Engine crankshaft magnetic powder inspection system and method based on hand-eye system

Abstract

According to the engine crankshaft magnetic powder inspection system and method based on the hand-eye system, manual operation is replaced with automatic equipment, the purposes of saving manpower, improving the detection precision and shortening the detection time are achieved, meanwhile, by means of unique clamping detection design, the stability of a crankshaft in the inspection process is ensured, and the reliability and accuracy of a detection result are improved.

Description

Technical Field

[0001] This article belongs to the field of non-destructive testing technology, and specifically relates to an engine crankshaft magnetic particle inspection system and method based on a hand-eye system, which can be widely used in crankshaft production and quality inspection links in the automotive, shipbuilding, aerospace and other industries. Background Art

[0002] In modern mechanical manufacturing, the crankshaft is a key component of the engine, and its quality directly affects the performance and reliability of the engine. Magnetic particle inspection, as an efficient and sensitive non-destructive testing method, is often used to detect cracks, folds and other defects on the crankshaft surface and near the surface. However, the current traditional crankshaft magnetic particle inspection operation has many problems. On the one hand, the manual operation method relies on a large number of professional inspectors, which has high labor costs, and long-term work can easily lead to visual fatigue, affecting the inspection accuracy. On the other hand, it is difficult to ensure the consistency of shooting angle and clarity when manually shooting with a handheld camera or observing with the naked eye. For crankshafts with complex structures, it is impossible to achieve all-round and high-precision inspection, and there are risks of blind spots and missed inspections. At the same time, manual inspection is inefficient and cannot meet the rapid inspection needs of large-scale production. Therefore, an automated and high-precision crankshaft magnetic particle inspection technology is urgently needed to solve the above problems.

[0003] This paper aims to overcome the shortcomings of existing technologies and provide an engine crankshaft magnetic particle inspection system and method based on a hand-eye system. By replacing manual operations with automated equipment, it achieves the goals of saving manpower, improving inspection accuracy, and shortening inspection time. At the same time, with the help of a unique clamping inspection design, it ensures the stability of the crankshaft during the inspection process and improves the reliability and accuracy of the inspection results. Summary of the Invention

[0004] To solve the above technical problems, this paper provides an engine crankshaft magnetic particle inspection system and method based on a hand-eye system, which aims to improve the accuracy of such inspections in the industry, enhance production efficiency and product quality.

[0005] An engine crankshaft magnetic particle inspection system based on a hand-eye system, comprising: a manipulator system, a camera shooting system, and a data processing and analysis unit;

[0006] The robot system includes a multi-axis high-precision robot, a position feedback device, and a ring light source installed at the end of the robot. The design of integrating the ring light source into the end of the robot ensures that the lighting conditions can stably follow the detection position during complex trajectory movement, providing a consistent and stable lighting environment for subsequent imaging, which is a key prerequisite for ensuring the accuracy of visual inspection.

[0007] The camera system includes an area array camera and an area array lens. The area array camera and area array lens are coaxially arranged below the annular light source. This layout effectively reduces shadow interference caused by the crankshaft surface geometry, provides a uniform UV illumination field, and significantly improves the contrast and clarity of magnetic powder accumulation defects in the image, creating excellent imaging conditions for high-precision image analysis.

[0008] The data processing and analysis unit is connected to the area array camera to ensure that all modules of the system work together and realize automated image analysis and processing;

[0009] The system operation logic is:

[0010] The manipulator carries a ring light source and a camera system and moves along a preset path, controlling the crankshaft's rotation and triggering multi-angle photography. This enables automated scanning of the crankshaft's surface from multiple angles and with full coverage, significantly reducing manual intervention.

[0011] The ring light source emits 365nm ultraviolet light to illuminate the crankshaft surface, automatically stimulating the magnetic powder fluorescence;

[0012] The images captured by the area array camera are input into the data processing and analysis unit for recognition and analysis, completing the core data flow closed loop from image acquisition to intelligent analysis.

[0013] The position feedback device of the manipulator includes:

[0014] The laser displacement sensor detects the position deviation of the end of the robot in real time and provides high-precision real-time feedback of the end position;

[0015] High-precision encoders record the joint angles of the robot in real time, providing accurate real-time feedback of the joint angles;

[0016] The position feedback device dynamically corrects the robot's motion trajectory through a closed-loop control system. It can detect motion deviations in real time and perform dynamic trajectory corrections, greatly improving the robot's motion accuracy, stability, and anti-interference capabilities, ensuring the precise execution of the preset scanning path, and guaranteeing the consistency of image acquisition positions and system reliability.

