Forge piece crack recognition and positioning method and system based on binocular vision

Abstract

The invention relates to the technical field of forging crack recognition, in particular to a forging crack recognition positioning method and system based on binocular vision, and the method comprises the steps: obtaining a first turbine disc static image set and a first turbine disc dynamic image set of a turbine disc in a static state and a motion state through a binocular camera and a binocular high-speed camera respectively; constructing a turbine disc crack recognition model to recognize the two image sets, and obtaining a recognition result by extracting a first static crack feature and a second dynamic crack feature and generating a first dynamic and static crack fusion feature; and the initial learning rate of the turbine disc crack recognition model is improved according to the crack recognition error, the actual crack data and a preset error threshold, and a self-adaptive learning rate based on the crack recognition error is obtained. According to the method, the turbine disc crack identification model is improved by effectively combining static and dynamic image information of the turbine disc and combining the self-adaptive learning rate based on the crack identification error, and the turbine disc crack identification and positioning accuracy is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of forging crack identification, and specifically provides a method and system for forging crack identification and positioning based on binocular vision. Background Art

[0002] The high-pressure turbine disk is one of the core components of an aero-engine, and its quality inspection is of great significance for improving the safety and service life of the aero-engine. To ensure the quality of the high-pressure turbine disk, accurate crack identification and positioning are required.

[0003] For the crack identification and positioning of high-pressure turbine disks, traditional crack identification and positioning methods often perform identification in a single state, without fully considering the comprehensive analysis of images of the turbine disk in both moving and stationary states, which is likely to limit the comprehensiveness of high-pressure turbine disk crack identification and positioning, thus resulting in limited accuracy of high-pressure turbine disk crack identification and positioning.

[0004] Therefore, a method and system for forging crack identification and positioning based on binocular vision are proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for forging crack identification and positioning based on binocular vision. The first static image set and the first dynamic image set of the turbine disk in the stationary state and the moving state are respectively obtained through a binocular camera and a binocular high-speed camera; a turbine disk crack identification model is constructed to identify these two image sets. By extracting the first static crack feature and the second dynamic crack feature, and generating the first static-dynamic crack fusion feature, the identification result is obtained; furthermore, the initial learning rate of the turbine disk crack identification model is improved according to the crack identification error, the actual crack data, and the preset error threshold, and an adaptive learning rate based on the crack identification error is obtained. This method effectively combines the static and dynamic image information of the turbine disk, and improves the turbine disk crack identification model by combining the adaptive learning rate based on the crack identification error, effectively improving the accuracy of turbine disk crack identification and positioning.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] A method for forging crack identification and positioning based on binocular vision, comprising:

[0008] S1. Respectively obtain the first static image set and the first dynamic image set of the high-pressure turbine disk in the stationary state and the moving state through a binocular camera and a binocular high-speed camera;

[0009] S2. Construct a turbine disk crack recognition model to recognize the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk. The turbine disk crack recognition model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, a fusion feature analysis layer, and an output layer; the image preprocessing layer preprocesses the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk respectively to obtain the second set of static images of the turbine disk and the second set of dynamic images of the turbine disk; the crack feature extraction layer extracts features from the second set of static images of the turbine disk and the second set of dynamic images of the turbine disk through a static crack feature recognition unit and a dynamic crack feature recognition unit respectively to obtain the first static crack feature and the second dynamic crack feature; the crack feature fusion layer fuses the first static crack feature and the second dynamic crack feature to obtain the first static-dynamic crack fusion feature; the fusion feature analysis layer is used to analyze based on the first static-dynamic crack fusion feature to obtain the crack location recognition result.

[0010] S3. Improve the initial learning rate of the turbine disk crack recognition model based on the crack location recognition result, the actual crack data of the turbine disk, and a preset error threshold to obtain an improved turbine disk crack recognition model. The improved turbine disk crack recognition model includes an adaptive learning rate based on the crack recognition error.

[0011] Preferably, the first set of static images of the turbine disk includes static images of the turbine disk; the first set of dynamic images of the turbine disk includes dynamic images of the turbine disk.

[0012] Preferably, the turbine disk crack recognition model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, and an output layer;

[0013] The input layer is used to input the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk into the turbine disk crack recognition model;

[0014] The image preprocessing layer is used to perform image preprocessing on the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk to obtain the second set of static images of the turbine disk and the second set of dynamic images of the turbine disk; the preprocessing process includes graying and denoising the first set of static images of the turbine disk and the second set of dynamic images of the turbine disk.

[0015] The crack feature extraction layer is used to extract features from the second set of static images of the turbine disk and the second set of dynamic images of the turbine disk according to the YOLO model respectively to obtain the first static crack feature and the second dynamic crack feature; the first static crack feature includes static crack length, static crack width, static crack position, and static crack area; the second dynamic crack feature includes dynamic crack length, dynamic crack width, dynamic crack position, and dynamic crack area.

