A method for extracting features of temperature-space curves changing with time in ultrasonic infrared technology

Abstract

The present invention discloses a method for extracting the characteristics of temperature-space curves changing with time in ultrasonic infrared technology, comprising the following steps: collecting infrared thermal image sequences of defective test pieces and performing background reduction processing to obtain temperature rise heat maps, binarizing the temperature rise heat maps and obtaining binary images; classifying the binary images, obtaining the spatial value time series and spatiotemporal correlation coefficient sequence of each type of coordinate point in the marked area, and obtaining the composite sequence of each type of coordinate point through splicing processing; constructing a defect recognition model based on spatial attention, channel attention and SE attention mechanisms, and training the defect recognition model based on the composite sequence of each type of coordinate point; extracting characteristics of the test piece based on the trained defect recognition model, and identifying the defects of the test piece. The present invention can help identify different types of points by calculating the correlation coefficient sequence of coordinate points at each time, thereby increasing the accuracy and efficiency of defect recognition.

Description

Technical Field

[0001] The invention belongs to the technical field of nondestructive testing, and in particular relates to a method for extracting characteristics of a temperature-space curve changing with time in ultrasonic infrared technology. Background Art

[0002] Commonly used thermal excitation methods in infrared thermal imaging technology include flash lamp excitation, laser excitation, ultrasonic excitation, etc. Compared with other excitation methods, ultrasonic excitation is selective excitation, which only excites the defect area and does not excite the non-defect area. At the same time, ultrasonic excitation does not consider the uniformity of thermal excitation, and can achieve better detection results for complex structure specimens. In addition, ultrasonic excitation is an internal excitation and body excitation method, with a small heating area, which is helpful for the detection of small defects.

[0003] Ultrasonic infrared thermal imaging technology is a non-destructive testing technology, which mainly includes ultrasonic thermal excitation and infrared detection. First, ultrasonic waves are used as the excitation source and acted on the surface of the material or specimen to be tested. Ultrasonic waves are coupled into the object to be tested through the interface. If there are defects in the material or specimen to be tested, the defects will generate heat due to the thermoelastic effect, friction effect, etc., which will cause the temperature of the defect and the adjacent position to rise. Use an infrared thermal imager to observe and record these temperature changes, and collect infrared thermal image sequences. These infrared thermal image sequences can be processed to obtain the information of the corresponding defects. Ultrasonic infrared thermal imaging technology is now widely used in quality inspection and defect identification of various materials.

[0004] In recent years, with the vigorous development of deep learning, nondestructive testing methods based on deep learning have been more and more applied to the field of nondestructive testing. Unlike traditional machine learning methods, deep learning methods do not directly extract features from images, but use multi-level network structures to gradually map shallow features in images to deep features, thereby obtaining feature vectors that can express more complex semantic information. However, images obtained using ultrasonic infrared thermal imaging technology are often accompanied by disadvantages such as low imaging contrast and blurred defect boundaries, which are not conducive to feature processing using deep learning methods. At present, defect recognition in the field of infrared thermal imaging usually uses two-dimensional images or one-dimensional temperature-time signals as input, either losing time information or losing spatial information. Therefore, it is urgent to propose a feature extraction method for the time-varying temperature-space curve in ultrasonic infrared technology, which can better extract the time and space features in the ultrasonic infrared thermal image sequence, which is of great significance for the use of deep learning methods for defect recognition in the field of nondestructive testing. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a method for extracting features of temperature-space curves changing over time in ultrasonic infrared technology. Through effective feature extraction, deep learning methods are used to qualitatively identify defects of the test piece to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides a method for extracting the characteristics of temperature-space curve changing with time in ultrasonic infrared technology, comprising the following steps:

[0007] Collecting an infrared thermal image sequence of a defective specimen and performing background subtraction processing to obtain a temperature rise thermal map, and performing binarization processing on the temperature rise thermal map to obtain a binary image;

[0008] The binary image is labeled by category, and the spatial value time series and spatiotemporal correlation coefficient series of each coordinate point in the labeled area are obtained, and the composite sequence of each coordinate point is obtained by splicing.

