Steel wire rope defect identification and quantitative analysis method based on deep transfer learning

Abstract

The invention relates to the technical field of fault diagnosis and intelligent manufacturing, in particular to a steel wire rope defect identification and quantitative analysis method based on deep transfer learning. According to the technical scheme, the method comprises the following steps that leakage magnetic field signals LF and magnetic flux signals LMA of a steel wire rope are synchronously collected through a double-channel magnetic probe, and multi-channel data high-speed synchronous collection is conducted on the signals; and carrying out double-signal feature extraction on the LF signal and the LMA signal, wherein the extracted features comprise defect image information # imgabs0 #, a defect enclosed area # imgabs1 #, a defect wave crest number # imgabs2 #, a defect width # imgabs3 # and a defect height # imgabs4 #. Through deep transfer learning and multi-modal signal fusion, the problems that qualitative analysis is inaccurate and quantitative means are deficient in traditional steel wire rope defect detection are solved, the method has the advantages of high precision, strong generalization, high efficiency and automation, innovative technical support is provided for safe operation and intelligent operation and maintenance of industrial equipment, and the method is worthy of popularization and application. And steel wire rope defect identification and quantification are facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis and intelligent manufacturing, and in particular to a method for identifying and quantitatively analyzing wire rope defects based on deep transfer learning. Background Art

[0002] In recent years, wire ropes have played an important role in mining, ports, construction and other fields. As a key material for equipment lifting, transportation, support, shock absorption and other links, it not only directly affects the equipment operation efficiency, but also has important significance for personnel safety and equipment integrity. However, due to its high strength and wear resistance, wire ropes are prone to wear, corrosion, fatigue failure and wire breakage during use, which seriously threaten equipment safety and personnel life and property. At present, there are still great challenges in the quantitative analysis of wire rope defects. Traditional methods can only perform qualitative identification and cannot accurately assess the actual degree of damage of defect points. In response to this technical bottleneck, the present invention proposes a quantitative intelligent analysis method for wire rope defects based on deep transfer learning, which aims to provide a scientific solution for wire rope defect detection and improve the safety and reliability of industrial equipment.

[0003] Therefore, this application proposes a wire rope defect recognition and quantitative analysis method based on deep transfer learning. Summary of the invention

[0004] The purpose of the present invention is to address the great challenges in the quantitative analysis of wire rope defects in the background technology. Traditional methods can only perform qualitative identification and cannot accurately assess the actual damage degree of the defect points. A wire rope defect identification and quantitative analysis method based on deep transfer learning is proposed.

[0005] The technical solution of the present invention is a method for identifying and quantitatively analyzing wire rope defects based on deep transfer learning, comprising the following steps: S1. Synchronously collect the leakage magnetic field signal LF and the magnetic flux signal LMA of the wire rope through a dual-channel magnetic probe, and perform multi-channel data high-speed synchronous acquisition on the signals; S2. Perform dual signal feature extraction on the LF signal and the LMA signal, wherein the extracted features include defect image information , Area enclosed by the defect , Number of defect peaks , Defect Width and defect height , generate a combined feature dataset ; S3, dividing the combined feature data set into a source domain data set and a target domain data set, wherein the source domain data set includes a defect type label and a defect degree label, and the target domain data set does not include a label; S4. Construct deep migration network A and deep migration network B for classification task and regression task training respectively: The deep migration network A migrates the classification loss function by: in, is the number of source domain samples, is the number of samples in the target domain, yes Any value, yes Any value, The source domain samples, The target domain samples, is the calculation of the gradient distance error within the sample, and H is the Hilbert space; Perform training and output defect type classification results: The deep migration network B migrates the regression loss function: in, is the number of source domain samples, is the number of samples in the target domain, yes Any value, yes Any value, The source domain samples, The target domain samples, is the in-sample gradient square error calculation; Detection box loss function: in, is the predicted area enclosed by the wire rope detection frame, The area enclosed by the actual wire rope defect is the loss amount obtained by calculating the cross area ratio between the predicted position and the actual position. ; Conduct joint training and output quantitative results of defect levels; S5. Deploy the trained deep migration network A and deep migration network B into the fusion model, extract features of the LF signal and LMA signal collected in real time, and then use the fusion model to synchronously perform defect type identification and defect degree quantitative analysis.

[0006] Optionally, the dual signal feature extraction in step S2 further includes: collecting LF signals and LMA signals through an induction coil and a Hall sensor respectively, and preprocessing the signals to eliminate noise interference.

[0007] Optionally, the transfer classification loss function and the transfer regression loss function reduce the feature distribution difference between the source domain and the target domain through domain adaptation technology.

[0008] Optionally, the detection box loss function is used to optimize the overlap accuracy between the predicted defect area and the actual defect area.

[0009] Optionally, in step S3, the source domain dataset and the target domain dataset are generated by feature fusion and label division, the classification task dataset includes defect type labels, and the regression task dataset includes defect degree labels.

