Detection and classification method of irregular thread defects

Abstract

The present invention discloses a detection and classification method for irregular thread defects, comprising the following steps: S1 data acquisition: using three eddy current probes for the thread root, thread foot, and thread side positions respectively, the trapezoidal thread is scanned in a spiral motion along the thread groove to acquire the original data of phase information; S2 data preprocessing: using multi-scale principal component analysis to denoise the collected original data; S3 feature extraction: performing feature extraction and forming a combined feature quantity; S4 using a regression model to predict the defect position, length, and depth under different paths; S5 defect hazard level classification. The present invention is applicable to the field of thread detection technology. Scanning is performed at different positions where defects may appear (thread root, thread side, thread foot), which can effectively provide information such as the position, shape, area, depth, etc. of irregular thread defects, and perform hazard level classification based on this information, which has higher guiding value for subsequent thread processing work.

Description

Technical Field

[0001] The present invention belongs to the technical field of thread detection, and in particular to a method for detecting and classifying irregular thread defects. Background Art

[0002] In production and life, threaded structures are everywhere. During use, they are inevitably affected by the environment and external forces, which will cause crack defects and affect the normal use of the threaded structure.

[0003] Some scholars have conducted research on methods for detecting and classifying thread defects, primarily focusing on machine vision solutions. However, machine vision solutions have high requirements for lighting conditions and thread surface cleanliness, and can only detect surface defects. Threaded structures with surface stains such as oil and rust (such as drill rod bolts) require a more tedious and time-consuming cleaning process to prepare them for visual inspection. In addition, some near-surface defects may exist in the thread structure, which cannot be detected using visual solutions. This patent uses eddy current detection to inspect thread defects.

[0004] Existing eddy current detection solutions for thread defects mostly aim to determine the presence of defects at the thread root. In reality, thread defects can occur not only at the root but also at the foot and flank. Furthermore, defects can have irregular shapes. Defects of varying locations, shapes, and depths can have varying degrees of impact on bolt performance and require different subsequent treatment methods. Existing technologies primarily focus on determining the presence of defects at the thread root. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and provide a method for detecting and classifying irregular thread defects.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] The detection and classification method of irregular thread defects includes the following steps:

[0008] S1 data acquisition: Using three eddy current probes at the thread root, thread foot, and thread flank positions, the trapezoidal thread is scanned helically along the thread groove to collect raw data of phase information;

[0009] S2 Data preprocessing: Use multi-scale principal component analysis to denoise the collected raw data;

[0010] S3 feature extraction: extract features and form combined feature quantities;

[0011] S4 uses regression models to predict defect location, length, and depth under different paths;

[0012] S5 defect hazard level classification.

[0013] Preferably, in the data collection step S1, the trapezoidal thread is scanned by performing a helical motion along the thread groove, with a total of 5 paths, and the original data X=(x1, x2, x3, x4, x5) of the phase information is collected.

[0014] Preferably, in the data preprocessing step S2, the preprocessed signal is Q=(q1, q2, q3, q4, q5).

[0015] Preferably, the feature extraction step S3 includes:

[0016] Extract the features of the tooth bottom signal q1 and form a combined feature:

[0017] (1) Perform wavelet packet decomposition on each preprocessed signal to obtain wavelet p 11 ~p 1n ;

[0018] (2) Perform empirical mode decomposition on each preprocessed signal and select the first m intrinsic modal components with the largest energy, named p 1(n+1) ~p 1(n+m) , where m is less than or equal to the minimum number of intrinsic modal components of a single signal after empirical mode decomposition;

[0019] (3) Extract each signal p obtained by decomposition in steps (1) and (2) 11 ~p 1(n+m) The statistical components of: standard deviation, signal energy, skewness coefficient, kurtosis, form a combined feature f1, f1 is a one-dimensional vector containing (n+m)*4 elements;

[0020] In the above manner, features of other signals are extracted respectively and combined feature quantities are formed. The original data X = (x1, x2, x3, x4, x5), and the obtained combined feature quantity is F = (f1, f2, f3, f4, f5).