[0016] The crack feature fusion layer obtains the first static and dynamic crack fusion feature by performing feature fusion on the first static crack feature and the second dynamic crack feature; the first static and dynamic crack fusion feature includes the comprehensive crack length, comprehensive crack width, comprehensive crack position, and comprehensive crack area;

[0017] The fusion feature analysis layer obtains the crack positioning and recognition result according to the first static and dynamic crack fusion feature; the crack positioning and recognition result includes the position, length, width, and area of each crack identified for the turbine disk;

[0018] The output layer is used to output the crack positioning and recognition result.

[0019] Preferably, the first static and dynamic crack fusion feature is:

[0020] TFCODS = ο1 * TFSC + ο2 * TFDC;

[0021] Wherein, TFCODS represents the first static and dynamic crack fusion feature; o1 represents the influence weight of the static crack feature; TFSC represents the first static crack feature; o2 represents the influence weight of the dynamic crack feature; TFDC represents the second dynamic crack feature.

[0022] Preferably, the adaptive learning rate based on the crack recognition error is:

[0023]

[0024] Wherein, ALRCRE represents the adaptive learning rate based on the crack recognition error; LRRE represents the initial learning rate; μ represents the learning rate adjustment factor; λ1 represents the first crack recognition error influence coefficient, reflecting the influence of the turbine disk crack position recognition error probability on the learning rate adjustment; LE represents the turbine disk crack position recognition error probability; λ2 represents the second crack recognition error influence coefficient, representing the influence of the crack length recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EL represents the crack length recognition error probability in the cracks with correct turbine disk crack position recognition; λ3 represents the third crack recognition error influence coefficient, representing the influence of the crack width recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EW represents the crack width recognition error probability in the cracks with correct turbine disk crack position recognition; λ4 represents the fourth crack recognition error influence coefficient, representing the influence of the crack area recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; ES represents the crack area recognition error probability in the cracks with correct turbine disk crack position recognition; ERE represents the preset error threshold.

[0025] A forging crack recognition and positioning system based on binocular vision, comprising:

[0026] An image acquisition module, configured to respectively acquire a first set of static images of the high-pressure turbine disk in a stationary state and a first set of dynamic images of the high-pressure turbine disk in a moving state through a binocular camera and a binocular high-speed camera;

[0027] A model analysis module, configured to construct a turbine disk crack identification model to identify the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk. The turbine disk crack identification model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, a fusion feature analysis layer, and an output layer. The image preprocessing layer respectively preprocesses the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk to obtain a second set of static images of the turbine disk and a second set of dynamic images of the turbine disk. The crack feature extraction layer respectively extracts features from the second set of static images of the turbine disk and the second set of dynamic images of the turbine disk through a static crack feature recognition unit and a dynamic crack feature recognition unit to obtain a first static crack feature and a second dynamic crack feature. The crack feature fusion layer fuses the first static crack feature and the second dynamic crack feature to obtain a first static and dynamic crack fusion feature. The fusion feature analysis layer is configured to analyze based on the first static and dynamic crack fusion feature to obtain a crack location identification result;

[0028] A model optimization module, which improves the initial learning rate of the turbine disk crack identification model based on the crack location identification result, the actual crack data of the turbine disk, and a preset error threshold to obtain an improved turbine disk crack identification model. The improved turbine disk crack identification model includes an adaptive learning rate based on the crack identification error.

[0029] Preferably, the first set of static images of the turbine disk includes static images of the stationary turbine disk; the first set of dynamic images of the turbine disk includes dynamic images of the moving turbine disk.

[0030] Preferably, the turbine disk crack identification model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, and an output layer;

[0031] The input layer is configured to input the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk into the turbine disk crack identification model;

[0032] The image preprocessing layer is configured to perform image preprocessing on the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk to obtain a second set of static images of the turbine disk and a second set of dynamic images of the turbine disk. The preprocessing process includes graying and denoising the first set of static images of the turbine disk and the second set of dynamic images of the turbine disk;

[0033] The crack feature extraction layer is used to extract features from the second turbine disk static image set and the second turbine disk dynamic image set respectively according to the YOLO model, and obtain the first static crack features and the second dynamic crack features; the first static crack features include static crack length, static crack width, static crack position and static crack area; the second dynamic crack features include dynamic crack length, dynamic crack width, dynamic crack position and dynamic crack area;

[0034] The crack feature fusion layer obtains the first static and dynamic crack fusion features by fusing the first static crack features and the second dynamic crack features; the first static and dynamic crack fusion features include comprehensive crack length, comprehensive crack width, comprehensive crack position and comprehensive crack area;

[0035] The fusion feature analysis layer obtains the crack location and recognition result according to the first static and dynamic crack fusion features; the crack location and recognition result includes the position, length, width and area of each crack identified for the turbine disk;

[0036] The output layer is used to output the crack location and recognition result.