[0010] Optionally, the deep migration network A and the deep migration network B adopt a convolutional neural network (CNN) architecture and share some feature extraction layer parameters based on a migration learning strategy.

[0011] Optionally, the fusion model in S5 synchronously outputs comprehensive evaluation results of defect type, defect location and defect severity through a multi-task collaborative reasoning mechanism.

[0012] Optionally, it also includes a real-time prediction stage: LF signals and LMA signals are synchronously acquired through a high-speed acquisition board, and input into a fusion model after feature extraction to generate a real-time identification and quantitative analysis report of wire rope surface defects.

[0013] Compared with the prior art, the present invention has at least one of the following beneficial technical effects: Based on the principle of deep transfer learning, multi-task model training is carried out using multi-channel magnetic signal features to enable the model to have multi-classification and regression capabilities. The present invention solves the problems of inaccurate qualitative analysis and lack of quantitative means in traditional wire rope defect detection through deep transfer learning and multi-modal signal fusion. It has the advantages of high precision, strong generalization, high efficiency and automation, and provides innovative technical support for the safe operation and intelligent operation and maintenance of industrial equipment, and facilitates the identification and quantification of wire rope defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart for model training of quantitative intelligent analysis method for wire rope defects based on deep transfer learning; Figure 2 This is the model reasoning flowchart of the quantitative intelligent analysis method for wire rope defects based on deep transfer learning. DETAILED DESCRIPTION

[0015] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0016] Example

[0017] Model training phase: Assume that the LF signal dataset is , the LMA signal data set is By extracting features from LF and LMA signals, defect image information of dual-channel signals is obtained. At the same time, calculate the area enclosed by the dual channel defect , count the number of defect peaks of dual-channel signals , statistics of dual channel signal defect width ,high Through feature fusion, the combined feature data set is obtained . According to whether the defect type and degree can be identified, it is divided into source domain data set and target domain data set. The one with defect label is a classification task, and the one with defect degree is a regression task. By merging the classification data set, a source domain classification data set with defect label and a target domain classification data set without defect label are obtained. The classification data set is input into the designed deep transfer network A for classification training to obtain a trained classification model. At the same time, by merging the regression data set, a source domain regression data set with defect degree and a target domain regression data set without defect degree are obtained. The regression data set is input into the designed deep transfer network B for regression training to obtain a trained regression model.

[0018] The specific process is as follows: Source domain dataset , tagged as , target domain dataset , where the deep migration network A defines the migration classification loss function: in, is the number of source domain samples, is the number of samples in the target domain, yes Any value, yes Any value, The source domain samples, The target domain samples, is the calculation of the gradient distance error within the sample, and H is the Hilbert space; Perform training and output defect type classification results: The deep transfer network B defines the transfer regression loss function: here is the number of source domain samples, is the number of samples in the target domain, yes Any value, yes Any value, The source domain samples, The target domain samples, is the in-sample gradient square error calculation; At the same time, in order to characterize the overlap accuracy of the predicted wire rope defect position and the actual wire rope defect position, an additional detection box loss function is added to the migration network B: in, is the predicted area enclosed by the wire rope detection frame, The area enclosed by the actual wire rope defect is the loss amount obtained by calculating the cross area ratio between the predicted position and the actual position. .

[0019] See also Figure 2 : New rope prediction stage: First, start the dual-channel magnetic probe, turn on the wire rope actuator, and the wire rope starts to run. Use a high-speed acquisition board to perform high-speed synchronous acquisition of multiple data channels, and use the induction coil and Hall sensor to collect LF and LMA signals at different positions when the wire rope is running. After feature extraction of real-time LF and LMA data, put them into the trained fusion model. The fusion model will input the data into the deep migration network A and the deep migration network B for model inference, and obtain the corresponding defect types and defect severity at different positions on the real-time wire rope.

[0020] In this embodiment, by fusing the dual-channel feature extraction of the leakage magnetic field signal (LF) and the magnetic flux signal (LMA), this method can fully capture the geometric characteristics (such as defect area, number of peaks, width, height) and magnetic properties of wire rope defects, making up for the limitations of single signal analysis. Combined with deep transfer learning technology, the source domain data (with labels) is used to optimize the model training of the target domain (without labels or with a small number of labels), which significantly improves the accuracy of defect type identification and quantitative analysis, especially when the target domain data is insufficient, it can still maintain high reliability.

[0021] Among them, the design transfer classification loss function and the transfer regression loss function , domain adaptation technology is used to narrow the feature distribution differences between the source domain and the target domain, so that the model can adapt to data changes under different working conditions. In addition, the network architecture of the shared feature extraction layer (such as CNN) further enhances the model's adaptability to new scenarios and ensures stable performance in complex industrial environments.