[0021] Preferably, in step S4, the regression model is an extreme gradient boosting tree model.

[0022] Preferably, in step S4, using a regression model to predict the defect position, length, and depth under different paths includes:

[0023] The combined feature quantity F is input into the trained extreme gradient boosting tree model to obtain the prediction results of the starting position, length and average depth of the defect in each path, where the starting position of the defect is (s1, s2, s3, s4, s5), the length is (l1, l2, l3, l4, l5), and the average depth is (d1, d2, d3, d4, d5).

[0024] Preferably, in step S5, the defect risk level classification includes:

[0025] Based on the predicted starting position, length and depth of the defect, the position, area, shape and depth of the defect are further determined. The five paths are expanded into a two-dimensional plane. The scanning path is marked with a dotted line, and the starting and ending positions and lengths of the defect predicted by the regression model under different paths are represented by solid lines. The ends of the solid lines are connected to obtain the defect shape, and the defect area can be calculated. According to the defect position, area, shape, depth and other information, the defect hazard level is graded and a corresponding treatment plan is formulated.

[0026] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0027] In the present invention, scanning is performed on different locations where defects may occur (bottom of the tooth, side of the tooth, and foot of the tooth), which can effectively provide information such as the location, shape, area, and depth of irregular thread defects, and classify the hazard level accordingly, which has a higher guiding value for subsequent thread processing.

[0028] In the present invention, multi-scale principal component analysis can effectively filter out the noise of the original signal; the combined feature quantity formed by the feature extraction method based on wavelet packet decomposition, empirical mode decomposition and high-order statistical components contains rich time-frequency information, and combined with the extreme gradient boosting tree model, the size information of the defect can be predicted more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flow chart of the detection and classification method of irregular thread defects of the present invention;

[0030] Figure 2 Schematic diagram of the scanning path in the detection and classification method of irregular thread defects of the present invention;

[0031] Figure 3 Schematic diagram of the method for evaluating the area and shape in the detection and classification method of irregular thread defects of the present invention;

[0032] Figure 4 It is a schematic diagram of the corresponding relationship between the eddy current probe coil structure and its detection position in the detection and classification method of irregular thread defects of the present invention. DETAILED DESCRIPTION

[0033] The following is combined with Figure 1-4 , further illustrating a specific embodiment of the method for detecting and classifying irregular thread defects of the present invention. The method for detecting and classifying irregular thread defects of the present invention is not limited to the description of the following embodiments.

[0034] Example 1:

[0035] This embodiment provides a specific implementation method of the detection and classification method of irregular thread defects, such as Figure 1-3 As shown, the following steps are included:

[0036] S1 data acquisition: Using three eddy current probes at the thread root, thread foot, and thread flank positions, the trapezoidal thread is scanned helically along the thread groove to collect raw data of phase information;

[0037] S2 Data preprocessing: Use multi-scale principal component analysis to denoise the collected raw data;

[0038] S3 feature extraction: extract features and form combined feature quantities;

[0039] S4 uses regression models to predict defect location, length, and depth under different paths;

[0040] S5 defect hazard level classification.

[0041] Furthermore, in step S1 data collection, the trapezoidal thread is scanned by performing a helical motion along the thread groove, with a total of 5 paths, and raw data X=(x1, x2, x3, x4, x5) of the phase information is collected.

[0042] Furthermore, in step S2 data preprocessing, the preprocessed signal is Q=(q1, q2, q3, q4, q5).