[0017] The data processing and analysis unit includes: image acquisition module, data processing module, display module and storage library;

[0018] The image acquisition module receives the original image from the area array camera;

[0019] The data processing module performs image preprocessing, defect recognition and result calculation;

[0020] The display module displays defect parameters and judgment conclusions in real time, ensuring the operation process is transparent and controllable;

[0021] The repository records original images, defect data, and inspection reports to achieve comprehensive data archiving. It records original images, structured defect data, and final inspection reports, providing complete support for quality traceability, process analysis, and report generation, ensuring the traceability of the inspection process and data integrity.

[0022] A magnetic particle flaw detection system for an engine crankshaft based on a hand-eye system, the method comprising the following steps:

[0023] S1: Path planning and motion control: Control the manipulator carrying the ring light source and camera shooting system to perform a preset number of lateral movements along the crankshaft axis. Between two adjacent lateral movements, the crankshaft is controlled to complete a preset number of self-rotations and trigger shooting, realizing automated path control and image acquisition triggering;

[0024] S2: UV illumination and image acquisition: A ring light source emits 365nm UV light to illuminate the crankshaft surface, and an area array camera captures the crankshaft surface image at the triggering moment, achieving standard UV illumination and synchronous image capture;

[0025] S3: Data processing and analysis: Receive the original image captured by the area array camera, perform image preprocessing, defect identification and result calculation, and perform core image intelligent analysis;

[0026] S4: Result display and storage: Real-time display of defect parameters and judgment conclusions, and recording of original images, defect data and test reports to complete result presentation and data management;

[0027] These steps form a standardized, clear and executable operational process, covering all aspects of path control, image acquisition, intelligent analysis, and result management, enhancing the practicality and repeatability of the method.

[0028] The path planning and motion control in step S1 are specifically as follows:

[0029] Perform three lateral movements along the crankshaft axis, with each movement distance being 1 / 3 of the crankshaft length, covering the crankshaft length direction;

[0030] Between two adjacent lateral movements, the crankshaft is controlled to complete five 72° rotations and trigger the shooting, covering the circumferential direction of the crankshaft;

[0031] Step S1 defines an optimized, immediately implementable scanning strategy, which theoretically ensures complete grid scanning coverage of the length and circumferential direction of the inspected surface of the engine crankshaft, ensuring comprehensiveness of the inspection while taking into account efficiency.

[0032] The data processing and analysis in step S3 specifically includes:

[0033] Image preprocessing: grayscale conversion, median filtering, Gaussian filtering, and histogram equalization are performed to effectively improve image quality and lay the foundation for subsequent recognition.

[0034] Defect identification: The pre-processed image is input into the channel convolutional neural network model, which outputs the defect type, location, and size parameters. This enables automated, high-precision identification and quantification of defects, significantly surpassing the efficiency, objectivity, and consistency of manual visual interpretation.

[0035] Result determination: The crankshaft's eligibility is determined based on defect parameters, and suggestions for rework, adjustment, or scrapping are generated if the crankshaft fails to meet the requirements. This forms a closed loop from detection to intelligent decision-making, greatly improving the practical value of the system.

[0036] The channel convolutional neural network model includes:

[0037] Multi-scale channel convolution layer: Uses small convolution kernels to extract local detail features and large convolution kernels to extract global features. Multi-scale convolution significantly enhances the model's recognition ability and generalization of defects of different sizes.

[0038] Pooling and sharing of multi-layer perceptron layers:

[0039] Perform the global maximum pooling operation on the feature map of each channel and output the maximum value of each channel Out c =max i,j (X c,i,j ), where c = 1, 2, ... C. This step aggregates information in the spatial dimension, thereby retaining the most significant features of each channel and reducing the redundancy of spatial information;

[0040] Perform global average pooling on the feature map of each channel and output the average value of each channel Finally, a vector of length c is output;

[0041] The feature vectors obtained by global maximum pooling and global average pooling are spliced ​​and then input into a shared multi-layer perceptron. The shared multi-layer perceptron consists of multiple fully connected layers and shares part of the network structure in different task branches. This allows the model to share feature representations between different tasks, improving the generalization ability and training efficiency of the model. By sharing part of the network structure, the shared multi-layer perceptron can share feature representations between different tasks, thereby improving the generalization ability and training efficiency of the model and reducing the repetitiveness of feature learning during the training process. The design of this structure enables the model to handle multi-task learning problems more efficiently and achieve better performance in both classification and regression tasks.