[0009] Based on spatial attention, channel attention and SE attention mechanisms, a defect recognition model is constructed, and the defect recognition model is trained based on the composite sequence of each coordinate point;

[0010] Based on the trained defect recognition model, features of the test piece are extracted to identify defects of the test piece.

[0011] Optionally, the process of obtaining the temperature rise thermal map includes: taking the infrared thermal map subsequence of the first 6 frames from the infrared thermal image sequence, and averaging it as the background thermal map; presetting a certain moment, taking the infrared thermal map subsequence of 6 consecutive frames starting from the preset moment, and averaging it as the foreground thermal map; subtracting the background thermal map from the foreground thermal map of all time frames to obtain the temperature rise thermal map.

[0012] Optionally, the process of obtaining the spatial value time series of each coordinate point includes: obtaining the temperature-space curve of each coordinate point in the temperature rise thermal map, obtaining the maximum temperature value of the temperature-space curve, performing maximum value normalization on the maximum temperature value, and obtaining a normalized spatial sequence; summing the normalized spatial sequence to obtain a summation result, and based on the ratio of the summation result to the maximum temperature value, obtaining the spatial value time series within the overall time period, presetting the time period, and selecting the spatial value time series of each coordinate point within the preset time period.

[0013] Optionally, the process of obtaining the spatiotemporal correlation coefficient sequence for each coordinate point includes: obtaining the standard deviation of the experimental data, fitting the standard deviation of the experimental data to obtain a fitted standard deviation; obtaining a Gaussian-like sequence of the normalized spatial sequence based on the fitted standard deviation; obtaining the spatiotemporal correlation coefficient sequence of the normalized spatial sequence and the Gaussian-like sequence within the overall time period, presetting the time period, and obtaining the spatiotemporal correlation coefficient sequence within the preset time period.

[0014] Optionally, the process of obtaining the standard deviation of experimental data includes: selecting different defect signals of the same defective specimen, corresponding the different defect signals to normal signal intensity, stronger signal intensity and weaker signal intensity, respectively, and obtaining normalized spatial sequences under three different signal intensities, and then obtaining a normalized spatial sequence curve, respectively fitting a Gaussian-like curve of the normalized spatial sequence curve, and then obtaining the standard deviation of the Gaussian-like curve, and using the average of the standard deviations of the three Gaussian-like curves as the standard deviation of the experimental data.

[0015] Optionally, the process of processing data based on the defect recognition model includes: inputting the input data into Conv1 convolution and Conv2 convolution for convolution, the output is convolved through Conv3 and Conv4 respectively, the output of Conv3 and Conv4 is sent to the spatial attention and channel attention network respectively, at the same time, the output of Conv3 and Conv4 is spliced ​​and sent to the first Bottleneck layer, the output of the first Bottleneck layer is sent to the Conv5 convolution operation; the output of Conv5 is spliced ​​with the output of the spatial attention and channel attention respectively, and sent to Conv6, Conv7 and Conv8 respectively, and Conv The output of v7 is then concatenated with the outputs of Conv6 and Conv8, and the concatenated results are sent to the second and third Bottleneck layers respectively; the outputs of the second and third Bottleneck layers are respectively passed to the SE attention mechanism, and the outputs of the two SE attention mechanisms are concatenated and sent to the fourth Bottleneck layer. After that, the output of the fourth Bottleneck layer is passed to a pooling layer for adaptive average pooling operation, and then the output of the pooling layer is passed to a fully connected layer for result mapping, and finally the SoftMax function is used to classify and output the recognition result.

[0016] Optionally, the Conv1 convolution and the Conv2 convolution both include a convolution with a convolution kernel of 1 and a padding of 1 and a convolution with a convolution kernel of 3 and a padding of 1.

[0017] Optionally, the categories of coordinate points in the binary image include: defect points, bad paint points, support column points, heat reflection points and noise points.

[0018] The present invention also provides an electronic device, comprising: a memory and a processor; the memory is used to store a program; the processor is used to execute the program to implement each step of the method for extracting the characteristics of the temperature-space curve changing with time in ultrasonic infrared technology.

[0019] The present invention also provides a readable storage medium on which a computer program is stored. When the computer program is executed by a processor, each step of the method for extracting the characteristics of the temperature-space curve changing with time in ultrasonic infrared technology is realized.