[0043] Furthermore, step S3 feature extraction includes:

[0044] Extract the features of the tooth bottom signal q1 and form a combined feature:

[0045] (1) Perform wavelet packet decomposition on each preprocessed signal to obtain wavelet p 11 ~p 1n ;

[0046] (2) Perform empirical mode decomposition on each preprocessed signal and select the first m intrinsic modal components with the largest energy, named p 1(n+1) ~p 1(n+m) , where m is less than or equal to the minimum number of intrinsic modal components of a single signal after empirical mode decomposition;

[0047] (3) Extract each signal p obtained by decomposition in steps (1) and (2) 11 ~p 1(n+m) The statistical components of: standard deviation, signal energy, skewness coefficient, kurtosis, form a combined feature f1, f1 is a one-dimensional vector containing (n+m)*4 elements;

[0048] In the above manner, features of other signals are extracted respectively and combined feature quantities are formed. The original data X = (x1, x2, x3, x4, x5), and the obtained combined feature quantity is F = (f1, f2, f3, f4, f5).

[0049] Furthermore, in step S4, the regression model is an extreme gradient boosting tree model.

[0050] Furthermore, in step S4, the defect location, length, and depth under different paths are predicted using a regression model, including:

[0051] The combined feature quantity F is input into the trained extreme gradient boosting tree model to obtain the prediction results of the starting position, length and average depth of the defect in each path, where the starting position of the defect is (s1, s2, s3, s4, s5), the length is (l1, l2, l3, l4, l5), and the average depth is (d1, d2, d3, d4, d5).

[0052] Furthermore, in step S5, defect risk level classification includes:

[0053] Based on the predicted starting position, length and depth of the defect, the position, area, shape and depth of the defect are further determined. The five paths are expanded into a two-dimensional plane. The scanning path is marked with a dotted line, and the starting and ending positions and lengths of the defect predicted by the regression model under different paths are represented by solid lines. The ends of the solid lines are connected to obtain the defect shape, and the defect area can be calculated. According to the defect position, area, shape, depth and other information, the defect hazard level is graded and a corresponding treatment plan is formulated.

[0054] Example 2:

[0055] This embodiment provides a specific implementation method of the detection and classification method of irregular thread defects, such as Figure 4 As shown, the following steps are included:

[0056] (21) Data acquisition: Three differential eddy current probes are used to scan the thread bottom, tooth side and tooth foot positions along the thread spiral line. The corresponding relationship between the eddy current probe coil structure and its detection position is as follows: Figure 4 As shown:

[0057] The instrument used was the MFLI digital lock-in amplifier from Zurich Instruments, Switzerland. This lock-in amplifier can collect phase information X during the probe scanning process. In this example, a total of 5,000 sets of experimental data were obtained, representing different defects of the same thread type. The measured length and depth data for each defect scan path were marked.

[0058] (22) Data preprocessing: Multiscale principal component analysis (MSPCA) is used to denoise the collected original signal X to obtain the preprocessed signal Q.

[0059] (23) Feature extraction: Feature extraction and forming combined feature quantities, including the following sub-steps:

[0060] (23.1) The pre-processed signal under each path is decomposed by wavelet packet to obtain wavelet p 11 ~p 1n

[0061] (23.2) Perform EMD (empirical mode decomposition) on each preprocessed signal and select the first m IMF (intrinsic mode) components with the largest energy, named p 1(n+1) ~p 1(n+m) , where m is less than or equal to the minimum number of IMFs after a single signal is decomposed by EMD.

[0062] (23.3) Extract each signal p obtained by decomposition in steps (23.1) and (23.2) 11 ~p 1(n+m) The statistical components of: standard deviation, signal energy, skewness coefficient, kurtosis, form a combined feature f1, f1 is a one-dimensional vector containing (n+m)*4 elements.

[0063] (24) Model training: An XGBoost regression model is established for the defect depth, length, and position of each scanning path. The model is evaluated using the ten-fold cross-validation method to find the XGBoost model with the best generalization performance for each path.

[0064] (25) Defect location, length, and depth prediction: The measured defect data that needs to be evaluated for hazard level is preprocessed and feature extracted as described above to obtain feature F. Feature F is input into the trained extreme gradient boosting tree model (XGBoost) to obtain the prediction results of the starting position, length, and average depth of the defect in each path.