[0037] Preferably, the first static and dynamic crack fusion features are:

[0038] TFCODS = ο1 * TFSC + ο2 * TFDC;

[0039] Wherein, TFCODS represents the first static and dynamic crack fusion features; o1 represents the influence weight of static crack features; TFSC represents the first static crack features; o2 represents the influence weight of dynamic crack features; TFDC represents the second dynamic crack features.

[0040] Preferably, the adaptive learning rate based on the crack recognition error is:

[0041]

[0042] Among them, ALRCRE represents the adaptive learning rate based on crack recognition error; LRRE represents the initial learning rate; μ represents the learning rate adjustment factor; λ1 represents the first crack recognition error influence coefficient, reflecting the influence of the crack position recognition error probability of the turbine disk on the learning rate adjustment; LE represents the crack position recognition error probability of the turbine disk; λ2 represents the second crack recognition error influence coefficient, representing the influence of the crack length recognition error in the cracks with correct crack position recognition of the turbine disk on the learning rate adjustment; EL represents the crack length recognition error probability in the cracks with correct crack position recognition of the turbine disk; λ3 represents the third crack recognition error influence coefficient, representing the influence of the crack width recognition error in the cracks with correct crack position recognition of the turbine disk on the learning rate adjustment; EW represents the crack width recognition error probability in the cracks with correct crack position recognition of the turbine disk; λ4 represents the fourth crack recognition error influence coefficient, representing the influence of the crack area recognition error in the cracks with correct crack position recognition of the turbine disk on the learning rate adjustment; ES represents the crack area recognition error probability in the cracks with correct crack position recognition of the turbine disk; ERE represents the preset error threshold.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. The present invention respectively obtains the first turbine disk static image set and the first turbine disk dynamic image set of the high-pressure turbine disk in a static state and a dynamic state through a binocular camera and a binocular high-speed camera; and constructs a turbine disk crack recognition model to identify the first turbine disk static image set and the first turbine disk dynamic image set. The crack feature extraction layer of the turbine disk crack recognition model respectively extracts the features of the second turbine disk static image set and the second turbine disk dynamic image set through a static crack feature recognition unit and a dynamic crack feature recognition unit, and respectively obtains the first static crack feature and the second dynamic crack feature; by extracting the static crack feature and the dynamic crack feature of the turbine disk image, a good data foundation is laid for subsequent feature fusion, thereby effectively improving the accuracy of crack recognition and positioning of the high-pressure turbine disk.

[0045] 2. The crack feature fusion layer of the turbine disk crack recognition model of the present invention fuses the first static crack feature and the second dynamic crack feature to obtain the first static and dynamic crack fusion feature; based on the static image information and dynamic image information of the turbine disk, the comprehensiveness of crack recognition and positioning of the high-pressure turbine disk can be improved, thereby effectively improving the accuracy of crack recognition and positioning of the high-pressure turbine disk.

[0046] 3. The present invention improves the initial learning rate of the turbine disk crack recognition model based on the crack positioning recognition result, the actual crack data of the turbine disk, and the preset error threshold, and obtains an adaptive learning rate based on the crack recognition error. By adaptively improving the initial learning rate of the turbine disk crack recognition model, the recognition accuracy of the turbine disk crack recognition model can be effectively improved, thereby effectively improving the accuracy of crack recognition and positioning of the high-pressure turbine disk. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic flowchart of a method for identifying and positioning forging cracks based on binocular vision provided by an embodiment of the present invention;

[0048] Figure 2 It is a schematic structural diagram of a turbine disk crack recognition model provided by an embodiment of the present invention;

[0049] Figure 3 It is a schematic structural diagram of a system for identifying and positioning forging cracks based on binocular vision provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0051] Embodiment 1

[0052] In order to improve the accuracy of crack recognition and positioning of the A high-pressure turbine disk, a method for identifying and positioning forging cracks based on binocular vision is applied;

[0053] Figure 1 It is a schematic flowchart of a method for identifying and positioning forging cracks based on binocular vision provided by an embodiment of the present invention, including:

[0054] S1. Respectively obtain a first set of static turbine disk images and a first set of dynamic turbine disk images of the high-pressure turbine disk in a static state and a dynamic state through a binocular camera and a binocular high-speed camera; wherein, the rotational speed of the turbine disk is 15,000 revolutions per minute; the frame rate of the binocular high-speed camera is 5,000 frames per second;

[0055] Further, the first set of static turbine disk images includes static turbine disk images; the first set of dynamic turbine disk images includes dynamic turbine disk images.