[0042] The training process of the channel convolutional neural network model includes:

[0043] Divide the preprocessed image data into training set, validation set and test set;

[0044] During the training process, the model parameters are adjusted through the back propagation algorithm and the stochastic gradient descent optimization algorithm to minimize the loss function. Where c is the number of categories. In this method, the defect judgment results are either qualified or unqualified. Therefore, i = 1 or i = 2, y i To mark the true probability distribution of qualified or unqualified, y^i is the probability value predicted by the model. The model is evaluated using the validation set. Based on the evaluation results, the number of convolution kernels and convolution kernel sizes in the channel convolution layer, the number of layers and neurons in the shared multi-layer perceptron, etc. in the model are adjusted to prevent model overfitting and ensure that the model has good generalization ability. When the model achieves satisfactory accuracy and recall rate on the test set, training is stopped, and finally a trained defect recognition model is obtained.

[0045] The inspection report generated in step S4 contains the defect type, location coordinates, size parameters, judgment conclusion and original image index path, and is synchronously stored in the repository. The generated report is detailed and clearly structured, meeting industrial quality inspection requirements. The original image index ensures data traceability, and synchronous persistent storage ensures the long-term traceability of inspection results and related evidence, providing complete support for quality analysis, process improvement and report review.

[0046] Beneficial effects:

[0047] Save manpower: This article uses robots and automated equipment to replace manual operations, greatly reducing the dependence on professional inspection personnel. The entire flaw detection process only requires a small number of people to monitor and operate the equipment, reducing labor costs, improving production efficiency, and avoiding inspection errors caused by factors such as manual fatigue and emotions.

[0048] High-precision inspection: The combination of high-resolution cameras and advanced image recognition algorithms enables high-precision inspection of subtle defects on the crankshaft surface. Compared with traditional manual inspection methods, this method can more accurately detect and identify various types of defects, with the minimum detectable defect size reaching the micron level. This effectively improves the accuracy and reliability of inspection and ensures that the crankshaft quality meets high standards.

[0049] Save time: The rapid movement of the robot and the high-speed shooting function of the camera enable the entire flaw detection process to be completed in a short time. Compared with traditional manual inspection methods, the inspection efficiency is increased several times, which greatly shortens the inspection cycle, meets the rapid inspection needs of large-scale production, and helps to improve the company's production efficiency and market competitiveness.

[0050] Advantages of clamping detection: The stable and reliable clamping device ensures the stability of the crankshaft during the flaw detection process, reduces the interference of crankshaft movement or shaking on the detection results, and further improves the accuracy of detection. At the same time, the adjustable clamping design can adapt to crankshafts of different specifications and models, improving the versatility and applicability of the equipment and reducing the company's equipment investment and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The present invention is a schematic diagram of the hardware portion of an engine crankshaft magnetic particle inspection system and method based on a hand-eye system;

[0052] Figure 2 The present invention is a method flow diagram of a magnetic particle flaw detection system and method for an engine crankshaft based on a hand-eye system;

[0053] In the figure, 1. Robot system, 101. Robot, 102. Ring light source, 2. Camera shooting system, 201. Area array camera, 202. Area array lens, 3. Data processing and analysis unit. DETAILED DESCRIPTION

[0054] In order to deepen the understanding of this article, this article will be further described in detail below with reference to examples and drawings. The examples are only used to explain this article and do not constitute a limitation on the scope of protection of this article.

[0055] Manipulator system 1, manipulator 101, annular light source 102, camera shooting system 2, area array camera 201, area array lens 202, data processing and analysis unit 3.

[0056] like Figure 1 、 2 shown

[0057] 1. Robotic arm system 1

[0058] Platform and Function: The multi-degree-of-freedom high-precision manipulator 101 is used as the camera mounting platform, which can flexibly adjust the position, angle and posture of the camera in space;

[0059] Inspection process: After the product is loaded to the designated shooting position, the robot 101 carries a camera to take pictures from multiple angles. After the crankshaft rotates at a certain angle, the camera continues to take pictures until a full 360-degree coverage of the crankshaft is completed. If a defect is detected, the equipment will alarm and manual re-inspection will be required. For standard products, the normal unloading motion program will be executed.

