A belt foreign matter recognition algorithm

Abstract

A belt foreign matter recognition algorithm, the recognition algorithm comprises the following steps: collecting belt normal operation image and belt image with foreign matter, constituting abnormal detection data set and target detection data set and marking; the abnormal detection data set is divided and the image abnormal detection model is trained, the belt normal sample is divided into training set sample and verification set sample, and then the labeled abnormal image is added to the verification set sample; the belt foreign matter detection model is trained by using the labeled target detection data set, and the detection model result is displayed; the image abnormal detection model and the belt foreign matter detection model are trained, and then online detection is implemented, and the image abnormal detection result is obtained; draw the detection area, calculate the foreign matter size, and judge whether the foreign matter size exceeds the set threshold value; when the foreign matter size exceeds the set threshold value is recognized, the system time corresponding to the foreign matter is recorded and the count is increased by one.

Description

Technical Field

[0001] This invention relates to a foreign object identification algorithm for belts. Background Technology

[0002] Coal mining involves separating valuable coal from underground or the surface and transporting it away from the site. Whether it is open-pit or underground mining, geological exploration is required first to determine the distribution range of coal-bearing strata, the number of mineable strata, the thickness of the strata, the dip angle, the reserves, as well as the geological structure, spontaneous combustion tendency, water, gas and other occurrence conditions and mining conditions. Then, the construction scale, number of mines, output and construction sequence of the mining area are rationally planned. According to the overall design of the mining area and the design of the mines, they are constructed one by one and then handed over to production.

[0003] In the coal mining and conveying process, the main conveyor belt is an essential and critical piece of equipment. During coal transport, various foreign objects, such as large pieces of coal, anchor bolts, wire mesh, and reinforcing bars, will be mixed in. These foreign objects can not only scratch the conveyor belt but also cause blockages at the coal drop point. This seriously affects the service life of the conveyor belt, reduces production efficiency, and leads to economic losses.

[0004] Currently, there are two main methods for detecting foreign objects on conveyor belts. One method relies on physical equipment, such as lidar and magnetic separators. This method is expensive, susceptible to environmental influences, and can only detect a limited number of foreign objects. The other method is vision-based. This method is simple to deploy and is a non-contact measurement method, thus it has been widely used. However, this method requires a large number of foreign object samples, which are difficult to obtain. Therefore, it is impossible to train a detection model for all categories, resulting in a limited range of detection types. The model can only detect fixed categories and cannot calculate the size of foreign objects, leading to the missed detection of certain categories of foreign objects. Missed foreign objects, especially large ones, may cause blockages at the coal feeder's chute during transportation, affecting the overall efficiency of coal conveying. Summary of the Invention

[0005] This invention provides a conveyor belt foreign object identification algorithm. The method is reasonably designed and achieves full-category detection of conveyor belt foreign objects by fusing abnormal image detection and target image detection. It integrates the detection results of multiple models related to conveyor belt foreign object detection for judgment. Compared with the current machine vision-based coal conveyor belt foreign object detection using a single model, it can effectively increase the number of detection categories, improve detection accuracy, and the model has the ability to learn on the job and has high robustness. At the same time, it can calculate the size of foreign objects and use logical strategies to provide early warning of coal drop blockage, thus solving the problems existing in the prior art.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] A foreign object detection algorithm for a belt, the algorithm comprising the following steps:

[0008] S1. Collect images of the belt operating normally and images of the belt with foreign objects, and construct an anomaly detection dataset and a target detection dataset, and then label them;

[0009] S2, the anomaly detection dataset is divided and an image anomaly detection model is trained. Normal belt samples are divided into training set samples and validation set samples, and then the labeled anomaly images are added to the validation set samples. The training set samples are used to train the anomaly detection model, and the validation set samples are used to evaluate the accuracy of the anomaly detection model.

[0010] S3 uses a labeled target detection dataset to train a belt foreign object detection model and displays the detection model results;

[0011] S4, the image anomaly detection model and the belt foreign object detection model are trained and then online detection is performed to obtain the image anomaly detection results;

[0012] S5, draw the detection area, calculate the size of the foreign object, and determine whether the size of the foreign object exceeds the set threshold;

[0013] S6. When a foreign object is detected whose size exceeds the set threshold, the system time corresponding to the foreign object is recorded and the count is incremented by one. The number of times the foreign object appears within the set time interval is counted. If the number of times exceeds the set threshold, an alarm message is issued.

