A multi-layer neural network optimization method for infrared thermal imaging defect recognition
Through multi-layer neural network optimization of infrared thermal imaging data, the problem of low signal-to-noise ratio and low depth in metal welding is solved, real-time defect detection of friction stir welding areas is realized, and the detection depth and accuracy are improved. Meets the need for welding parameter optimization.
Patent Information
- Application Number
- CN202210798796.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-08
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-07-08
AI Technical Summary
Infrared thermal imaging technology has problems with low signal-to-noise ratio and low depth in metal welding, making it difficult to effectively detect surface or internal defects in aluminum alloy welding, especially in high thermal conductivity materials, which limits its application. Depth and accuracy.
Using multi-layer neural network optimization method, pre-processing of infrared thermal imaging data, flaw feature extraction and image data preparation, establishing an infrared imaging multi-layer neural network model, and improving welding through the optimization method of combining cross entropy loss function and L2 regularization SNR and flaw detection depth in the area.
Real-time defect detection in the friction stir welding area is realized, the depth and accuracy of welding defect detection is improved, the signal-to-noise ratio of infrared imaging is improved, the need for optimization of friction stir welding parameters is met, and the detection efficiency is improved.
Smart Images

Figure CN115147379B_ABST
Abstract
Description
[Technical field]
[0001] The invention relates to a metal welding defect detection method, in particular to a multi-layer neural network optimization method for infrared thermal imaging defect recognition, and belongs to the technical field of infrared detection. [Background technology]
[0002] Aluminum alloys have excellent specific strength and specific stiffness, good casting properties and plastic processing properties, and are widely used in aerospace, transportation, construction, electromechanical and other fields. With the continuous expansion of aluminum alloy applications and the increasing demand for reducing manufacturing costs, people are looking for a feasible and cost-effective aluminum alloy welding technology. Among them, stir friction welding uses friction heat to melt and extrude the material to form a dense solid phase weld, which is particularly suitable for the welding of aluminum alloys. However, mechanized and automated stir friction welding is prone to introduce surface or internal defects under improper operating conditions. Therefore, in order to improve the quality of stir friction welding and optimize the operating parameters, it is urgent to provide real-time high-speed non-destructive detection technology for stir friction welding defects.
[0003] Commonly used non-destructive testing technologies include eddy current method, ultrasonic method, X-ray and infrared thermal imaging. The eddy current method is suitable for micron-level surface defects, the ultrasonic method is limited to small-range contact detection, and X-ray technology is not suitable for on-site real-time detection due to cost and safety. Infrared thermal imaging has outstanding advantages such as fast response, no disturbance, non-contact, and all-weather target detection. It is suitable for surface or internal defect detection in stir friction welding areas. However, the contrast of thermal signals in high thermal conductivity aluminum alloy materials is low, which brings great challenges to defect positioning and internal defect detection of thick plate materials. The application of infrared thermal imaging technology in the metal field is limited by low signal-to-noise ratio and low depth.
[0004] Therefore, in order to solve the above technical problems, it is necessary to provide an innovative multi-layer neural network optimization method for infrared thermal imaging defect recognition to overcome the above defects in the prior art. [Summary of the invention]
[0005] The purpose of the present invention is to provide a multi-layer neural network optimization method for infrared thermal imaging defect recognition, which can perform real-time defect detection on the stir friction welding effect and improve the signal-to-noise ratio of infrared thermal imaging defects / non-defects.
[0006] To achieve the above object, the technical solution adopted by the present invention is: a multi-layer neural network optimization method for infrared thermal imaging defect recognition, which includes the following process steps:
[0007] 1), thermal imaging data pre-processing;
[0008] 2), Friction stir welding defect feature extraction;
[0009] 3), welding defect / non-defect image data preparation;
[0010] 4), Multi-layer neural network modeling for infrared imaging;
[0011] 5), Multi-layer neural network model optimization.
[0012] The multi-layer neural network optimization method for infrared thermal imaging defect recognition of the present invention is further as follows: the specific steps of step 1) are:
[0013] 1-1), convert the original RGB image data into a 64-bit grayscale image, identify and crop the target area, remove the invalid background, and use Gaussian filtering to increase the contrast between defective pixels and non-defective pixels. The sample kernel is 13×13 and the standard deviation is 3.0; the Gaussian filtered image sequence is G(j,i,t), where G is the image at time t, j and i are the column number and row number respectively; the partial derivative of the image data after Gaussian filtering with respect to time is
[0014] 1-2), convert the three-dimensional image data into a two-dimensional matrix N×P; where the number of rows N is the number of infrared image pixels, including spatial changes; the number of columns P records the number of infrared images, including temporal changes.