[0020] Compared with the prior art, the present invention has the following advantages and technical effects:

[0021] (1) The present invention extracts the temperature-space sequence of the coordinate point and calculates its maximum normalized mean at a certain moment. The space values ​​at multiple times constitute the space value time series. For different types of points, their thermal reactions after ultrasonic excitation are different, and their space value time series curves also have different characteristics. By extracting the space value sequence, which contains not only spatial information but also time information, different types of points can be significantly distinguished;

[0022] (2) The present invention obtains the degree of similarity between the type of point and the defect point at that moment by fitting a quasi-Gaussian sequence of the normalized spatial sequence of the defect point at a certain time, and calculating the correlation coefficient with the normalized spatial sequence of the type of point to be identified at that moment. Since the normalized spatial sequence curve characteristics of the bad paint point and the defect point are similar, both are similar to the Gaussian curve, while the normalized spatial sequence curve characteristics of other types of points are very different from the Gaussian curve, the correlation coefficient sequence of the coordinate points at each time can help identify different types of points. At the same time, it can make up for the problem that the defect points are mistakenly judged as other types when the pixel points are classified only according to the spatial value sequence, thereby increasing the accuracy and efficiency of defect recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0024] Figure 1 A flow chart of a method for extracting characteristics of a temperature-space curve changing with time in an ultrasonic infrared technology according to an embodiment of the present invention;

[0025] Figure 2 It is a comparison diagram of normalized spatial sequence curves of five types of points in an embodiment of the present invention;

[0026] Figure 3 A comparison diagram of the normalized spatial sequence curves of badpaint points and defect points changing over time according to an embodiment of the present invention;

[0027] Figure 4 It is a comparison diagram of maximum normalized mean sequence curves of five types of points in an embodiment of the present invention;

[0028] Figure 5A comparison diagram of five types of point space coefficient sequence curves according to an embodiment of the present invention;

[0029] Figure 6 A graph showing the change of the standard deviation of the Gaussian-like curves of the normalized spatial sequence at the defect points of three test pieces taken in an embodiment of the present invention over time;

[0030] Figure 7 A comparison diagram of composite sequence curves of five types of points in an embodiment of the present invention;

[0031] Figure 8 A deep learning network structure diagram of an embodiment of the present invention;

[0032] Fig. 9 This is a structural diagram of the attention mechanism used in the deep learning network of an embodiment of the present invention;

[0033] Fig.10 It is a binary image after the threshold processing of the background temperature rise-reduced heat map corresponding to the tested object in an embodiment of the present invention;

[0034] Fig.11 This is a diagram of defect recognition results of the tested piece according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0036] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0037] Embodiment 1

[0038] In this embodiment, 153 samples made of titanium and chromium-nickel-inconel alloy were used, and the sample size was approximately: 200 mm long, 30 mm wide, and 6 to 8 mm thick. Before the experiment, each sample was manually spray-painted to eliminate surface heat reflection. There were several unknown defects and pseudo defects caused by spray painting in the 153 samples.

[0039] like Figure 1 As shown, this embodiment proposes a method for extracting features of temperature-space curves changing with time in ultrasonic infrared technology, comprising the following steps:

[0040] Infrared thermal image preprocessing:

[0041] Ultrasonic thermal excitation is performed on the test piece to collect the infrared thermal image sequence of the test piece, where the infrared thermal image sequence is recorded as D∈Rh×w×n , where h and w are the height and width of each frame of infrared thermal image, n represents the total number of frames of infrared thermal image sequence, n=120 in the embodiment, and T(x,y,t) represents the temperature of coordinate (x,y) at time t.

[0042] Background reduction:

[0043] In ultrasonic excitation infrared thermal imaging technology, defects generate heat under the action of ultrasound, but the heat is usually very small. When doing experiments at room temperature, the heat difference is difficult to be observed in the original thermal image. Therefore, it is necessary to perform background subtraction on the ultrasonic infrared thermal image sequence. First, take the infrared thermal image subsequence of the first 6 frames, average it and use it as the background thermal image. Similarly, for the thermal image at time t, take the average of the 6 consecutive frames of thermal images starting from the thermal image at time t as the foreground thermal image, and subtract the background thermal image from the foreground thermal image of all time frames to obtain the temperature rise thermal image sequence:

[0044]

[0045] in, is a subsequence of 6 consecutive infrared thermal images starting from the thermal image at time t, It is the infrared thermal image subsequence of the first 6 frames.