[0056] S2. Construct a turbine disk crack recognition model to recognize the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk. The turbine disk crack recognition model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, a fused feature analysis layer, and an output layer; the image preprocessing layer preprocesses the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk respectively to obtain the second set of static images of the turbine disk and the second set of dynamic images of the turbine disk; the crack feature extraction layer extracts features from the second set of static images of the turbine disk and the second set of dynamic images of the turbine disk through a static crack feature recognition unit and a dynamic crack feature recognition unit respectively to obtain the first static crack feature and the second dynamic crack feature; the crack feature fusion layer fuses the first static crack feature and the second dynamic crack feature to obtain the first static-dynamic crack fused feature; the fused feature analysis layer is used to analyze according to the first static-dynamic crack fused feature to obtain the crack location recognition result;

[0057] Further, the turbine disk crack recognition model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, and an output layer; as Figure 2 is a schematic structural diagram of a turbine disk crack recognition model provided by an embodiment of the present invention;

[0058] The input layer is used to input the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk into the turbine disk crack recognition model;

[0059] The image preprocessing layer is used to perform image preprocessing on the first set of static images of the turbine disk and the first set of dynamic images of the turbine disk to obtain the second set of static images of the turbine disk and the second set of dynamic images of the turbine disk; the preprocessing process includes graying and denoising the first set of static images of the turbine disk and the second set of dynamic images of the turbine disk;

[0060] The crack feature extraction layer is used to extract features from the second set of static images of the turbine disk and the second set of dynamic images of the turbine disk respectively according to the YOLO model to obtain the first static crack feature and the second dynamic crack feature; the first static crack feature includes static crack length, static crack width, static crack position, and static crack area; the second dynamic crack feature includes dynamic crack length, dynamic crack width, dynamic crack position, and dynamic crack area;

[0061] The crack feature fusion layer obtains the first static-dynamic crack fused feature by performing feature fusion on the first static crack feature and the second dynamic crack feature; the first static-dynamic crack fused feature includes comprehensive crack length, comprehensive crack width, comprehensive crack position, and comprehensive crack area;

[0062] The fusion feature analysis layer obtains the crack location and recognition result according to the first static and dynamic crack fusion feature; the crack location and recognition result includes the position, length, width, and area of each crack identified for the turbine disk.

[0063] The output layer is used to output the crack location and recognition result.

[0064] In this embodiment, a binocular camera and a binocular high-speed camera are respectively used to obtain a first set of static turbine disk images and a first set of dynamic turbine disk images of the high-pressure turbine disk in a static state and a dynamic state; and a turbine disk crack recognition model is constructed to recognize the first set of static turbine disk images and the first set of dynamic turbine disk images. The crack feature extraction layer of the turbine disk crack recognition model extracts the features of the second set of static turbine disk images and the second set of dynamic turbine disk images through a static crack feature recognition unit and a dynamic crack feature recognition unit respectively, and obtains a first static crack feature and a second dynamic crack feature; by extracting the static crack feature and the dynamic crack feature of the turbine disk image, a good data foundation is laid for later feature fusion, thereby effectively improving the accuracy of crack recognition and location of the high-pressure turbine disk.

[0065] Further, the first static and dynamic crack fusion feature is:

[0066] TFCODS = ο1 * TFSC + ο2 * TFDC;

[0067] Wherein, TFCODS represents the first static and dynamic crack fusion feature; o1 represents the influence weight of the static crack feature; TFSC represents the first static crack feature; o2 represents the influence weight of the dynamic crack feature; TFDC represents the second dynamic crack feature.

[0068] In this embodiment, the crack feature fusion layer of the turbine disk crack recognition model performs feature fusion on the first static crack feature and the second dynamic crack feature to obtain the first static and dynamic crack fusion feature; based on the static image information and dynamic image information of the turbine disk, the comprehensiveness of crack recognition and location of the high-pressure turbine disk can be improved, thereby effectively improving the accuracy of crack recognition and location of the high-pressure turbine disk.

[0069] S3. Based on the crack location and recognition result, the actual crack data of the turbine disk, and a preset error threshold, the initial learning rate of the turbine disk crack recognition model is improved to obtain an improved turbine disk crack recognition model, and the improved turbine disk crack recognition model includes an adaptive learning rate based on the crack recognition error.

[0070] Further, first, obtain the error rate of the crack location recognition of the A high-pressure turbine disk, and perform error rate statistics on the cracks with correct location recognition to obtain the error rates of length, width, and area.

[0071] Further, the adaptive learning rate based on the crack recognition error is as follows:

[0072]

[0073] where ALRCRE represents the adaptive learning rate based on the crack recognition error; LRRE represents the initial learning rate; μ represents the learning rate adjustment factor; λ1 represents the first crack recognition error influence coefficient, reflecting the influence of the turbine disk crack position recognition error probability on the learning rate adjustment; LE represents the turbine disk crack position recognition error probability; λ2 represents the second crack recognition error influence coefficient, representing the influence of the crack length recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EL represents the crack length recognition error probability in the cracks with correct turbine disk crack position recognition; λ3 represents the third crack recognition error influence coefficient, representing the influence of the crack width recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EW represents the crack width recognition error probability in the cracks with correct turbine disk crack position recognition; λ4 represents the fourth crack recognition error influence coefficient, representing the influence of the crack area recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; ES represents the crack area recognition error probability in the cracks with correct turbine disk crack position recognition; and ERE represents the preset error threshold.