[0015] The multi-layer neural network optimization method for infrared thermal imaging defect recognition of the present invention is further as follows: the specific steps of step 2) are:
[0016] 2-1), linear modeling of thermal imaging signals, compressing the instantaneous thermal imaging image sequence into four images while retaining spatial and temporal information;
[0017] 2-2), a univariate polynomial is used to perform Gaussian filtering on each pixel. The grayscale value after filtering has the following function with respect to time: G(j, i, t)≈a0(j, i)+a1(j, i)t, where a0(j, i) and a1(j, i) are polynomial parameters; the grayscale value slope has the following function with respect to time: Where b0(j, i) and b1(j, i) are polynomial parameters: Thus, the three-dimensional grayscale data after Gaussian filtering is converted into two polynomial coefficient graphs formed by polynomial coefficients, namely a0(j, i) and a1(j, i);
[0018] 2-3), the least square method is used to reweight the polynomial parameters (a0, a1, b0, b1). When the accuracy reaches 10 -3 Stop the reweighting calculation when , and obtain the corresponding weight matrix and bias matrix, whose image grayscale range is [0,255].
[0019] The multi-layer neural network optimization method for infrared thermal imaging defect recognition of the present invention further comprises the following steps:
[0020] 3-1), the data preparation process includes data labeling, conversion, sampling and segmentation: First, the data is classified and labeled using the location and size of the defects; then, the input data features are arranged in columns, and the z-score normalization method is used to reduce the Euclidean distance between features so that the mean and unit standard deviation of each feature are zero;
[0021] 3-2) Create a data set with non-defective and defective data samples in a ratio of 1:1;
[0022] 3-3), randomly divide the data set into two equal parts to ensure sufficient training and test sets.
[0023] The multi-layer neural network optimization method for infrared thermal imaging defect recognition of the present invention further comprises the following steps:
[0024] 4-1), a binary classification method [0,1] is used for supervised learning, where 0 represents no defect and 1 represents defect;
[0025] 4-2), the input layer consists of 10 neurons, and the hidden layer includes a nonlinear activation function The output layer consists of one neuron and a logistic activation function composition.
[0026] The multi-layer neural network optimization method for infrared thermal imaging defect recognition of the present invention further comprises: the specific steps of step 5) are:
[0027] 5-1), randomly select the initial weights of the neural network and use the cross entropy loss function as follows: Where n is the data point, y i and z i are the predicted value and target value of the i-th input vector respectively;
[0028] 5-2), use L2 regularization to perform weight penalty, as follows: Where ω is the model weight vector;
[0029] 5-3), the complete damage function after combining the cross entropy damage function and the L2 regularization method is as follows: total =J cc +ρJ reg , where ρ is the regularization parameter and ρ ≥ 0;
[0030] 5-4), the scaled conjugate gradient back propagation method is used to minimize the damage function and perform weight optimization, which has two stopping criteria, namely the accuracy of the objective function and the weight at the solution;
[0031] 5-5), K-fold cross validation is used to improve the selection of the global minimum.
[0032] The multi-layer neural network optimization method for infrared thermal imaging defect recognition of the present invention can also be: in step 5-3), ρ = 0.225; in step 5-4), take 10 -3 is the stopping criterion; in step 5-4), K=10.
[0033] Compared with the prior art, the present invention has the following beneficial effects:
[0034] 1. The multi-layer neural network optimization method for infrared thermal imaging defect recognition of the present invention detects and identifies the thermal imaging signal of the stir friction welding area in real time by establishing a multi-layer neural network model, thereby improving the signal-to-noise ratio and flaw detection depth of infrared imaging defects / non-defects in the welding area, and improving the depth and accuracy of welding defect detection;
[0035] 2. The present invention optimizes weights through continuous iteration of the neural network model, and can efficiently and accurately perform real-time detection and feedback of welding defects, thereby improving detection efficiency, thereby laying a foundation for optimizing stir friction welding parameters and better meeting practical application needs.