[0046] The temperature rise heat map obtained by background subtraction is characterized by the fact that no heat is generated in the normal defect-free area. After the foreground heat map is subtracted from the background heat map, the result is noise. However, the foreground heat map of the defect area is subtracted from the background heat map. Since the background heat map is before ultrasonic excitation, it is a pure noise image. The result obtained after subtraction directly reflects the amount of heat generated. Therefore, the contrast of the defect signal in the heat map can be effectively improved by background subtraction processing. Subsequent processing is based on the temperature rise heat map, i.e., f(x, y, t).

[0047] Get the temperature-space curve:

[0048] For a pixel point (x0, y0) in the heat map sequence, its pixel value at time t is f(x0, y0, t), and the temperature-space curve corresponding to the pixel point is defined as the temperature value of 10 pixels on the left and right (l0=9 in the embodiment) centered at x0, expressed as:

[0049]

[0050] For the coordinate points in the boundary range, since there may be less than 9 pixels on the left or right side, they need to be replaced with the pixels at the boundary.

[0051] Find the maximum normalized mean:

[0052] Max Normalization:

[0053] Find the maximum temperature value of the temperature-space curve corresponding to the pixel point (x0, y0) to be processed

[0054] make

[0055] right Perform maximum normalization to obtain a normalized spatial sequence

[0056]

[0057] When the coordinate points (x0, y0) belong to different categories, their normalized spatial sequence curves have different characteristics, such as Figure 2 As shown in the figure, for bad paint points and defect points, the curve characteristics of the two are similar to the Gaussian curve. At the same time, by analyzing the normalized spatial sequence curves of various types of points at different times, it can be found that with the advancement of time, the curve of the defect point shows a trend of gradually widening, while the curve of the bad paint point remains almost unchanged, as shown in the figure. Figure 3 As shown in the figure, for the support column (post), heat reflection (heatReflection) point and noise (noise) point, the curve characteristics have a certain randomness, and the similarity with the Gaussian curve is small, and the value of the normalized spatial sequence of the noise point will increase with time. This change is reflected in the sequence curve that the curve will "rise" with time, while the trend of the sequence curve of the heat reflection point and the support column point changing with time is relatively random;

[0058] Normalized mean sum:

[0059] After summing the normalized spatial sequence, we get

[0060]

[0061] After maximum normalization, for defect signals, the temperature-space curves at different times all become similar to Gaussian curves, which is equivalent to the independence of signals at different times. However, for infrared thermal imaging technology, it is based on heat conduction theory, and it not only follows the law of heat conduction in space, but also in time. The normalized average contains certain spatial heat conduction information. In order to further reflect the time information, it is necessary not only to use the normalized average of all time points, but also to divide these values ​​by the maximum temperature value at the corresponding moment to obtain the maximum normalized mean, hereinafter referred to as the spatial value time series.

[0062]

[0063] Through the above processing, a complete spatial value time series is obtained. However, due to further division by the maximum value of the temperature-space curve, for the thermal map before and after ultrasonic excitation, it may not only be the noise background area, but the defect area signal may also be approximately equal to the noise. Therefore, the results obtained by using this value as the divisor may be many and may also fluctuate greatly. Therefore, only a section of the spatial value time series is taken for subsequent processing. It is assumed that the spatial values ​​within the range of time t1 to t2 constitute the spatial value time series of the coordinate point (x0, y0)

[0064] In this embodiment, t1=30, t2=100 are taken to observe the spatial value time series of different types of points, such as Figure 4 As shown in the figure, the spatial value of defective points will gradually increase with time, while for bad paint points, their spatial value will basically remain unchanged with time; at the same time, the spatial value of noise points will also increase with time, but there will not be a large fluctuation; in addition, the spatial values ​​of heat reflection points and support column points tend to fluctuate greatly with time; it is easy to know that most heat reflection points and support column points can be identified by analyzing the spatial value, and at the same time, the bad paint points and defective points with similar characteristics of some spatial sequence curves can be effectively distinguished;

