A detection device that uses AI visual recognition technology to quickly measure strip deviation

Abstract

The present invention discloses a detection device for quickly measuring strip deviation using AI visual recognition technology, relating to the field of strip processing technology, comprising a device bracket, a linear array camera and a visual recognition system, the visual recognition system comprising an image acquisition unit, an image analysis unit, an environmental compensation unit, a deviation prediction unit, and a result output unit; the present invention collects real-time image data of the edge area of the strip through a linear array camera in real time, analyzes the real-time image data through a deep learning model, outputs pixel-level coordinates of the strip edge and image abnormality feature points, and then obtains distance data and inclination angle data, fuses and corrects the data with the output of the Kalman filter and the AI processing module, obtains corrected image data, establishes a strip deviation feature model based on a convolutional neural network, inputs the corrected image data into the strip deviation feature model, extracts the strip deviation feature, and calculates the strip deviation probability and sends it to the result output unit.

Description

Technical Field

[0001] The present invention relates to the technical field of strip steel processing, and in particular to a detection device for rapidly measuring strip steel deviation using AI visual recognition technology. Background Art

[0002] Strip steel is a thin, narrow steel plate in long lengths, usually supplied in coils. Compared to plate steel, strip steel offers advantages such as high dimensional accuracy, good surface quality, and ease of use. Strip steel is widely used in the production of welded steel pipes, cold-formed steel billets, and in the manufacture of products such as bicycle frames, wheel rims, clamps, washers, springs, cable armor, saw blades, razor blades, and baling iron sheets. Strip steel production is a complex process involving multiple steps, including raw material processing, smelting, rolling, quenching, tempering, and annealing.

[0003] During the hot rolling production process, it is very easy for the center of the strip to deviate from the center of the roller, which is often called deviation in industrial production. Many scholars at home and abroad have conducted relevant research on the factors that cause deviation and the methods of correction, such as: strip shape defects, the influence of roller geometry, axial movement of the conveyor roller, and process parameters of the unit. Most hot rolling production lines are equipped with a side guide device in front of the coiler to clamp the strip and align it with the coiler's pinch rollers for correction, but it can only correct the strip entering the side guide part, and cannot completely avoid the deviation phenomenon. When the hot rolled strip before entering the coiler deviates, it will mainly cause surface defects of the strip and hole defects formed on the surface of the fallen strip;

[0004] Traditional detection methods rely on photoelectric sensors or manual inspections, which suffer from issues such as insufficient accuracy, slow response, and susceptibility to environmental interference (such as dust and dim light). Traditional machine vision methods (such as edge detection and object detection) lack robustness in complex working conditions and are prone to false alarms or missed detections.

[0005] In view of the above technical defects, a solution is now proposed. Summary of the Invention

[0006] The purpose of the present invention is to establish a strip deviation feature model based on a convolutional neural network, input the corrected image data into the strip deviation feature model, extract the strip deviation features, and calculate the strip deviation probability. During the strip conveying process, the strip image is acquired by a linear array camera, and the strip deviation prediction is realized through high-precision image processing and visual recognition. It is possible to intervene in time before the strip deviates, thereby fundamentally solving the production losses caused by strip deviation.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a detection device for rapidly measuring strip deviation using AI visual recognition technology, comprising a device bracket, a linear array camera, and a visual recognition system, wherein the device bracket is fixed to the top surface of a frame, an angle adjustment mechanism is installed on the outer surface of the device bracket, and two linear array cameras are respectively connected to the outer surfaces of the angle adjustment mechanism;

[0008] The visual recognition system includes an image acquisition unit, an image analysis unit, an environment compensation unit, a deviation prediction unit, and a result output unit;

[0009] The image acquisition unit uses a linear array camera to collect real-time image data of the strip edge area. Its baseline distance is dynamically matched with the strip width, and high-frequency pulse synchronization control is used to achieve image temporal and spatial alignment.

[0010] The image analysis unit includes an AI processing module and an image analysis module. The AI processing module is used to integrate the edge computing unit and has a built-in deep learning model. The deep learning model is implemented based on the fusion architecture of the residual network and the attention mechanism. The input is the real-time image data output by the image acquisition unit, and the output is the pixel-level coordinates of the strip edge and the image abnormality feature points, which are sent to the image analysis module.