[0074] In this embodiment, the initial learning rate of the turbine disk crack recognition model is improved based on the crack location recognition result, the actual crack data of the turbine disk, and the preset error threshold, and the adaptive learning rate based on the crack recognition error is obtained. By adaptively improving the initial learning rate of the turbine disk crack recognition model, the recognition accuracy of the turbine disk crack recognition model can be effectively improved, thereby effectively improving the accuracy of high-pressure turbine disk crack recognition and positioning.

[0075] To verify the effectiveness of a forging crack recognition and positioning method based on binocular vision provided in this embodiment, three different methods are proposed for comparative experiments. For the crack recognition of multiple A high-pressure turbine disks by different methods, the average values of the crack recognition and positioning accuracy rates of multiple A high-pressure turbine disks are compared to obtain the average accuracy rate, including Method 1, Method 2, and Method 3 respectively; Method 1 is a forging crack recognition and positioning method based on binocular vision provided in this embodiment; Method 2 does not consider combining dynamic and static crack features on the basis of Method 1; Method 3 does not consider the adaptive learning rate improvement of the turbine disk crack recognition model; the specific comparison results are shown in Table 1;

[0076] Table 1 Average accuracy rates of crack recognition and positioning of multiple A high-pressure turbine disks by different methods

[0077] Method Average accuracy Method 1 97％ Method 2 90％ Method 3 91

[0078] As can be seen from Table 1, a method for identifying and locating forging cracks based on binocular vision provided in this embodiment shows a certain effectiveness in processing the task of identifying and locating cracks in the A high-pressure turbine disk.

[0079] In this embodiment, a binocular camera and a binocular high-speed camera are used to obtain the first set of static turbine disk images and the first set of dynamic turbine disk images of the turbine disk in a static state and a dynamic state respectively; a turbine disk crack recognition model is constructed to recognize these two sets of images. By extracting the first static crack feature and the second dynamic crack feature, and generating the first static and dynamic crack fusion feature, the recognition result is obtained; furthermore, the initial learning rate of the turbine disk crack recognition model is improved according to the crack recognition error, the actual crack data, and the preset error threshold, and an adaptive learning rate based on the crack recognition error is obtained. This method effectively combines the static and dynamic image information of the turbine disk, and improves the turbine disk crack recognition model by combining the adaptive learning rate based on the crack recognition error, effectively improving the accuracy of turbine disk crack recognition and location.

[0080] Embodiment 2

[0081] In order to improve the accuracy of crack identification and location of the B high-pressure turbine disk, a forging crack identification and location system based on binocular vision is applied;

[0082] Figure 3 The following is a schematic structural diagram of a forging crack identification and location system based on binocular vision provided in an embodiment of the present invention, including:

[0083] A forging crack identification and location system based on binocular vision, including:

[0084] An image acquisition module, configured to obtain the first set of static turbine disk images and the first set of dynamic turbine disk images of the high-pressure turbine disk in a static state and a dynamic state respectively through a binocular camera and a binocular high-speed camera;

[0085] Further, the first set of static turbine disk images includes static turbine disk images; the first set of dynamic turbine disk images includes dynamic turbine disk images; the rotation speed of the turbine disk is 12,000 revolutions per minute; the frame rate of the binocular high-speed camera is 5,000 frames per second;

[0086] The model analysis module is used to construct a turbine disk crack identification model to identify the first set of static turbine disk images and the first set of dynamic turbine disk images. The turbine disk crack identification model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, a fusion feature analysis layer, and an output layer. The image preprocessing layer preprocesses the first set of static turbine disk images and the first set of dynamic turbine disk images respectively to obtain a second set of static turbine disk images and a second set of dynamic turbine disk images. The crack feature extraction layer extracts features from the second set of static turbine disk images and the second set of dynamic turbine disk images respectively through a static crack feature recognition unit and a dynamic crack feature recognition unit to obtain a first static crack feature and a second dynamic crack feature. The crack feature fusion layer fuses the first static crack feature and the second dynamic crack feature to obtain a first static-dynamic crack fusion feature. The fusion feature analysis layer is used to analyze based on the first static-dynamic crack fusion feature to obtain a crack location identification result.

[0087] Further, the turbine disk crack identification model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, and an output layer. Figure 2 It is a schematic structural diagram of a turbine disk crack identification model provided by an embodiment of the present invention.

[0088] The input layer is used to input the first set of static turbine disk images and the first set of dynamic turbine disk images into the turbine disk crack identification model.

[0089] The image preprocessing layer is used to perform image preprocessing on the first set of static turbine disk images and the first set of dynamic turbine disk images to obtain a second set of static turbine disk images and a second set of dynamic turbine disk images. The preprocessing process includes grayscale conversion and denoising of the first set of static turbine disk images and the second set of dynamic turbine disk images.