[0065] Find the spatial correlation coefficient:

[0066] The spatial value time series obtained through the above steps contains not only spatial information but also time information. The basic idea is that the defect signal is similar to the shape of a Gaussian curve, and the greater the heat diffusion over time, that is, the greater the standard deviation of the Gaussian curve. Reflected in the curve, there is a feature that the normalized mean increases over time. Therefore, the present invention uses the spatial value time series obtained by the above processing as a feature for defect identification. However, in ultrasonic excitation infrared thermal imaging technology, there are various interference signals. These interference signals may have similar spatial value time series as defect signals. For example, for heat reflection interference signals, although their temperature-space curve and Gaussian curve may be very different, their mean may be similar to that of the defect signal. Therefore, the present invention proposes to further calculate the correlation coefficient between the temperature-space curve and the Gaussian curve, and use the spatial value time series as a feature signal for defect identification.

[0067] Find the relationship between the standard deviation of the experimental data temperature-space curve and time:

[0068] For the normalized spatial sequence of coordinates (x0, y0) in the heat map When (x0, y0) belongs to different regions, its normalized spatial sequence curve has different characteristics; the normalized spatial sequence curves of typical defect points and poor paint points are similar to the Gaussian curve, and the degree of change of the sequence curves of the two over time has certain differences. This difference is Figure 3 is reflected in.

[0069] For the normalized spatial sequence curves of noise points, heat reflection points and support column points, although their curve characteristics are quite different from those of the Gaussian curve, when the spatial value operation is performed on them, the spatial values ​​obtained will be equal to the spatial values ​​of the bad paint points and defect points, resulting in the spatial value sequences of some non-defective and non-bad paint points being similar to the spatial value sequences of defective and bad paint points. If the pixel points are classified only based on the spatial value sequence, the classification accuracy will be low and many types of points will be misjudged. This problem can be effectively solved by further extracting the characteristics of the temperature-space sequence curve changing over time.

[0070] Since the normalized spatial sequence curves of bad paint spray points and defect points are similar to Gaussian curves, and the normalized spatial sequence curves of defect points tend to become wider over time, the correlation coefficient between the normalized spatial sequence of type points and the Gaussian-like sequence of defect points is calculated at a certain time. The characteristics of the temperature-space sequence curve changing with time can be further obtained from the correlation coefficient sequence of type points at different times. Figure 5 As shown in the figure, the correlation coefficients of bad paint points and defective points are larger than those of other types of points, and there is no negative correlation coefficient. At the same time, because the normalized spatial sequence curve of bad paint points basically does not change with time, it can be seen that the correlation coefficient obtained with the Gaussian curve whose standard deviation increases with time will decrease with time. For noise points, heat reflection points and support column points, because their normalized spatial sequence curves fluctuate up and down and are very different from the Gaussian curve, the correlation coefficients obtained are both positive and negative.

[0071] To find the correlation coefficient between the temperature-space curve and the Gaussian curve, it is necessary to first determine the standard deviation of the Gaussian curve, and then simulate the Gaussian curve corresponding to the temperature-space curve based on the standard deviation. From the above analysis, it can be seen that the heat diffusion at the defect is not only reflected in space, but also in time, so the standard deviation of the temperature-space curve at different times is different. In order to obtain the relationship between the standard deviation of the temperature-space curve and time, methods such as theoretical simulation and finite element simulation can be used, but these methods are only approximations of actual experiments. In order to obtain standard deviation data that is closer to the actual experimental value, the present invention directly extracts different experimental data and calculates its standard deviation.

[0072] Defective specimens are taken from titanium and Inconel alloy samples respectively, and the temperature rise sequences at their defect locations are analyzed. Since the specimen materials are different, the maximum temperature rise at the defect locations, hereinafter referred to as the temperature rise value, will also vary; at the same time, it is found that even if the specimen materials are the same, the temperature rise values at their defect locations will also vary due to reasons such as defect size and depth. Two defect pixel points with similar temperature rise values in the two types of specimens are taken, and their normalized spatial sequences are analyzed. It is found that the normalized spatial sequence curves at the same moment are not very different, that is, the corresponding standard deviation differences are not very large. Therefore, it can be approximately considered that the relationship between the standard deviation and time change is independent of the specimen material.