[0060] Path Planning: To achieve comprehensive imaging of all crankshaft components and precise magnetic particle inspection, the system comprehensively considers multiple parameters, including the crankshaft's contour position, the camera's focal length, and the optical brightfield imaging angle. These parameters are then integrated with the robot's 101 motion freedom. This results in a crankshaft magnetic particle inspection path planning method based on multi-source information from the environment, the object being measured, and the robot's 101, accurately planning the optimal optical imaging path.

[0061] Specific motion control: During the process of the camera moving horizontally three times continuously, each time the distance of the movement is one-third of the crankshaft length, the crankshaft product is synchronously controlled to rotate five times, each time rotating 72 degrees, to ensure that the camera fully covers all parts of the crankshaft and obtains complete and clear image data;

[0062] High-precision position feedback and control:

[0063] Feedback device: Equipped with a high-precision position feedback device that integrates advanced sensor technology, including laser displacement sensors and high-precision encoders;

[0064] Real-time monitoring and compensation: Sensors work together to collect real-time motion status data such as the position, speed, and acceleration of the robot 101, and compare it with the preset ideal trajectory in real time. Once a position error is detected, the feedback device quickly feeds the error information back to the control system.

[0065] Closed-loop control: The control system immediately activates the error compensation mechanism, calculates and adjusts the driving torque or voltage signal of each joint according to the error size and direction, corrects the deviation in time, and returns to the ideal trajectory.

[0066] Effect verification: This closed-loop control method ensures the accuracy and repeatability of the shooting position, improves the quality and reliability of image acquisition, and actual tests verify that its repeated positioning accuracy is controlled within an extremely small range, ensuring the stability and consistency of the detection results.

[0067] 2. Camera shooting system 2

[0068] Core equipment: Use high-resolution, high-frame-rate industrial cameras, matched with professional optical lenses, to build an efficient image acquisition system to quickly and clearly capture magnetic powder display images on the crankshaft surface;

[0069] Data transmission: The camera is equipped with a high-speed data transmission interface to achieve real-time transmission of captured images to subsequent data processing systems, improving the consistency and efficiency of the inspection process;

[0070] UV irradiation technology:

[0071] Principle: Adopting 365nm ultraviolet light irradiation technology verified by experiments and theory;

[0072] Function: The ultraviolet light in this wavelength band produces a unique optical interaction with the magnetic powder and the crankshaft surface, which makes the absorption and reflection characteristics of the magnetic powder significantly different from those of the crankshaft matrix, thereby enhancing the contrast between the magnetic powder and the crankshaft surface;

[0073] Effect: The enhanced contrast makes the magnetic particles adsorbed on the defective parts more obvious under ultraviolet light. Even the magnetic particles at the tiny defects can be clearly displayed, creating favorable conditions for the camera to accurately capture the tiny defects.

[0074] 3. Data Processing and Analysis Unit 3

[0075] An image recognition and analysis method is constructed based on a deep learning algorithm. The specific processing steps are as follows:

[0076] Data collection and annotation:

[0077] Collect a large number of crankshaft magnetic particle inspection images with defects of different types and degrees to form an original data set;

[0078] Professional inspectors manually mark defects in images, including their precise locations and specific types, such as cracks, pores, and bruises.

[0079] Follow unified annotation standards and formats, clarify defect geometry representation methods, type definition details, and recording formats, ensure the accuracy and consistency of annotation data, and generate high-quality annotation data sets;

[0080] Data preprocessing:

[0081] Grayscale processing: convert color images into grayscale images to reduce data dimensions and computational complexity;

[0082] Filtering and denoising: Use median filtering to remove salt and pepper noise, and use Gaussian filtering to smooth the image and eliminate random noise;

[0083] Enhance contrast: Enhance image contrast through histogram equalization to make the defect area and background area clearer;

[0084] Model building and training:

[0085] Architecture: This model uses a channel convolutional neural network as its basic architecture to build an image defect recognition model based on channel convolution and pooling. It includes multiple channel convolution layers, pooling layers, global maximum pooling layers, global average pooling layers, and fully connected layers, and also introduces a shared multi-layer perceptron.

[0086] Feature extraction: Each channel convolution layer independently performs convolution operations on each channel of the input feature map, carefully extracting features of different channel dimensions. A multi-scale channel encoding strategy is adopted: small-size kernels focus on local details, while large-size kernels capture global context.