[0090] The crack feature extraction layer is used to extract features from the second set of static turbine disk images and the second set of dynamic turbine disk images respectively according to the YOLO model to obtain a first static crack feature and a second dynamic crack feature. The first static crack feature includes static crack length, static crack width, static crack position, and static crack area. The second dynamic crack feature includes dynamic crack length, dynamic crack width, dynamic crack position, and dynamic crack area.

[0091] The crack feature fusion layer obtains a first static-dynamic crack fusion feature by performing feature fusion on the first static crack feature and the second dynamic crack feature. The first static-dynamic crack fusion feature includes comprehensive crack length, comprehensive crack width, comprehensive crack position, and comprehensive crack area.

[0092] The fusion feature analysis layer obtains a crack location and recognition result based on the first static and dynamic crack fusion feature; the crack location and recognition result includes the positions, lengths, widths, and areas of the cracks identified for the turbine disk.

[0093] The output layer is used to output the crack location and recognition result.

[0094] In this embodiment, a binocular camera and a binocular high-speed camera are used to obtain a first set of static turbine disk images and a first set of dynamic turbine disk images of the high-pressure turbine disk in a static state and a dynamic state respectively; and a turbine disk crack recognition model is constructed to recognize the first set of static turbine disk images and the first set of dynamic turbine disk images. The crack feature extraction layer of the turbine disk crack recognition model extracts the features of the second set of static turbine disk images and the second set of dynamic turbine disk images through a static crack feature recognition unit and a dynamic crack feature recognition unit respectively, and obtains a first static crack feature and a second dynamic crack feature; by extracting the static crack feature and the dynamic crack feature of the turbine disk image, a good data foundation is laid for subsequent feature fusion, thereby effectively improving the accuracy of crack recognition and location of the high-pressure turbine disk.

[0095] Further, the first static and dynamic crack fusion feature is:

[0096] TFCODS = ο1 * TFSC + ο2 * TFDC;

[0097] Wherein, TFCODS represents the first static and dynamic crack fusion feature; o1 represents the influence weight of the static crack feature; TFSC represents the first static crack feature; o2 represents the influence weight of the dynamic crack feature; TFDC represents the second dynamic crack feature.

[0098] In this embodiment, the crack feature fusion layer of the turbine disk crack recognition model performs feature fusion on the first static crack feature and the second dynamic crack feature to obtain the first static and dynamic crack fusion feature; based on the static image information and dynamic image information of the turbine disk, the comprehensiveness of crack recognition and location of the high-pressure turbine disk can be improved, thereby effectively improving the accuracy of crack recognition and location of the high-pressure turbine disk.

[0099] A model optimization module improves the initial learning rate of the turbine disk crack recognition model based on the crack location and recognition result, the actual crack data of the turbine disk, and a preset error threshold, and obtains an improved turbine disk crack recognition model. The improved turbine disk crack recognition model includes an adaptive learning rate based on the crack recognition error.

[0100] Further, first, the error rate of the crack location recognition of the B high-pressure turbine disk is obtained, and the error rates of the lengths, widths, and areas are statistically obtained for the cracks with correct location recognition.

[0101] Further, the adaptive learning rate based on the crack recognition error is as follows:

[0102]

[0103] where ALRCRE represents the adaptive learning rate based on the crack recognition error; LRRE represents the initial learning rate; μ represents the learning rate adjustment factor; λ1 represents the first crack recognition error influence coefficient, reflecting the influence of the turbine disk crack position recognition error probability on the learning rate adjustment; LE represents the turbine disk crack position recognition error probability; λ2 represents the second crack recognition error influence coefficient, representing the influence of the crack length recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EL represents the crack length recognition error probability in the cracks with correct turbine disk crack position recognition; λ3 represents the third crack recognition error influence coefficient, representing the influence of the crack width recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EW represents the crack width recognition error probability in the cracks with correct turbine disk crack position recognition; λ4 represents the fourth crack recognition error influence coefficient, representing the influence of the crack area recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; ES represents the crack area recognition error probability in the cracks with correct turbine disk crack position recognition; and ERE represents the preset error threshold.

[0104] In this embodiment, the initial learning rate of the turbine disk crack recognition model is improved based on the crack location recognition result, the actual crack data of the turbine disk, and the preset error threshold to obtain the adaptive learning rate based on the crack recognition error. By adaptively improving the initial learning rate of the turbine disk crack recognition model, the recognition accuracy of the turbine disk crack recognition model can be effectively improved, thereby effectively improving the accuracy of crack recognition and location of the high-pressure turbine disk.

[0105] To verify the effectiveness of a forging crack recognition and location system based on binocular vision provided in this embodiment, three different systems are proposed for comparative experiments. For the crack recognition of multiple B high-pressure turbine disks by different systems, the average value of the crack recognition and location accuracy rates of multiple B high-pressure turbine disks is compared to obtain the average accuracy rate. The three different systems include System 1, System 2, and System 3 respectively; System 1 is a forging crack recognition and location system based on binocular vision provided in this embodiment; System 2 does not consider combining dynamic and static crack features on the basis of System 1; System 3 does not consider the adaptive learning rate improvement of the turbine disk crack recognition model; the specific comparison results are shown in Table 2;

[0106] Table 2 Average accuracy rates of different systems for crack recognition and location of multiple B high-pressure turbine disks

[0107] System Average accuracy System 1 96％ System 2 89％ System 3 90％

[0108] As can be seen from Table 2, a forging crack recognition and positioning system based on binocular vision provided by this embodiment shows a certain effectiveness in processing the B high-pressure turbine disk crack recognition and positioning task.