[0011] The image analysis module is used to obtain a preset feature point number threshold, and if the abnormal feature points of the image are greater than or equal to the feature point number threshold, generate an image correction signal and send the image correction signal to the environment compensation unit;

[0012] The environmental compensation unit includes a laser ranging unit and an image correction module. The laser ranging unit is used to monitor the distance between the strip and the linear array camera and the inclination of the strip conveyor roller in real time, and integrate the distance data and inclination data into a data set and send it to the image correction module.

[0013] The image correction module is used to obtain and process the image correction signal, fuse and correct the data set with the output of the AI processing module through the Kalman filter to eliminate the influence of strip vibration and mechanical installation errors, and obtain the corrected image data and send it to the deviation prediction unit;

[0014] The deviation prediction unit is used to obtain historical production data of hot-rolled strip and perform preprocessing, construct a hot-rolled strip image dataset as training samples, establish a strip deviation feature model based on a convolutional neural network, input the corrected image data into the strip deviation feature model, extract the strip deviation features, and calculate the strip deviation probability and send it to the result output unit.

[0015] Furthermore, the device bracket includes a self-aligning ball bearing and a vertical support shaft. The two vertical support shafts are symmetrically arranged on the top surface of the frame adjacent to the conveying path along the width direction of the strip steel. The two self-aligning ball bearings are respectively connected to the top surface of the vertical support shafts. A horizontal support shaft is commonly installed between the two self-aligning ball bearings, and an angle adjustment mechanism is installed on the outer surface of the horizontal support shaft.

[0016] Furthermore, the angle adjustment mechanism includes a fixing seat and a polarization filter assembly, the fixing seat is fixed on the end surface of the horizontal support shaft, two driving motors are fixed inside the fixing seat, and the two linear array cameras are respectively fixed on the outer surface of the output end of the driving motor.

[0017] Furthermore, the polarization filter assembly includes a pressure sensor and an elastic sensing block, and the polarization angle is dynamically adjusted through an electric rotation mechanism to eliminate the reflective interference on the surface of the strip steel. The outer surface of the driving motor is provided with a gear ring, and the outer surface of the fixed seat is provided with a sliding groove. The inner wall of the sliding groove is movably connected with a movable rack, and the movable rack is engaged with the gear ring. The pressure sensor is fixed on the end surface of the movable rack, and the elastic sensing block is fixed on the outer surface of the fixed seat. The pressure sensor and the elastic sensing block are in contact with each other.

[0018] Furthermore, the image acquisition unit further includes a high-frequency pulse LED array and a light source controller;

[0019] The light emission frequency of the high-frequency pulse LED array is synchronized with the line frequency of the linear array camera;

[0020] The light source controller has a built-in closed-loop feedback circuit, which adjusts the output power in real time according to the ambient light intensity to ensure that the image signal-to-noise ratio is less than or equal to 40dB.

[0021] Furthermore, the specific process of outputting the pixel-level coordinates of the strip edge and the image abnormality feature points is as follows:

[0022] S101. Build a deep learning model based on the improved YOLOv5 architecture and embed a deformable convolutional layer in the backbone network to adapt to the geometric deformation of the strip edge;

[0023] S102: Introduce a direction classification branch in the detection head part and synchronously output the deviation direction determination result;

[0024] S103. The loss function is designed as a weighted combination of the deviation regression loss and the direction classification loss, and the weight coefficient is automatically determined by Bayesian optimization;

[0025] S104. Synthetic data augmentation technology is used during the training of the deep learning model, including simulation scenarios of dynamic lighting, dust occlusion, and changes in strip surface texture;

[0026] S105. Input the real-time image data into the deep learning model, extract the edge feature points of the strip steel, construct the three-dimensional point cloud data, obtain the pixel-level coordinates of the edge of the strip steel, and perform feature comparison based on the standard strip steel image to obtain the image abnormal feature points.