[0022] In addition, a multi-task learning framework is used to simultaneously train the classification model (defect type identification) and the regression model (defect degree quantification), and the collaborative reasoning of the two types of tasks is achieved through the fusion model. This design not only reduces the computational overhead of repeated feature extraction, but also improves the overall efficiency through parameter sharing, while outputting a comprehensive evaluation result of the defect type, location, and severity, meeting the needs of multi-dimensional information in industrial inspection. Introducing the detection box loss function By optimizing the overlap accuracy (IoU) between the predicted defect area and the real area, the defect position can be accurately located. Combined with the quantitative analysis of the defect degree label in the regression task, the damage degree (such as the number of broken wires and the wear depth) can be scientifically evaluated to provide quantifiable data support for maintenance decisions.

[0023] It is worth noting that the present invention can complete dynamic detection during the operation of the wire rope by synchronously acquiring LF and LMA signals through a high-speed acquisition board, and performing real-time reasoning based on a trained fusion model. This technology greatly shortens the detection cycle and reduces the reliance on manual inspections. It is suitable for continuous operation requirements in scenarios such as mines and ports, and significantly improves the safety and operation and maintenance efficiency of industrial equipment. Through high-precision defect identification and quantitative analysis, potential faults can be warned in advance to avoid sudden fracture accidents caused by defect accumulation. At the same time, accurate damage assessment can guide targeted maintenance (such as partial replacement rather than overall scrapping), extend the service life of the wire rope, and reduce enterprise maintenance costs.

[0024] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A wire rope defect recognition and quantitative analysis method based on deep transfer learning, characterized in that: The following steps are involved: S1. synchronously collect the leakage magnetic field signal LF and the magnetic flux signal LMA of the wire rope through a dual-channel magnetic probe, and perform multi-channel data high-speed synchronous acquisition on the signals; S2. Perform dual signal feature extraction on the LF signal and the LMA signal, wherein the extracted features include defect image information , Area enclosed by the defect , Number of defect peaks , Defect Width and defect height , generate a combined feature dataset ; S3, dividing the combined feature data set into a source domain data set and a target domain data set; S4, build deep transfer network A and deep transfer network B, which are used for classification task and regression task training respectively; S5. Deploy the trained deep migration network A and deep migration network B into the fusion model, extract features of the LF signal and LMA signal collected in real time, and then use the fusion model to synchronously perform defect type identification and defect degree quantitative analysis.

2. According to the method for steel wire rope defect recognition and quantitative analysis based on deep transfer learning in claim 1, it is characterized in that: The dual signal feature extraction in step S2 further includes: collecting LF signals and LMA signals through an induction coil and a Hall sensor respectively, and preprocessing the signals to eliminate noise interference.

3. According to the method for steel wire rope defect identification and quantitative analysis based on deep transfer learning in claim 1, it is characterized in that: The migration classification loss function and the migration regression loss function reduce the feature distribution difference between the source domain and the target domain through domain adaptation technology.

4. The method for identifying and quantitatively analyzing wire rope defects based on deep transfer learning according to claim 1 is characterized in that: The detection box loss function is used to optimize the overlap accuracy between the predicted defect area and the real defect area.

5. The method for identifying and quantitatively analyzing wire rope defects based on deep transfer learning according to claim 1 is characterized in that: In step S3, the source domain dataset includes defect type labels and defect degree labels, the target domain dataset does not include labels, the source domain dataset and the target domain dataset are generated by feature fusion and label division, the classification task dataset includes defect type labels, and the regression task dataset includes defect degree labels.

6. The method for identifying and quantitatively analyzing wire rope defects based on deep transfer learning according to claim 1 is characterized in that: The deep migration network A migrates the classification loss function by: in, is the number of source domain samples, is the number of samples in the target domain, yes Any value, yes Any value, The source domain samples, The target domain samples, is the calculation of the gradient distance error within the sample, and H is the Hilbert space; Perform training and output defect type classification results: The deep migration network B migrates the regression loss function: in, is the number of source domain samples, is the number of samples in the target domain, yes Any value, yes Any value, The source domain samples, The target domain samples, is the in-sample gradient square error calculation; Detection box loss function: in, is the predicted area enclosed by the wire rope detection frame, is the area enclosed by the actual wire rope defect, and the loss is obtained by calculating the cross area ratio between the predicted position and the actual position. ; Conduct joint training and output quantitative results of defect levels; The deep migration network A and the deep migration network B adopt a convolutional neural network architecture and share some feature extraction layer parameters based on a migration learning strategy.

7. The method for identifying and quantitatively analyzing wire rope defects based on deep transfer learning according to claim 1 is characterized in that: The fusion model in S5 synchronously outputs comprehensive evaluation results of defect type, defect location and defect severity through a multi-task collaborative reasoning mechanism.

8. The method for identifying and quantitatively analyzing wire rope defects based on deep transfer learning according to claim 1 is characterized in that: It also includes a real-time prediction stage: the LF signal and LMA signal are synchronously acquired through a high-speed acquisition board, and after feature extraction, they are input into the fusion model to generate a real-time identification and quantitative analysis report of wire rope surface defects.

Images