[0109] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying and locating forging cracks based on binocular vision, characterized in that, Including: S1. Obtain a first set of static turbine disk images and a first set of dynamic turbine disk images of the high-pressure turbine disk in a static state and a dynamic state respectively through a binocular camera and a binocular high-speed camera; S2. Construct a turbine disk crack recognition model to recognize the first set of static turbine disk images and the first set of dynamic turbine disk images. The turbine disk crack recognition model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, a fusion feature analysis layer, and an output layer; the image preprocessing layer preprocesses the first set of static turbine disk images and the first set of dynamic turbine disk images respectively to obtain a second set of static turbine disk images and a second set of dynamic turbine disk images; the crack feature extraction layer extracts features from the second set of static turbine disk images and the second set of dynamic turbine disk images respectively through a static crack feature recognition unit and a dynamic crack feature recognition unit to obtain a first static crack feature and a second dynamic crack feature; the crack feature fusion layer fuses the first static crack feature and the second dynamic crack feature to obtain a first static and dynamic crack fusion feature; the fusion feature analysis layer is used to analyze according to the first static and dynamic crack fusion feature to obtain a crack location recognition result; S3. Improve the initial learning rate of the turbine disk crack recognition model based on the crack location recognition result, the actual crack data of the turbine disk, and a preset error threshold to obtain an improved turbine disk crack recognition model, and the improved turbine disk crack recognition model includes an adaptive learning rate based on the crack recognition error.

2. The method for identifying and positioning forging cracks based on binocular vision according to claim 1, wherein: The first set of static turbine disk images includes static turbine disk images; the first set of dynamic turbine disk images includes dynamic turbine disk images.

3. A method for identifying and locating forging cracks based on binocular vision according to claim 1, characterized in that: The turbine disk crack recognition model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, and an output layer; The input layer is used to input the first set of static turbine disk images and the first set of dynamic turbine disk images into the turbine disk crack recognition model; The image preprocessing layer is used to perform image preprocessing on the first set of static turbine disk images and the first set of dynamic turbine disk images to obtain a second set of static turbine disk images and a second set of dynamic turbine disk images; The preprocessing process includes graying and denoising the first set of static turbine disk images and the second set of dynamic turbine disk images; The crack feature extraction layer is used to extract features from the second set of static turbine disk images and the second set of dynamic turbine disk images respectively according to the YOLO model to obtain a first static crack feature and a second dynamic crack feature; the first static crack feature includes static crack length, static crack width, static crack position, and static crack area; the second dynamic crack feature includes dynamic crack length, dynamic crack width, dynamic crack position, and dynamic crack area; The crack feature fusion layer fuses the first static crack feature and the second dynamic crack feature to obtain a first static and dynamic crack fusion feature; the first static and dynamic crack fusion feature includes comprehensive crack length, comprehensive crack width, comprehensive crack position, and comprehensive crack area; The fusion feature analysis layer obtains a crack location and recognition result based on the first static and dynamic crack fusion feature; the crack location and recognition result includes the positions, lengths, widths, and areas of the cracks identified for the turbine disk. The output layer is used to output the crack location and recognition result.

4. The method for identifying and positioning forging cracks based on binocular vision according to claim 3, characterized in that: The first static and dynamic crack fusion feature is: TFCODS = ο1 * TFSC + ο2 * TFDC; where TFCODS represents the first static and dynamic crack fusion feature; o1 represents the influence weight of the static crack feature; TFSC represents the first static crack feature; o2 represents the influence weight of the dynamic crack feature; TFDC represents the second dynamic crack feature.

5. A method for identifying and locating forging cracks based on binocular vision according to claim 1, characterized in that: The adaptive learning rate based on the crack recognition error is: where ALRCRE represents the adaptive learning rate based on the crack recognition error; LRRE represents the initial learning rate; μ represents the learning rate adjustment factor; λ1 represents the first crack recognition error influence coefficient, reflecting the influence of the turbine disk crack position recognition error probability on the learning rate adjustment; LE represents the turbine disk crack position recognition error probability; λ2 represents the second crack recognition error influence coefficient, representing the influence of the crack length recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EL represents the crack length recognition error probability in the cracks with correct turbine disk crack position recognition; λ3 represents the third crack recognition error influence coefficient, representing the influence of the crack width recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EW represents the crack width recognition error probability in the cracks with correct turbine disk crack position recognition; λ4 represents the fourth crack recognition error influence coefficient, representing the influence of the crack area recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; ES represents the crack area recognition error probability in the cracks with correct turbine disk crack position recognition; ERE represents the preset error threshold.