[0027] Furthermore, the specific process of outputting the strip deviation probability is as follows:

[0028] S201. Acquire historical production data of hot-rolled strip steel, wherein the historical production data includes standard images of the strip steel, a strip steel conveying path, and a straight line of the strip steel edge; integrate historical operation data of a road roller as training samples; and split the training samples into a training set and a test set in a ratio of 8:2;

[0029] S202, constructing a strip steel bias feature model based on a convolutional neural network, downloading a weight file and loading it onto the corresponding network to initialize migration network parameters;

[0030] S203. Modify the last fully connected layer of the network, keep the input unchanged, set the output to the strip edge straight line, initialize the weight of the last layer, use the gradient descent algorithm for learning, and use fixed step size decay to optimize the training parameters, retrain the entire network, and obtain the strip bias feature model;

[0031] S204, during the training process, small batches of training samples are randomly and non-repeatedly extracted from the training set for training. The extraction of all training samples constitutes one training cycle. The training is completed after a certain number of iterations, and a strip bias feature model is obtained;

[0032] S205. Input the corrected image data into the strip deviation feature model to output the strip edge straight line, extract the potential roller using the Hough transform line detection algorithm, and then perform deviation warning and deviation amount prediction based on the position coordinates of the strip edge straight line screened out based on the position between the strip standard image and the roller, to obtain the strip deflection probability.

[0033] Furthermore, the result output unit is used to obtain a preset strip deviation judgment interval and make result judgment based on the strip deviation probability. If the strip deviation probability exceeds the limited range of the strip deviation judgment interval, a strip deviation signal is generated and sent to the industrial control system for visual display and voice reminder.

[0034] Furthermore, it also includes a control interface unit for providing a standardized communication protocol interface, outputting the deviation detection results to the industrial control system at a refresh rate greater than or equal to 1kHz, supporting the PID control parameter adaptive adjustment function, and dynamically adjusting the response threshold of the correction actuator according to the deviation amount.

[0035] Furthermore, the environmental compensation unit further includes a temperature compensation module and a vibration compensation module, wherein:

[0036] The temperature compensation module uses a thermistor array to monitor the temperature field distribution of the camera installation area in real time, and corrects the thermal expansion and contraction errors of the optical components through a lookup table method;

[0037] The vibration compensation module uses an accelerometer to collect the vibration spectrum of the mechanical structure and suppresses interference signals in the resonance frequency band through a notch filter.

[0038] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0039] This detection device uses AI visual recognition technology to quickly measure strip deviation. It collects real-time image data of the strip edge area through a line array camera, analyzes the real-time image data through a deep learning model, outputs the pixel-level coordinates of the strip edge and image abnormality feature points, and then obtains distance data and inclination data. It fuses and corrects the output of the Kalman filter and the AI processing module to obtain corrected image data, establishes a strip deviation feature model based on a convolutional neural network, inputs the corrected image data into the strip deviation feature model, extracts the strip deviation feature, and calculates the strip deviation probability and sends it to the result output unit. During the strip conveying process, the strip image is obtained through the line array camera, and the strip deviation prediction is realized through high-precision image processing and visual recognition. It can intervene in time before the strip deviates, and fundamentally solve the production losses caused by strip deviation. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Shows a schematic diagram of the overall external structure of the present invention;

[0041] Figure 2 Another schematic diagram of the overall external structure of the present invention is shown;

[0042] Figure 3 Shows a schematic diagram of the structure of the visual recognition system of the present invention;

[0043] Legend: 1. Linear array camera; 2. Vertical support shaft; 3. Self-aligning ball bearing; 4. Horizontal support shaft; 5. Fixed seat; 6. Drive motor; 7. Gear ring; 8. Sliding slot; 9. Moving rack; 10. Pressure sensor; 11. Elastic sensing block. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] Example 1:

[0046] like Figure 1-Figure 2 As shown, a detection device for quickly measuring strip deviation using AI visual recognition technology includes a device bracket, a linear array camera 1, and a visual recognition system. The device bracket is fixed to the top surface of the frame, and an angle adjustment mechanism is installed on the outer surface of the device bracket. Two linear array cameras 1 are respectively connected to the outer surfaces of the angle adjustment mechanism.