[0087] Pooling and Sharing Multilayer Perceptron:

[0088] Global maximum pooling: Assume that the feature map has c channels, and the size of each channel is H×W, where H is the height of the feature map, W is the width of the feature map, and c is the number of channels. After global maximum pooling, the output is the maximum value of each channel, that is: Out c =max i,j (X c,i,j ), where c = 1, 2, ... C. The purpose of this operation is to aggregate information in the spatial dimension, thereby retaining the most significant features of each channel and reducing the redundancy of spatial information.

[0089] Global average pooling: Global average pooling calculates the average value of the entire feature map of each channel, that is, Finally, a vector of length c is output.

[0090] The feature vectors obtained by the two pooling methods are concatenated and input into a shared multi-layer perceptron. The shared multi-layer perceptron consists of multiple fully connected layers and shares part of the network structure between different task branches, enabling the model to share feature representations between different tasks, improving generalization ability and training efficiency.

[0091] train:

[0092] Data division: After preprocessing, the image data is divided into 80% training set, 10% validation set, and 10% test set;

[0093] Optimization: During the training process, the back propagation algorithm and the stochastic gradient descent optimization algorithm are used to continuously adjust the model parameters to minimize the loss function (formula). Where c is the number of categories. There are two kinds of labeling results in this system, so i=1 or i=2, y i (symbol) is the probability distribution of marking unqualified and qualified, and y^i (symbol) is the probability value predicted by the model;

[0094] Tuning and preventing overfitting: Use the validation set to evaluate the model and adjust parameters, such as the number and size of convolution kernels in the channel convolution layer, the number of layers and neurons in the shared multilayer perceptron, to prevent overfitting and ensure good generalization.

[0095] Termination condition: When the model achieves satisfactory accuracy and recall on the test set, training is stopped to obtain the final defect recognition model;

[0096] Image recognition and analysis:

[0097] Input the real-time image transmitted by the camera into the trained model;

[0098] Input normalization: normalize the image pixel values ​​to the range of (-1,1);

[0099] Feature extraction: Extract defect shape, texture, edge and other features through multi-layer channel convolution;

[0100] Pooling: Use the pooling layer to process the feature maps of each channel, extract the maximum value and average value, reduce noise sensitivity and computational complexity, and enhance representation capabilities;

[0101] Classification and positioning: The extracted features are classified and positioned through a shared multi-layer perceptron layer, and parameter information such as defect type, location, length, and width is output;

[0102] Result output and report generation:

[0103] Conduct a comprehensive comparison and analysis of the defect information output by the model with the preset quality standards;

[0104] Decision logic: Defect parameters within the allowable range are considered qualified; those outside the range are considered unqualified, and treatment suggestions such as rework, adjustment, or scrapping are made based on the defect type, quantity, and severity.

[0105] Report generation: Automatically generate detailed inspection reports containing defect type, location, size, and parameter information, as well as image data from the inspection process, ensuring process traceability and transparency.

[0106] Implementation Example

[0107] 1. Loading method: The crankshaft is fixed on the fixture and carried to the magnetic powder spraying position at a constant speed by the mechanical module.

[0108] 2. Magnetic powder application and magnetization: Apply magnetic powder evenly on the crankshaft surface by spraying or pouring, and then magnetize the crankshaft through a magnetizing device to generate a leakage magnetic field at the defective part to absorb the magnetic powder, forming an obvious magnetic trace display.

[0109] 3. Shooting and image acquisition: The robot 101 carries the camera and moves around the crankshaft to shoot according to the preset motion program. During the shooting process, the camera captures multi-angle and all-round images of the magnetic powder display area on the crankshaft surface, and transmits the collected image data to the data processing and analysis system in real time.

[0110] 4. Data analysis and result determination: The data processing and analysis system processes and analyzes the received image data, identifies and measures the defect information on the crankshaft surface, automatically determines whether the crankshaft is qualified according to the quality standards, and generates a detailed inspection report.

[0111] 5. Unloading and subsequent processing: After the inspection is completed, the clamping device releases the crankshaft, and the robot 101 moves the crankshaft to the unloading area. The crankshaft is classified and processed according to the inspection results. Qualified products enter the next production link, and unqualified products are returned for repair or scrapped.

[0112] The above description is only a preferred embodiment of this article and is not intended to limit this article. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of this article should be included in the scope of protection of this article.