[0073] Furthermore, several different defect signals of one of the materials, corresponding to normal signal intensity, stronger signal intensity, and weaker signal intensity respectively, are taken, and the normalized spatial sequences of these three signals at time t are obtained according to the above steps. Furthermore, the normalized spatial sequence curves are obtained, and the Gaussian-like curves G1(u, w1 t ), G2(u, w2 t ), G3(u, w3 t ) of their normalized spatial sequence curves are respectively fitted, where the standard deviations of the Gaussian-like curves are w1 t , w2 t , w3 t , where u = 9;

[0074] The data from frame t1 to t2 with relatively stable temperature rise at the defect point are taken for analysis, 1 ≤ t1 < t2 ≤ n - 5. In this embodiment, t1 = 30 and t2 = 100 are taken;

[0075] Let where the sequence curves of W1, W2, and W3 are as shown in Figure 6 ; By comparing the three signals with different intensities, we find that the corresponding standard deviation-time curve rules are the same: increasing with time and tending to saturation after a certain time; in addition, the standard deviation-time curves of the normal signal and the stronger signal are basically the same, and the standard deviation value of the weaker signal is relatively small. Since the standard deviation-time curve change rules of different intensity signals are the same, and for defect signal recognition, this change rule is more important. Therefore, in the present invention, the average value of the standard deviations of these three signals is used as the standard deviation of the experimental data, and the standard deviations at all moments are calculated to obtain the sequence

[0076]

[0077] Calculate the correlation coefficient:

[0078] Since there is a certain fluctuation in the standard deviation time series obtained from the experimental data, therefore, for Make a polynomial fit to get w(t), and use w(t) as the function of the standard deviation of the Gaussian curve of the normalized spatial sequence at the defect point with respect to time t; the Gaussian sequence of the normalized spatial sequence of the defect pixel point at time t is obtained from w(t) and the fixed value u

[0079]

[0080] Find the normalized spatial sequence of the point (x0, y0) in the infrared thermal image at time t and Find the Pearson correlation coefficient make X i but It can be expressed by the following formula:

[0081]

[0082] in are the means of X and Y respectively;

[0083] Take the correlation coefficients from time t1 to t2 to form the spatiotemporal correlation coefficient sequence of point (x0, y0) In this embodiment, t1=30, t2=100:

[0084]

[0085] Deep learning recognition:

[0086] Dataset preparation:

[0087] The defective specimens are divided into training specimens and test specimens in a ratio of 7:3. The infrared thermal images are preprocessed according to the above steps to obtain infrared thermal images after background subtraction. The images after background subtraction are threshold processed and the temperature rise thermal images after background subtraction are segmented to remove most of the background and noise to obtain binary images.

[0088] The binary image is labeled and the regions of interest are divided into five categories, including defect, bad paint, support column, heat reflection and noise. Defect, bad paint, support column and heat reflection are shown as white areas, and noise is shown as black areas.

[0089] According to the above steps, calculate the spatial value time series of the coordinate point (x0, y0) in the specimen marking area at time t And the correlation coefficient series The composite sequence is obtained by splicing the two

[0090] The comparison of composite sequence curves of five types of points is as follows Figure 7 As shown, the composite sequences of all coordinate points in the annotated areas of the training specimens and the test specimens are calculated respectively, and they are used as the training data and test data of the deep learning model respectively.