[0047] The device bracket includes a self-aligning ball bearing 3 and a vertical support shaft 2. The two vertical support shafts are symmetrically arranged on the top surface of the frame adjacent to the conveying path along the width direction of the strip. The two self-aligning ball bearings 3 are respectively connected to the top surface of the vertical support shaft 2. A horizontal support shaft 4 is installed between the two self-aligning ball bearings 3, and an angle adjustment mechanism is installed on the outer surface of the horizontal support shaft 4.

[0048] The angle adjustment mechanism includes a fixing seat 5 and a polarization filter assembly. The fixing seat 5 is fixed to the end surface of the horizontal support shaft 4. Two drive motors 6 are fixed inside the fixing seat 5. The two line array cameras 1 are respectively fixed to the outer surface of the output end of the drive motor 6.

[0049] The polarization filter assembly includes a pressure sensor 10 and an elastic sensing block 11. The polarization angle is dynamically adjusted through an electric rotation mechanism to eliminate the reflective interference on the strip surface. The outer surface of the drive motor 6 is sleeved with a gear ring 7, and the outer surface of the fixed seat 5 is provided with a sliding groove 8. The inner wall of the sliding groove 8 is movably connected to a moving rack 9. The moving rack 9 is engaged with the gear ring 7. The pressure sensor 10 is fixed to the end surface of the moving rack 9, and the elastic sensing block 11 is fixed to the outer surface of the fixed seat 5. The pressure sensor 10 and the elastic sensing block 11 are in contact with each other.

[0050] Working principle: The line array camera 1 is vertically fixed on the frame of the strip transport path through the vertical support shaft 2. The real-time image data of the edge area of the strip is collected by the line array camera 1. During the detection process, the self-aligning ball bearing 3 keeps the line array camera 1 in a horizontal state. The drive motor 6 drives the line array camera 1 to rotate and adjust the shooting angle. During the rotation, the gear ring 7 drives the movable rack 9 to move along the sliding slot 8, and then presses the elastic sensing block 11. The pressure change is detected by the pressure sensor 10, and then the adjustment angle of the line array camera 1 is reflected, which is convenient for precise adjustment.

[0051] Example 2:

[0052] like Figure 3 As shown, a detection device for quickly measuring strip deviation using AI visual recognition technology includes a device bracket, a linear array camera 1 and a visual recognition system. The visual recognition system includes an image acquisition unit, an image analysis unit, an environmental compensation unit, a deviation prediction unit, and a result output unit.

[0053] The image acquisition unit collects real-time image data of the strip edge area through the linear array camera 1. Its baseline distance is dynamically matched with the strip width, and the image is aligned in time and space through high-frequency pulse synchronization control.

[0054] The image acquisition unit also includes a high-frequency pulse LED array and a light source controller;

[0055] The light emission frequency of the high-frequency pulse LED array is synchronized with the line frequency of the linear array camera 1;

[0056] The light source controller has a built-in closed-loop feedback loop, which adjusts the output power in real time according to the ambient light intensity to ensure that the image signal-to-noise ratio is less than or equal to 40dB.

[0057] The image analysis unit includes an AI processing module and an image analysis module. The AI processing module is used to integrate the edge computing unit and has a built-in deep learning model. The deep learning model is implemented by integrating the residual network with the attention mechanism. The input is the real-time image data output by the image acquisition unit. The output is the pixel-level coordinates of the strip edge and the image abnormality feature points and sends them to the image analysis module.

[0058] The specific process of outputting the pixel-level coordinates of the strip edge and the image abnormality feature points is as follows:

[0059] S101. Build a deep learning model based on the improved YOLOv5 architecture and embed a deformable convolutional layer in the backbone network to adapt to the geometric deformation of the strip edge;

[0060] S102: Introduce a direction classification branch in the detection head part and synchronously output the deviation direction determination result;

[0061] S103. The loss function is designed as a weighted combination of the deviation regression loss and the direction classification loss, and the weight coefficient is automatically determined by Bayesian optimization;

[0062] S104. Synthetic data augmentation technology is used during deep learning model training, including simulation scenarios of dynamic lighting, dust occlusion, and changes in strip surface texture;

[0063] S105. Input the real-time image data into the deep learning model, extract the edge feature points of the strip steel, construct the three-dimensional point cloud data, obtain the pixel-level coordinates of the edge of the strip steel, and perform feature comparison based on the standard strip steel image to obtain the image abnormal feature points.