Claims

1. A magnetic particle inspection system for engine crankshafts based on a hand-eye system, characterized in that: include: Robotic arm system, camera shooting system and data processing and analysis unit; The manipulator system includes a multi-axis high-precision manipulator, a position feedback device and a ring light source installed at the end of the manipulator; The camera shooting system includes an area array camera and an area array lens, and the area array camera and the area array lens are coaxially arranged below the annular light source; The data processing and analysis unit is in communication with the area array camera; The system operation logic is: The manipulator carries the annular light source and the camera shooting system and moves along a preset path, controls the crankshaft to rotate and triggers multi-angle shooting; The annular light source emits 365nm ultraviolet light to irradiate the surface of the crankshaft; The images collected by the area array camera are input into the data processing and analysis unit for recognition and analysis.

2. The engine crankshaft magnetic particle inspection system based on the hand-eye system according to claim 1 is characterized in that: The position feedback device of the manipulator comprises: The laser displacement sensor detects the position deviation of the end of the robot in real time; High-precision encoder records the robot joint angle in real time; The position feedback device dynamically corrects the motion trajectory of the manipulator through the control system.

3. The engine crankshaft magnetic particle inspection system based on the hand-eye system according to claim 1 or 2, characterized in that: The data processing and analysis unit includes: an image acquisition module, a data processing module, a display module and a storage library; The image acquisition module receives the original image from the area array camera; The data processing module performs image preprocessing, defect recognition and result calculation; The display module displays the defect parameters and judgment conclusions in real time; The storage library records original images, defect data and inspection reports.

4. A magnetic particle flaw detection method for an engine crankshaft based on a hand-eye system, characterized in that: The method comprises the following steps: S1: Path planning and motion control: Control the manipulator carrying the ring light source and the camera shooting system to perform a preset number of lateral movements along the crankshaft axis. Between two adjacent lateral movements, control the crankshaft to complete a preset number of self-rotations and trigger shooting; S2: UV illumination and image acquisition: 365 nm UV light is emitted by the annular light source to illuminate the crankshaft surface, and the area array camera is used to capture an image of the crankshaft surface at the triggering moment; S3: Data processing and analysis: receiving the original image captured by the area array camera, performing image preprocessing, defect recognition and result calculation; S4: Result display and storage: Real-time display of defect parameters and judgment conclusions, and recording of original images, defect data and inspection reports.

5. The method for magnetic particle flaw detection of an engine crankshaft based on a hand-eye system according to claim 4, characterized in that: The path planning and motion control in step S1 are specifically as follows: Perform three lateral movements along the crankshaft axis, with each movement distance being 1 / 3 of the crankshaft length; Between two adjacent lateral movements, the crankshaft is controlled to complete five 72° rotations and trigger shooting.

6. The method for magnetic particle flaw detection of an engine crankshaft based on a hand-eye system according to claim 4 or 5, characterized in that: The data processing and analysis in step S3 specifically includes: Image preprocessing: performing grayscale, median filtering, Gaussian filtering, and histogram equalization operations; Defect recognition: The preprocessed image is input into the channel convolutional neural network model, which outputs the defect type, location and size parameters; Result determination: The crankshaft's eligibility is determined based on the defect parameters, and suggestions for rework, adjustment, or scrapping are generated if the crankshaft fails to meet the requirements.

7. The method for magnetic particle flaw detection of an engine crankshaft based on a hand-eye system according to claim 6, characterized in that: The channel convolutional neural network model includes: Multi-scale channel convolution layer: Use small convolution kernels to extract local detail features and large convolution kernels to extract global features; Pooling and shared multi-layer perceptron layer: Perform global maximum pooling operation on the feature map of each channel and output the maximum value of each channel Out c =max i,j (X c,i,j ), where c = 1, 2, ... C; and perform global average pooling on the feature map of each channel, outputting the average value of each channel Finally, a vector of length c is output.

8. The method for magnetic particle flaw detection of an engine crankshaft based on a hand-eye system according to claim 7, characterized in that: The training process of the channel convolutional neural network model includes: Divide the preprocessed image data into training set, validation set and test set; During the training process, the model parameters are adjusted through the back propagation algorithm and the stochastic gradient descent optimization algorithm to minimize the loss function. Where c is the number of categories. In this method, the defect judgment results are either qualified or unqualified. Therefore, i = 1 or i = 2, y i is the true probability distribution of passing or failing, and y^i is the probability value predicted by the model.

9. A magnetic particle flaw detection method for an engine crankshaft based on a hand-eye system according to any one of claims 4 to 8, characterized in that: The inspection report generated in step S4 includes the defect type, location coordinates, size parameters, judgment conclusion and original image index path, and is synchronously stored in the storage library.

Images