[0065] (26) Defect hazard level classification: Based on the defect starting position and length prediction under the five paths, the location, area, and shape of the defect can be further determined. Generally speaking, different defect locations have different hazard levels and different treatment methods; the larger the defect area, the higher the defect hazard level; the larger the average defect depth, the higher the defect hazard level. Irregular defects are classified according to their hazard level based on their shape, depth, location, and area.

[0066] Working principle, such as Figure 1-4 As shown:

[0067] First, for the thread structure to be inspected, eddy current detection data of thread defects of different positions and shapes are collected to construct a data set. The length of the defect in each set of data at the tooth bottom, tooth foot, and tooth side along the spiral direction is marked.

[0068] Secondly, the raw data in the eddy current detection data set are preprocessed and feature extracted to obtain the data set combined feature quantity F.

[0069] Then, an extreme gradient boosting tree (XGBoost) regression model was established. The function of the model was to predict the length and average depth of defects along the spiral direction at the tooth bottom, tooth foot, and tooth side.

[0070] Then, the extreme gradient boosting tree (XGBoost) model is used to give the location of defects under different paths in the measured data as well as the defect length L and average depth D.

[0071] Finally, the defect area and shape are further calculated, and the danger level of the defects under the measured data is classified accordingly.

[0072] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for detecting and classifying irregular thread defects, characterized in that: The following steps are involved: S1 data acquisition: Using three eddy current probes at the thread root, thread foot, and thread flank positions, the trapezoidal thread is scanned helically along the thread groove to collect raw data of phase information; S2 Data preprocessing: Use multi-scale principal component analysis to denoise the collected raw data; S3 feature extraction: extract features and form combined feature quantities; S4 uses regression models to predict defect location, length, and depth under different paths; S5 defect hazard level classification; In step S1, in data acquisition, the trapezoidal thread is scanned in a helical motion along the thread groove, with a total of 5 paths, and the raw data of the phase information X = (x1, x2, x3, x4, x5) is acquired; In step S2 data preprocessing, the preprocessed signal is Q = (q1, q2, q3, q4, q5); Step S3: feature extraction, including: Extract the features of the tooth bottom signal q1 and form a combined feature: (1) Perform wavelet packet decomposition on each preprocessed signal to obtain wavelet p 11 ~p 1n ; (2) Perform empirical mode decomposition on each preprocessed signal and select the first m intrinsic modal components with the largest energy, named p 1(n+1) ~p 1(n+m) , where m is less than or equal to the minimum number of intrinsic modal components of a single signal after empirical mode decomposition; (3) Extract each signal p obtained by decomposition in steps (1) and (2) 11 ~p 1(n+m) The statistical components of: standard deviation, signal energy, skewness coefficient, kurtosis, form a combined feature f1, f1 is a one-dimensional vector containing (n+m)*4 elements; By using the above method, feature extraction is performed on other signals respectively and combined feature quantities are formed. The original data X = (x1, x2, x3, x4, x5), and the obtained combined feature quantity is F = (f1, f2, f3, f4, f5); In step S4, the regression model is an extreme gradient boosting tree model; In step S4, the regression model is used to predict the defect location, length, and depth under different paths, including: The combined feature F is input into the trained extreme gradient boosting tree model to obtain the prediction results of the starting position, length and average depth of the defect in each path. The starting position of the defect is (s1, s2, s3, s4, s5), the length is (l1, l2, l3, l4, l5), and the average depth is (d1, d2, d3, d4, d5); In step S5, the defect risk level classification includes: Based on the predicted starting position, length and depth of the defect, the position, area, shape and depth of the defect are further determined. The five paths are expanded into a two-dimensional plane. The scanning path is marked with a dotted line, and the starting and ending positions and lengths of the defect predicted by the regression model under different paths are represented by solid lines. The ends of the solid lines are connected to obtain the defect shape, and the defect area is calculated. According to the position, area, shape and depth information of the defect, the hazard level of the defect is graded, and a corresponding treatment plan is formulated.

Images