[0064] The image analysis module is used to obtain a preset feature point number threshold. If the number of abnormal feature points in the image is greater than or equal to the feature point number threshold, an image correction signal is generated and sent to the environmental compensation unit.

[0065] The environmental compensation unit includes a laser ranging unit and an image correction module. The laser ranging unit is used to monitor the distance between the strip and the linear array camera and the inclination of the strip conveyor roller in real time. The distance data and inclination data are integrated into a data set and sent to the image correction module.

[0066] The environmental compensation unit also includes a temperature compensation module and a vibration compensation module, wherein:

[0067] The temperature compensation module uses a thermistor array to monitor the temperature distribution of the camera installation area in real time, and corrects the thermal expansion and contraction errors of the optical components through a lookup table method;

[0068] The vibration compensation module uses an accelerometer to collect the vibration spectrum of the mechanical structure and suppresses the interference signal in the resonant frequency band through a notch filter.

[0069] The image correction module is used to obtain and process image correction signals. It fuses and corrects the data set with the output of the AI processing module through a Kalman filter to eliminate the influence of strip vibration and mechanical installation errors. The corrected image data is then sent to the deviation prediction unit.

[0070] The deviation prediction unit is used to obtain and preprocess historical production data of hot-rolled strip steel, construct a hot-rolled strip steel image dataset as training samples, establish a strip steel deviation feature model based on a convolutional neural network, input the corrected image data into the strip steel deviation feature model, extract the strip steel deviation features, and calculate the strip steel deviation probability and send it to the result output unit;

[0071] The specific process of outputting strip deviation probability is as follows:

[0072] S201. Acquire historical production data of hot-rolled strip steel, including standard images of the strip steel, the strip steel conveying path, and the straight lines of the strip steel edges. Integrate historical operation data of a road roller as training samples, and split the training samples into a training set and a test set in a ratio of 8:2.

[0073] S202, constructing a strip steel bias feature model based on a convolutional neural network, downloading a weight file and loading it onto the corresponding network to initialize migration network parameters;

[0074] S203. Modify the last fully connected layer of the network, keep the input unchanged, set the output to the strip edge straight line, initialize the weight of the last layer, use the gradient descent algorithm for learning, and use fixed step size decay to optimize the training parameters, retrain the entire network, and obtain the strip bias feature model;

[0075] S204, during the training process, small batches of training samples are randomly and non-repeatedly extracted from the training set for training. The extraction of all training samples constitutes one training cycle. The training is completed after a certain number of iterations, and a strip bias feature model is obtained;

[0076] S205. Input the corrected image data into the strip deviation feature model to output the strip edge straight line, extract the potential roller using the Hough transform line detection algorithm, and then perform deviation warning and deviation amount prediction based on the position coordinates of the strip edge straight line screened out based on the position between the strip standard image and the roller, to obtain the strip deflection probability.

[0077] The result output unit is used to obtain the preset strip deviation judgment interval and make result judgments based on the strip deviation probability. If the strip deviation probability exceeds the limited range of the strip deviation judgment interval, a strip deviation signal is generated and sent to the industrial control system for visual display and voice reminder.

[0078] It also includes a control interface unit for providing a standardized communication protocol interface, outputting the deviation detection results to the industrial control system at a refresh rate greater than or equal to 1kHz, supporting the adaptive adjustment function of PID control parameters, and dynamically adjusting the response threshold of the correction actuator according to the deviation amount.

[0079] The present invention collects real-time image data of the edge area of the strip through a line array camera 1, analyzes the real-time image data through a deep learning model, outputs the pixel-level coordinates of the strip edge and the image abnormality feature points, and then obtains distance data and inclination data. The data is fused and corrected with the output of the Kalman filter and the AI processing module to obtain corrected image data, establishes a strip deviation feature model based on a convolutional neural network, inputs the corrected image data into the strip deviation feature model, extracts the strip deviation feature, and calculates the strip deviation probability and sends it to the result output unit. During the strip conveying process, the strip image is acquired by the line array camera, and the strip deviation prediction is realized through high-precision image processing and visual recognition. It can intervene in time before the strip deviates, and fundamentally solve the production loss caused by the strip deviation.