Brief Description of the Drawings
[0036] Figure 1 It is a flow chart of the multi-layer neural network optimization method for infrared thermal imaging defect recognition of the present invention. [Specific implementation method]
[0037] Please refer to the instruction manual Figure 1 As shown, the present invention relates to a multi-layer neural network optimization method for infrared thermal imaging defect recognition, which includes the following process steps:
[0038] 1) Pre-processing of thermal imaging data; the specific steps are:
[0039] 1-1), convert the original RGB image data into a 64-bit grayscale image, identify and crop the target area, remove the invalid background, and use Gaussian filtering to increase the contrast between defective pixels and non-defective pixels. The sample kernel is 13×13 and the standard deviation is 3.0; the Gaussian filtered image sequence is G(j,i,t), where G is the image at time t, j and i are the column number and row number respectively; the partial derivative of the image data after Gaussian filtering with respect to time is
[0040] 1-2), convert the three-dimensional image data into a two-dimensional matrix N×P; where the number of rows N is the number of infrared image pixels, including spatial changes; the number of columns P records the number of infrared images, including temporal changes.
[0041] 2) Extraction of friction stir welding defect features; the specific steps are:
[0042] 2-1), linear modeling of thermal imaging signals, compressing the instantaneous thermal imaging image sequence into four images while retaining spatial and temporal information.
[0043] 2-2), a univariate polynomial is used to perform Gaussian filtering on each pixel. The grayscale value after filtering has the following function with respect to time: G(j, i, t)≈a0(j, i)+a1(j, i)t, where a0(j, i) and a1(j, i) are polynomial parameters; the grayscale value slope has the following function with respect to time: Where b0(j, i) and b1(j, i) are polynomial parameters: Thus, the three-dimensional grayscale data after Gaussian filtering is converted into two polynomial coefficient graphs formed by polynomial coefficients, namely a0(j, i) and a1(j, i).
[0044] 2-3), the least square method is used to reweight the polynomial parameters (a0, a1, b0, b1). When the accuracy reaches 10 -3 Stop the reweighting calculation when , and obtain the corresponding weight matrix and bias matrix, whose image grayscale range is [0,255].
[0045] 3) Preparation of welding defect / non-defect image data; the specific steps are as follows:
[0046] 3-1), the data preparation process includes data labeling, conversion, sampling and segmentation: first, the data is classified and labeled using the location and size of the defects; then, the input data features are arranged in columns, and the z-score normalization method is used to reduce the Euclidean distance between features so that the mean and unit standard deviation of each feature are zero.
[0047] 3-2), create a data set with non-defective and defective data samples in a ratio of 1:1.
[0048] 3-3), randomly divide the data set into two equal parts to ensure sufficient training and test sets.
[0049] 4) Infrared imaging multi-layer neural network modeling; the specific steps are:
[0050] 4-1), a binary classification method [0,1] is used for supervised learning, that is, 0 represents no defect and 1 represents defect.
[0051] 4-2), the input layer consists of 10 neurons, and the hidden layer includes a nonlinear activation function The output layer consists of one neuron and a logistic activation function composition.
[0052] 5), multi-layer neural network model optimization; the specific steps are:
[0053] 5-1), randomly select the initial weights of the neural network and use the cross entropy loss function as follows: Where n is the data point, y i and z i are the predicted value and target value of the i-th input vector respectively.
[0054] 5-2), use L2 regularization to perform weight penalty, as follows: where ω is the model weight vector.
[0055] 5-3), the complete damage function after combining the cross entropy damage function and the L2 regularization method is as follows: total =J cc +ρJ reg , where ρ is a regularization parameter, and ρ≥0. In this implementation, ρ=0.225.
[0056] 5-4), the scaled conjugate gradient back propagation method is used to minimize the damage function and perform weight optimization, which has two stopping criteria, namely the accuracy of the objective function and the weight at the solution. In this embodiment, 10 -3 Stop criteria.
[0057] 5-5), K-fold cross validation is used to improve the selection of the global minimum. In this implementation, K=10.
[0058] In summary, the multi-layer neural network optimization method of the present invention can perform real-time detection and identification of thermal imaging signals in the friction stir welding area, improve the signal-to-noise ratio and flaw detection depth of metal thermal imaging images, and realize the application of infrared imaging in metal welding defect detection; it can be continuously iterated to improve detection efficiency.