[0091] Defect recognition model construction and training:

[0092] The model includes multiple convolutional layers, pooling layers, attention mechanisms, and fully connected layers;

[0093] The overall network structure is as follows Figure 8 As shown in the figure, the input data first undergoes a convolution with a kernel of 1 and padding of 1 and a convolution with a kernel of 3 and padding of 1. The purpose of padding is to keep the spatial size of the output feature map the same as the input. After that, it undergoes a convolution with a kernel of 1 to reduce the number of channels, and then is normalized through a batch normalization layer and a ReLU activation function is applied. These operations together constitute Conv1_1x1 and Conv1_3x3 convolutions; similarly, Conv2_1x1 and Conv2_3x3 operations are performed. Then, the output of Conv2 is convolved through Conv3_1x1 and Conv4_1x1, respectively, and batch normalization and activation functions are applied. Next, the outputs of Conv3 and Conv4 are sent to the SpatialAttention (SA) and ChannelAttention (CA) networks, respectively. At the same time, the outputs of Conv3 and Conv4 are concatenated and fed into the first Bottleneck layer (abbreviated as Bott1), which contains 1x1 convolution to reduce the number of channels.

[0094] The output of the first Bottleneck layer is sent to the Conv5 convolution operation, which includes three one-dimensional convolution layers with different convolution kernels, batch normalization operations and activation functions to extract and enhance the features of the input data; the output of Conv5 is spliced ​​with the output of spatial attention and channel attention, and sent to Conv6 and Conv8 respectively; in addition, the output of Conv5 is also input to Conv7, and the output of Conv7 is spliced ​​with the output of Conv6 and Conv8 respectively, and the spliced ​​results are sent to the second Bottleneck layer (abbreviated as Bott2) and the third Bottleneck layer (abbreviated as Bott3) to reduce the amount of calculation of network parameters; then the outputs of the second and third Bottleneck layers are passed to SE-Attention (Squeeze-and-Excitation The outputs of the two SE-Attention networks are concatenated and sent to the fourth Bottleneck layer (Bott4 for short). The output of the fourth Bottleneck layer is then passed to a pooling layer for adaptive average pooling, and the input tensor is adaptively pooled into an output tensor of a specified size. Finally, the output of the pooling layer is passed to a fully connected layer to map the pooled features to the final classification results. Finally, the SoftMax function is used to convert the probabilities of each category into a one-hot encoding format for classification output.

[0095] Three attention feature extraction networks are used, namely spatial attention (SpatialAttention), channel attention (ChannelAttention) and SE attention (Squeeze-and-Excitation Networks) mechanisms. The network structure is as follows Fig. 9 .

[0096] When training the defect recognition model, the processed training data is loaded into the model for training, and the category is determined based on the one-hot encoding derived from the model. The number of training rounds is set, and the model is optimized by updating the network parameters in each round of iteration.

[0097] Defect Identification:

[0098] During the test, first generate an all-zero matrix M of the same size as the temperature rise heat map, extract the composite sequence of all points from the test sample, and input them into the trained model in sequence. If a pixel is judged as a defect, take out the position of the horizontal and vertical coordinates in the binary image to which the sample belongs, and set the corresponding position in the all-zero matrix M to 255. From the image point of view, the corresponding point of the all-zero matrix image is turned white.

[0099] After the above steps, the identified target area is a defect. If there is a defect in the specimen, the image of the matrix M is white at the defect location. If there is no defect in the specimen, the image of the matrix M is all black.

[0100] In this embodiment, a test specimen with defects is selected from the chromium-nickel-iron alloy specimen class, and its background-reduced temperature rise thermal map is processed by threshold value and is shown in the attached figure. Fig.10 As shown in the figure, the white part in the center of the image is a defect; after the above defect detection steps, the final result is as shown in the attached figure. Fig.11 As shown, the defects on the specimen were accurately identified.