[0080] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by technical personnel in this field for each set of sample data; as long as it does not affect the proportional relationship between the parameter and the quantized value.

[0081] In the two embodiments provided in this application, it should be understood that the disclosed devices and systems can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the module division is merely a logical function division. In actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not implemented. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interface, and the indirect coupling or communication connection of devices or modules may be electrical, mechanical or other forms.

[0082] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A detection device for rapidly measuring strip deviation using AI visual recognition technology, comprising a device bracket, a linear array camera (1) and a visual recognition system, characterized in that: The device bracket is fixed on the top surface of the frame, an angle adjustment mechanism is installed on the outer surface of the device bracket, and the two line array cameras (1) are respectively connected to the outer surface of the angle adjustment mechanism; The visual recognition system includes an image acquisition unit, an image analysis unit, an environment compensation unit, a deviation prediction unit, and a result output unit; The image acquisition unit collects real-time image data of the edge area of the strip through a linear array camera (1); The image analysis unit includes an AI processing module and an image analysis module. The AI processing module is used to integrate the edge computing unit and has a built-in deep learning model. The deep learning model is implemented based on the fusion architecture of the residual network and the attention mechanism. The input is the real-time image data output by the image acquisition unit, and the output is the pixel-level coordinates of the strip edge and the image abnormality feature points, which are sent to the image analysis module. The specific process of outputting the pixel-level coordinates of the strip edge and the image abnormality feature points is as follows: S101. Build a deep learning model based on the improved YOLOv5 architecture and embed a deformable convolutional layer in the backbone network to adapt to the geometric deformation of the strip edge; S102: Introduce a direction classification branch in the detection head part and synchronously output the deviation direction determination result; S103. The loss function is designed as a weighted combination of the deviation regression loss and the direction classification loss, and the weight coefficient is automatically determined by Bayesian optimization; S104. Synthetic data augmentation technology is used during the training of the deep learning model, including simulation scenarios of dynamic lighting, dust occlusion, and changes in strip surface texture; S105. Input the real-time image data into the deep learning model, extract the edge feature points of the steel strip, construct the three-dimensional point cloud data, obtain the pixel-level coordinates of the steel strip edge, and perform feature comparison based on the standard steel strip image to obtain the image abnormal feature points; The image analysis module is used to obtain a preset feature point number threshold, and if the abnormal feature points of the image are greater than or equal to the feature point number threshold, generate an image correction signal and send the image correction signal to the environment compensation unit; The environmental compensation unit includes a laser distance measuring unit and an image correction module. The laser distance measuring unit is used to monitor the distance between the strip and the linear array camera (1) and the inclination of the strip conveyor roller in real time, and obtains distance data and inclination data, integrates them into a data set, and sends them to the image correction module. The image correction module is used to obtain and process the image correction signal, fuse and correct the data set with the output of the AI processing module through the Kalman filter, and send the corrected image data to the deviation prediction unit; The deviation prediction unit is used to obtain and preprocess historical production data of hot-rolled strip steel, construct a hot-rolled strip steel image dataset as training samples, establish a strip steel deviation feature model based on a convolutional neural network, input the corrected image data into the strip steel deviation feature model, extract the strip steel deviation features, and calculate the strip steel deviation probability and send it to the result output unit; The specific process of outputting strip deviation probability is as follows: S201. Acquire historical production data of hot-rolled strip steel, wherein the historical production data includes standard images of the strip steel, a strip steel conveying path, and a straight line of the strip steel edge; integrate historical operation data of a road roller as training samples; and split the training samples into a training set and a test set in a ratio of 8:2; S202, constructing a strip steel bias feature model based on a convolutional neural network, downloading a weight file and loading it onto the corresponding network to initialize migration network parameters; S203. Modify the last fully connected layer of the network, keep the input unchanged, set the output to the strip edge straight line, initialize the weight of the last layer, use the gradient descent algorithm for learning, and use fixed step size decay to optimize the training parameters, retrain the entire network, and obtain the strip bias feature model; S204, during the training process, small batches of training samples are randomly and non-repeatedly extracted from the training set for training. The extraction of all training samples constitutes one training cycle. The training is completed after a certain number of iterations, and a strip bias feature model is obtained; S205. Input the corrected image data into the strip deviation feature model to output the strip edge straight line, extract the potential roller using the Hough transform line detection algorithm, and then perform deviation warning and deviation amount prediction based on the position coordinates of the strip edge straight line screened out based on the position between the strip standard image and the roller, to obtain the strip deflection probability.