[0014] The steps involved in drawing the detection area, calculating the size of the foreign object, and determining whether the size of the foreign object exceeds a set threshold are as follows:

[0015] S5.1, Draw the detection area and obtain the coordinates of the two intersection points with the belt. and ;

[0016] S5.2, Define the coordinates of the upper left corner of the detection area. and the coordinates of the upper right corner The coordinates of the upper left corner of the detection frame are obtained based on the belt foreign object detection model. and the coordinates of the bottom right corner ;

[0017] S5.3, based on the belt pixel width EH at the corresponding height of the detection box and the actual width of the belt. The size of the foreign object can then be determined, i.e.:

[0018] Foreign object size = W pd ×max{(x r -x l ), (y r -y l )} / EH;

[0019] During online detection, images of the belt are captured using an intrinsically safe camera. These images are then input into a patchcore network model to obtain feature vectors for training samples. The Mahalanobis distance between the real-time image and these feature vectors is calculated to obtain anomaly detection results, which include the location and confidence level of any foreign object. Based on this, foreign object detection is performed on the belt to obtain the current image's foreign object detection results, including the category, confidence level, and location of the foreign object. The results are then fused and evaluated.

[0020] If both the anomaly detection model and the belt foreign object detection model identify foreign objects and the confidence level is greater than the set threshold, the size of the foreign object is calculated, an alarm image is uploaded, and the count is performed. If the anomaly detection model does not identify foreign objects or identifies foreign objects but the confidence level is less than the set threshold, the image is considered to be free of foreign objects.

[0021] If the anomaly detection model identifies a foreign object and the confidence level is greater than the set threshold, but the target detection model does not identify the foreign object, it means that this type of foreign object has never appeared before. Save this image, manually annotate it, and add it to the anomaly detection dataset and target detection dataset in step one to optimize the model so that the model has the ability to detect this type of foreign object.

[0022] If the anomaly detection model identifies a foreign object with a confidence level greater than the set threshold, while the target detection model identifies a foreign object with a confidence level less than the set threshold, the product of the two detection confidence levels is calculated. If the product is greater than the set threshold, the size of the foreign object is calculated, an alarm image is uploaded, and the count is performed. Otherwise, the detection continues in the next frame.

[0023] The image field of view includes the belt and the rollers on both sides. The belt occupies 60% of the image area and is centered. The image resolution is not less than 1920*1080.

[0024] The anomaly detection dataset includes images of empty conveyor belts, images of conveyor belts with loose coal but no foreign objects, and images of conveyor belts with foreign objects. The collected anomaly images are segmented and labeled to identify large pieces of coal, anchor bolts, wires, and other foreign objects.

[0025] The target detection dataset contains images of foreign objects on belts, which are labeled using the deep learning image labeling software CVAT. The label type is a rectangular box, and the label category is large coal pieces, anchor bolts, and other foreign objects.

[0026] The anomaly detection model uses the patchcore algorithm to detect foreign objects in the belt.

[0027] The belt foreign object detection model is YoloV5-m.

[0028] The foreign object identification device of the foreign object identification algorithm includes an image acquisition device, a communication transmission device, a data processing device, and a data storage device;

[0029] The image acquisition device is an intrinsically safe underground camera that can capture high-definition images during the conveyor belt operation. Dust-proof material is used on the surface of the camera to prevent the adhesion of underground dust and its impact on image quality.

[0030] The communication transmission equipment includes a 5G transmitter, a Wi-Fi transmitter, a USB transmitter, and an Ethernet transmitter;

[0031] The data processing device is an AI edge computing server, which has a certain computing power and memory to complete tasks such as image anomaly detection based on patchcore and belt foreign object detection based on YOLO.

[0032] The data storage device is an SD card or a hard drive, used to store images identified as foreign objects.

[0033] This invention employs the aforementioned structure, pre-collecting images of normal coal conveying and foreign objects on the conveyor belt using an intrinsically safe camera to establish an image dataset; fusing and judging the multi-model detection results related to foreign object detection using a multi-model element fusion detection method; transmitting data and exchanging information through various types of communication devices; storing the identified foreign object images through a data storage device; training the conveyor belt foreign object detection model using the pre-collected target detection dataset; and displaying the detection model results. This invention boasts advantages of stability, efficiency, safety, and reliability. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the process of the present invention.

[0035] Figure 2 This is a schematic diagram illustrating the calculation of foreign object dimensions according to the present invention.