[0101] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for extracting the characteristics of temperature-space curve changing with time in ultrasonic infrared technology, characterized in that: The following steps are involved: Collecting an infrared thermal image sequence of a defective specimen and performing background subtraction processing to obtain a temperature rise thermal map, and performing binarization processing on the temperature rise thermal map to obtain a binary image; The binary image is labeled by category, and the spatial value time series and spatiotemporal correlation coefficient series of each coordinate point in the labeled area are obtained, and the composite sequence of each coordinate point is obtained by splicing. Based on spatial attention, channel attention and SE attention mechanisms, a defect recognition model is constructed, and the defect recognition model is trained based on the composite sequence of each coordinate point; Extract features of the test piece based on the trained defect recognition model to identify defects of the test piece; The process of obtaining the spatial value time series of each coordinate point includes: obtaining the temperature-space curve of each coordinate point in the temperature rise thermal map, obtaining the maximum temperature value of the temperature-space curve, performing maximum value normalization on the maximum temperature value, and obtaining a normalized spatial sequence; summing the normalized spatial sequence to obtain a summation result, and based on the ratio of the summation result to the maximum temperature value, obtaining the spatial value time series within the overall time period, presetting the time period, and selecting the spatial value time series of each coordinate point within the preset time period; The process of obtaining the spatiotemporal correlation coefficient sequence of each coordinate point includes: obtaining the standard deviation of the experimental data, performing fitting processing on the standard deviation of the experimental data to obtain the fitted standard deviation; obtaining a Gaussian-like sequence of the normalized spatial sequence based on the fitted standard deviation; obtaining the spatiotemporal correlation coefficient sequence of the normalized spatial sequence and the Gaussian-like sequence within the overall time period, presetting the time period, and obtaining the spatiotemporal correlation coefficient sequence within the preset time period; The process of building a defect recognition model includes: inputting the input data into Conv1 convolution and Conv2 convolution for convolution, and the output is convolved through Conv3 and Conv4 respectively, and the outputs of Conv3 and Conv4 are sent to the spatial attention and channel attention networks respectively. At the same time, the outputs of Conv3 and Conv4 are spliced ​​and sent to the first Bottleneck layer, and the output of the first Bottleneck layer is sent to the Conv5 convolution operation; the output of Conv5 is spliced ​​with the outputs of spatial attention and channel attention respectively, and sent to Conv6, Conv7 and Conv8 respectively, and the input of Conv7 is The outputs are then spliced ​​with the outputs of Conv6 and Conv8, and the spliced ​​results are sent to the second and third Bottleneck layers respectively; the outputs of the second and third Bottleneck layers are respectively passed to the SE attention mechanism, and the outputs of the two SE attention mechanisms are spliced ​​and sent to the fourth Bottleneck layer, and then the output of the fourth Bottleneck layer is passed to a pooling layer for adaptive average pooling operation, and then the output of the pooling layer is passed to a fully connected layer for result mapping, and finally the SoftMax function is used to classify and output the recognition result; The Conv1 convolution and the Conv2 convolution both include a convolution with a convolution kernel of 1 and a padding of 1 and a convolution with a convolution kernel of 3 and a padding of 1.

2. The method for extracting the characteristics of temperature-space curve changing with time in ultrasonic infrared technology according to claim 1 is characterized in that: The process of obtaining the temperature rise thermal map includes: taking the infrared thermal map subsequence of the first 6 frames from the infrared thermal image sequence, and taking the average as the background thermal map; presetting a certain moment, taking the infrared thermal map subsequence of 6 consecutive frames starting from the preset moment, and taking the average as the foreground thermal map; subtracting the background thermal map from the foreground thermal map of all time frames to obtain the temperature rise thermal map.

3. The method for extracting the characteristics of temperature-space curve changing with time in ultrasonic infrared technology according to claim 1 is characterized in that: The process of obtaining the standard deviation of experimental data includes: selecting different defect signals of the same defective specimen, corresponding the different defect signals to normal signal intensity, strong signal intensity and weak signal intensity respectively, and obtaining normalized spatial sequences under three different signal intensities, and then obtaining a normalized spatial sequence curve, respectively fitting a Gaussian-like curve of the normalized spatial sequence curve, and then obtaining the standard deviation of the Gaussian-like curve, and taking the average of the standard deviations of the three Gaussian-like curves as the standard deviation of the experimental data.

4. The method for extracting the characteristics of temperature-space curve changing with time in ultrasonic infrared technology according to claim 1 is characterized in that: The coordinate point categories in the binary image include: defect points, poorly painted points, support column points, heat reflection points and noise points.

5. An electronic device, characterized in that: include: Memory and processor; The memory is used to store programs; The processor is used to execute the program to implement each step of the method for extracting the characteristics of the temperature-space curve changing with time in the ultrasonic infrared technology as described in any one of claims 1-4.

6. A readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the method for extracting the characteristics of the temperature-space curve changing with time in the ultrasonic infrared technology according to any one of claims 1 to 4 is implemented.

Images