6. A forging crack recognition and positioning system based on binocular vision, characterized in that, It includes: An image acquisition module, which is used to obtain a first turbine disk static image set and a first turbine disk dynamic image set of the high-pressure turbine disk in a static state and a dynamic state respectively through a binocular camera and a binocular high-speed camera. A model analysis module, which is used to construct a turbine disk crack recognition model to recognize the first turbine disk static image set and the first turbine disk dynamic image set. The turbine disk crack recognition model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, a fusion feature analysis layer, and an output layer; the image preprocessing layer preprocesses the first turbine disk static image set and the first turbine disk dynamic image set respectively to obtain a second turbine disk static image set and a second turbine disk dynamic image set; the crack feature extraction layer extracts features from the second turbine disk static image set and the second turbine disk dynamic image set respectively through a static crack feature recognition unit and a dynamic crack feature recognition unit to obtain a first static crack feature and a second dynamic crack feature; the crack feature fusion layer fuses the first static crack feature and the second dynamic crack feature to obtain a first static and dynamic crack fusion feature; the fusion feature analysis layer is used to analyze based on the first static and dynamic crack fusion feature to obtain a crack location and recognition result. The model optimization module improves the initial learning rate of the turbine disk crack recognition model based on the crack location recognition result, the actual crack data of the turbine disk, and a preset error threshold, and obtains an improved turbine disk crack recognition model. The improved turbine disk crack recognition model includes an adaptive learning rate based on the crack recognition error.

7. The forging crack recognition and positioning system based on binocular vision according to claim 6, characterized in that: The first set of static turbine disk images includes static turbine disk images; the first set of dynamic turbine disk images includes dynamic turbine disk images.

8. A forging crack recognition and positioning system based on binocular vision according to claim 6, characterized in that: The turbine disk crack recognition model includes an input layer, an image preprocessing layer, a crack feature extraction layer, a crack feature fusion layer, and an output layer; The input layer is used to input the first set of static turbine disk images and the first set of dynamic turbine disk images into the turbine disk crack recognition model; The image preprocessing layer is used to perform image preprocessing on the first set of static turbine disk images and the first set of dynamic turbine disk images to obtain a second set of static turbine disk images and a second set of dynamic turbine disk images; The preprocessing process includes grayscale conversion and denoising of the first set of static turbine disk images and the second set of dynamic turbine disk images; The crack feature extraction layer is used to extract features from the second set of static turbine disk images and the second set of dynamic turbine disk images respectively according to the YOLO model to obtain a first static crack feature and a second dynamic crack feature; the first static crack feature includes static crack length, static crack width, static crack position, and static crack area; the second dynamic crack feature includes dynamic crack length, dynamic crack width, dynamic crack position, and dynamic crack area; The crack feature fusion layer obtains a first static and dynamic crack fusion feature by fusing the first static crack feature and the second dynamic crack feature; the first static and dynamic crack fusion feature includes comprehensive crack length, comprehensive crack width, comprehensive crack position, and comprehensive crack area; The fusion feature analysis layer obtains a crack location recognition result according to the first static and dynamic crack fusion feature; the crack location recognition result includes the position, length, width, and area of each crack identified for the turbine disk; The output layer is used to output the crack location recognition result.

9. The forging crack recognition and positioning system based on binocular vision according to claim 8, characterized in that: The first static and dynamic crack fusion feature is: TFCODS = ο1 * TFSC + ο2 * TFDC; wherein, TFCODS represents the first static and dynamic crack fusion feature; o1 represents the influence weight of the static crack feature; TFSC represents the first static crack feature; o2 represents the influence weight of the dynamic crack feature; TFDC represents the second dynamic crack feature.

10. The forging crack recognition and positioning system based on binocular vision according to claim 6, characterized in that: The adaptive learning rate based on the crack recognition error is: Among them, ALRCRE represents the adaptive learning rate based on crack recognition error; LRRE represents the initial learning rate; μ represents the learning rate adjustment factor; λ1 represents the first crack recognition error influence coefficient, reflecting the influence of the turbine disk crack position recognition error probability on the learning rate adjustment; LE represents the turbine disk crack position recognition error probability; λ2 represents the second crack recognition error influence coefficient, indicating the influence of the crack length recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EL represents the crack length recognition error probability in the cracks with correct turbine disk crack position recognition; λ3 represents the third crack recognition error influence coefficient, indicating the influence of the crack width recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; EW represents the crack width recognition error probability in the cracks with correct turbine disk crack position recognition; λ4 represents the fourth crack recognition error influence coefficient, indicating the influence of the crack area recognition error in the cracks with correct turbine disk crack position recognition on the learning rate adjustment; ES represents the crack area recognition error probability in the cracks with correct turbine disk crack position recognition; ERE represents the preset error threshold.

Images