

[0078] (1) Continuous wavelet transform can provide frequency information at different scales, which helps to analyze the changing trend and characteristics of the signal more finely, provide the classification model with relevant information of the signal in the frequency domain, and more comprehensively describe the time-frequency changes of the signal.

[0079] (2) The one-dimensional data is converted into a time-spectrogram and classified using an image classification model. This can better utilize the advantages of deep learning models in extracting features. The time-spectrogram is more discriminative than the original features in the one-dimensional data. It can show the frequency changes of the data in different time periods, which helps the model capture the dynamic characteristics of the data.

[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A wire rope damage width identification method based on residual network multi-channel fusion, comprising an offline training step and an online detection step, characterized in that: The offline training steps include: S10, using steel wire ropes with the same damage depth and different damage widths as training samples, and obtaining magnetic leakage data of the training samples using a Hall sensor array; S20, using continuous wavelet transform to obtain each wavelet component of the magnetic leakage data, and then using sliding average filtering to denoise each wavelet component to obtain denoised data; S30, converting the denoised data into a time-spectrogram, taking the time-spectrogram as input and the damaged width of the wire rope as output, training an image classification model based on a residual network to obtain a trained image classification model based on a residual network; The online detection steps include: S40, using the Hall sensor array to obtain the magnetic leakage data of the steel wire rope to be detected, and performing continuous wavelet transform and sliding average filtering denoising on the magnetic leakage data to obtain denoised data; S50, inputting the denoised data into an image classification model based on a residual network to obtain a damage width of the corresponding steel wire rope output by the image classification model based on the residual network.

2. The wire rope damage width identification method based on residual network multi-channel fusion according to claim 1 is characterized in that: In S10, a multi-channel Hall sensor array is used to detect the magnetic leakage data of the wire rope, and the magnetic leakage data is converted into a voltage signal to obtain a curve graph of time and voltage.

3. The wire rope damage width identification method based on residual network multi-channel fusion according to claim 2 is characterized in that: In S20, the voltage signal in the time-voltage curve graph is transformed using wavelet functions of different scales and positions to obtain local features of the signal to achieve simultaneous analysis in the time domain and the frequency domain; Assume that the function ψ(t)∈L 2 (R) satisfies the following three conditions (a), (b), and (c): ∫ψ(t)dt=0 (a) ∫|ψ(t)|dt<+∞ (b) In formula (a), ψ(t) is a function that alternates between positive and negative, and its function value has equal energy in the positive and negative parts; formula (b) indicates that ψ(t) is a wavelet, and its oscillation waveform rapidly decays to zero as the independent variable increases; formula (c) is to enable the inverse transformation of wavelet transform; Based on the above expression, for any signal f(t)∈L 2 (R) defines the expression of continuous wavelet transform as shown in formula 2: Where a>0, b is the position parameter, t is the time, is the complex conjugate of ψ(t), and CWT stands for continuous wavelet transform.

4. The wire rope damage width identification method based on residual network multi-channel fusion according to claim 3 is characterized in that: In S20, the damaged and undamaged parts of the wire rope are respectively selected from the time-voltage curve diagram, and the center frequency α of the damaged part and the center frequency β of the undamaged part are respectively obtained by Fourier transform. The relative frequency factor k is calculated according to formula (3). k is used to measure the relationship between the high-frequency characteristic frequency and the low-frequency characteristic frequency, and the wavelet scale is measured by k. In continuous wavelet transform, it is necessary to select a suitable wavelet mother function ψ(t) according to different signals, which is usually an oscillation function with zero mean and fast decay. Combined with the data collected from the leakage magnetic signal of wire rope defects, the characteristics of the Cmor wavelet function are used to extract the features of wire rope defects. The expression of Cmor wavelet mother function is shown in formula (4): Among them, ω0 is the frequency parameter used to control the frequency of the wavelet function, i is the imaginary unit, π is the circumference of a circle, and the constant is a complex exponential, is a Gaussian function; The Cmor wavelet mother function can decompose the signal at different time or space scales, provide the characteristics of the signal at different scales, and capture the local details and global advantages of the signal when processing one-dimensional data.

5. The wire rope damage width identification method based on residual network multi-channel fusion according to claim 4 is characterized in that: In S20, the sliding average filtering denoising is to regard the sampling values ​​obtained in a sampling period as a queue with a fixed length, put the new data at the end of the queue and discard the data at the head of the queue, and then perform arithmetic averaging on the data in the queue to obtain denoised data.

6. The wire rope damage width identification method based on residual network multi-channel fusion according to claim 5 is characterized in that: After the sliding average filter denoising, before converting the time spectrum, the data of multiple channels are subjected to wavelet decomposition at the same transformation scale, and the decomposed wavelets are added to obtain the average value to obtain the denoised data of multi-channel fusion.

7. The wire rope damage width identification method based on residual network multi-channel fusion according to claim 6 is characterized in that: In S30, the denoised data after multi-channel fusion is evenly cut with a period of 1 s, and the cut data is converted into a time spectrum diagram.

8. The wire rope damage width identification method based on residual network multi-channel fusion according to claim 7 is characterized in that: The image classification model is a ResNet50 model, which includes 5 stages; the first stage consists of a convolutional layer of size 7×7 and a maximum pooling layer; the subsequent 4 stages use the first bottleneck layer and the second bottleneck layer.

9. The wire rope damage width identification method based on residual network multi-channel fusion according to claim 8 is characterized in that: The number of input and output channels of the second bottleneck layer is the same. The second bottleneck layer performs a convolution operation on the input and then directly adds the result to the input, so that the number of input channels matches the number of output channels. The number of input and output channels of the first bottleneck layer is different. The input of the first bottleneck layer is divided into two paths, one path is sampled through a convolution operation, and the other path is used to make the input and output dimensions different through a convolution operation.

10. A wire rope damage width identification system based on residual network multi-channel fusion, characterized in that: Used to implement the method according to any one of claims 1 to 9.

Images