2. The detection device for rapidly measuring strip deviation using AI visual recognition technology according to claim 1 is characterized in that: The device bracket includes a self-aligning ball bearing (3) and a vertical support shaft (2), the two vertical support shafts (2) are symmetrically arranged on the top surface of the frame adjacent to the conveying path along the width direction of the strip steel, the two self-aligning ball bearings (3) are respectively connected to the top surface of the vertical support shaft (2), a horizontal support shaft (4) is installed between the two self-aligning ball bearings (3), and an angle adjustment mechanism is installed on the outer surface of the horizontal support shaft (4).

3. The detection device for rapidly measuring strip deviation using AI visual recognition technology according to claim 1 is characterized in that: The angle adjustment mechanism comprises a fixing seat (5) and a polarization filter assembly, wherein the fixing seat (5) is fixed to the end surface of the horizontal support shaft (4), two drive motors (6) are fixed inside the fixing seat (5), and the two linear array cameras (1) are respectively fixed to the outer surfaces of the output ends of the drive motors (6).

4. The detection device for rapidly measuring strip deviation using AI visual recognition technology according to claim 3 is characterized in that: The polarization filter assembly includes a pressure sensor (10) and an elastic sensing block (11); the outer surface of the driving motor (6) is sleeved with a gear ring (7); the outer surface of the fixed seat (5) is provided with a sliding long groove (8); the inner wall of the sliding long groove (8) is movably connected to a moving rack (9); the moving rack (9) and the gear ring (7) are meshed with each other; the pressure sensor (10) is fixed to the end surface of the moving rack (9); the elastic sensing block (11) is fixed to the outer surface of the fixed seat (5); and the pressure sensor (10) and the elastic sensing block (11) are in contact with each other.

5. The detection device for rapidly measuring strip deviation using AI visual recognition technology according to claim 1 is characterized in that: The image acquisition unit also includes a high-frequency pulse LED array and a light source controller; The light emission frequency of the high-frequency pulse LED array is synchronized with the line frequency of the linear array camera (1); The light source controller has a built-in closed-loop feedback circuit, which adjusts the output power in real time according to the ambient light intensity to ensure that the image signal-to-noise ratio is less than or equal to 40dB.

6. The detection device for rapidly measuring strip deviation using AI visual recognition technology according to claim 1 is characterized in that: The result output unit is used to obtain a preset strip deviation judgment interval and make a result judgment based on the strip deviation probability. If the strip deviation probability exceeds the limited range of the strip deviation judgment interval, a strip deviation signal is generated and sent to the industrial control system for visual display and voice reminder.

7. The detection device for rapidly measuring strip deviation using AI visual recognition technology according to claim 1 is characterized in that: It also includes a control interface unit for providing a standardized communication protocol interface, outputting the deviation detection results to the industrial control system at a refresh rate greater than or equal to 1kHz, supporting the adaptive adjustment function of PID control parameters, and dynamically adjusting the response threshold of the correction actuator according to the deviation amount.

8. The detection device for rapidly measuring strip deviation using AI visual recognition technology according to claim 1 is characterized in that: The environmental compensation unit further includes a temperature compensation module and a vibration compensation module, wherein: The temperature compensation module uses a thermistor array to monitor the temperature field distribution of the camera installation area in real time, and corrects the thermal expansion and contraction errors of the optical components through a lookup table method; The vibration compensation module uses an accelerometer to collect the vibration spectrum of the mechanical structure and suppresses interference signals in the resonance frequency band through a notch filter.

Images