[0059] The above specific implementation methods are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A multi-layer neural network optimization method for infrared thermal imaging defect recognition, Features: The process steps include: 1), thermal imaging data pre-processing; The specific steps are: 1-1), convert the original RGB image data into a 64-bit grayscale image, identify and crop the target area, remove the invalid background, and use Gaussian filtering to increase the contrast between defective pixels and non-defective pixels. The sample kernel is 13×13 and the standard deviation is 3.0; the Gaussian filtered image sequence is G(j,i,t), where G is the image at time t, j and i are the column number and row number respectively; the partial derivative of the image data after Gaussian filtering with respect to time is 1-2), convert the three-dimensional image data into a two-dimensional matrix N×P; wherein the number of rows N is the number of infrared image pixels, including spatial changes; the number of columns P records the number of infrared images, including temporal changes; 2) Extraction of friction stir welding defect features; the specific steps are: 2-1), linear modeling of thermal imaging signals, compressing the instantaneous thermal imaging image sequence into four images while retaining spatial and temporal information; 2-2), a univariate polynomial is used to perform Gaussian filtering on each pixel, and the grayscale value after filtering is a function of time as follows: G(j, i, t)≈a 0 (j,i)+a1(j,i)t, where a 0 (j, i) and a 1 (j, i) are polynomial parameters; the function of gray value slope changing with time is as follows: where b 0 (j, i) and b 1 (j, i) are polynomial parameters: Therefore, the three-dimensional grayscale data after Gaussian filtering is converted into two polynomial coefficient graphs formed by polynomial coefficients, namely a 0 (j, i) and a 1 (j, i); 2-3), the least square method is used to calculate the polynomial parameters (a 0 , a 1 , b 0 , b 1 ) is re-weighted, and when the accuracy reaches 10 -3 Stop the reweighting calculation when , and obtain the corresponding weight matrix and bias matrix, whose image grayscale range is [0,255]; 3), welding defect / non-defect image data preparation; 4), Multi-layer neural network modeling for infrared imaging; 5), Multi-layer neural network model optimization.
2. The multi-layer neural network optimization method for infrared thermal imaging defect recognition according to claim 1, Features: The specific steps of step 3) are: 3-1), the data preparation process includes data labeling, conversion, sampling and segmentation: First, the data is classified and labeled using the location and size of the defects; then, the input data features are arranged in columns, and the z-score normalization method is used to reduce the Euclidean distance between features so that the mean and unit standard deviation of each feature are zero; 3-2) Create a data set with non-defective and defective data samples in a ratio of 1:1; 3-3), randomly divide the data set into two equal parts to ensure sufficient training and test sets.
3. The multi-layer neural network optimization method for infrared thermal imaging defect recognition according to claim 1, Features: The specific steps of step 4) are: 4-1), a binary classification method [0,1] is used for supervised learning, where 0 represents no defect and 1 represents defect; 4-2), the input layer consists of 10 neurons, and the hidden layer includes a nonlinear activation function The output layer consists of one neuron and a logistic activation function composition.
4. The multi-layer neural network optimization method for infrared thermal imaging defect recognition according to claim 1, Features: The specific steps of step 5) are: 5-1), randomly select the initial weights of the neural network and use the cross entropy loss function as follows: Where n is the data point, y i and z i are the predicted value and target value of the i-th input vector respectively; 5-2), use L2 regularization to perform weight penalty, as follows: Where ω is the model weight vector; 5-3), the complete damage function after combining the cross entropy damage function and the L2 regularization method is as follows: total =J cc +ρJ reg , where ρ is the regularization parameter and ρ ≥ 0; 5-4), the scaled conjugate gradient back propagation method is used to minimize the damage function and perform weight optimization, which has two stopping criteria, namely the accuracy of the objective function and the weight at the solution; 5-5), K-fold cross validation is used to improve the selection of the global minimum.
5. The multi-layer neural network optimization method for infrared thermal imaging defect recognition according to claim 4, Features: In step 5-3), ρ = 0.225; in step 5-4), take 10 -3 is the stopping criterion; in step 5-5), K=10.
Citation Information
Patent Citations
Glass flaw detection method based on variational auto-encoder
CN113344903A
End-to-end saliency mapping via probability distribution prediction
US20170308770A1
Method of measuring blemishes on skin
WO2008028893A1