[0036] Figure 3 This is a schematic diagram of the foreign object identification device of the present invention. Detailed Implementation

[0037] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0038] like Figure 1-3 As shown, a foreign object detection algorithm for a belt includes the following steps:

[0039] S1. Collect images of the belt operating normally and images of the belt with foreign objects, and construct an anomaly detection dataset and a target detection dataset, and then label them;

[0040] S2, the anomaly detection dataset is divided and an image anomaly detection model is trained. Normal belt samples are divided into training set samples and validation set samples, and then the labeled anomaly images are added to the validation set samples. The training set samples are used to train the anomaly detection model, and the validation set samples are used to evaluate the accuracy of the anomaly detection model.

[0041] S3 uses a labeled target detection dataset to train a belt foreign object detection model and displays the detection model results;

[0042] S4, the image anomaly detection model and the belt foreign object detection model are trained and then online detection is performed to obtain the image anomaly detection results;

[0043] S5, draw the detection area, calculate the size of the foreign object, and determine whether the size of the foreign object exceeds the set threshold;

[0044] S6. When a foreign object is detected whose size exceeds the set threshold, the system time corresponding to the foreign object is recorded and the count is incremented by one. The number of times the foreign object appears within the set time interval is counted. If the number of times exceeds the set threshold, an alarm message is issued.

[0045] The steps involved in drawing the detection area, calculating the size of the foreign object, and determining whether the size of the foreign object exceeds a set threshold are as follows:

[0046] S5.1, Draw the detection area and obtain the coordinates of the two intersection points with the belt. and ;

[0047] S5.2, Define the coordinates of the upper left corner of the detection area. and the coordinates of the upper right corner The coordinates of the upper left corner of the detection frame are obtained based on the belt foreign object detection model. and the coordinates of the bottom right corner ;

[0048] S5.3, based on the belt pixel width EH at the corresponding height of the detection box and the actual width of the belt. The size of the foreign object can then be determined, i.e.:

[0049] Foreign object size = W pd ×max{(x r -x l ), (y r -y l )} / EH;

[0050] During online detection, images of the belt are captured using an intrinsically safe camera. These images are then input into a patchcore network model to obtain feature vectors for training samples. The Mahalanobis distance between the real-time image and these feature vectors is calculated to obtain anomaly detection results, which include the location and confidence level of any foreign object. Based on this, foreign object detection is performed on the belt to obtain the current image's foreign object detection results, including the category, confidence level, and location of the foreign object. The results are then fused and evaluated.

[0051] If both the anomaly detection model and the belt foreign object detection model identify foreign objects and the confidence level is greater than the set threshold, the size of the foreign object is calculated, an alarm image is uploaded, and the count is performed. If the anomaly detection model does not identify foreign objects or identifies foreign objects but the confidence level is less than the set threshold, the image is considered to be free of foreign objects.

[0052] If the anomaly detection model identifies a foreign object and the confidence level is greater than the set threshold, but the target detection model does not identify the foreign object, it means that this type of foreign object has never appeared before. Save this image, manually annotate it, and add it to the anomaly detection dataset and target detection dataset in step one to optimize the model so that the model has the ability to detect this type of foreign object.

[0053] If the anomaly detection model identifies a foreign object with a confidence level greater than the set threshold, while the target detection model identifies a foreign object with a confidence level less than the set threshold, the product of the two detection confidence levels is calculated. If the product is greater than the set threshold, the size of the foreign object is calculated, an alarm image is uploaded, and the count is performed. Otherwise, the detection continues in the next frame.

[0054] The image field of view includes the belt and the rollers on both sides. The belt occupies 60% of the image area and is centered. The image resolution is not less than 1920*1080.

[0055] The anomaly detection dataset includes images of empty conveyor belts, images of conveyor belts with loose coal but no foreign objects, and images of conveyor belts with foreign objects. The collected anomaly images are segmented and labeled to identify large pieces of coal, anchor bolts, wires, and other foreign objects.

[0056] The target detection dataset contains images of foreign objects on belts, which are labeled using the deep learning image labeling software CVAT. The label type is a rectangular box, and the label category is large coal pieces, anchor bolts, and other foreign objects.

[0057] The anomaly detection model uses the patchcore algorithm to detect foreign objects in the belt.

[0058] The belt foreign object detection model is YoloV5-m.

[0059] The foreign object identification device of the foreign object identification algorithm includes an image acquisition device, a communication transmission device, a data processing device, and a data storage device;

[0060] The image acquisition device is an intrinsically safe underground camera that can capture high-definition images during the conveyor belt operation. Dust-proof material is used on the surface of the camera to prevent the adhesion of underground dust and its impact on image quality.

[0061] The communication transmission equipment includes a 5G transmitter, a Wi-Fi transmitter, a USB transmitter, and an Ethernet transmitter;

[0062] The data processing device is an AI edge computing server, which has a certain computing power and memory to complete tasks such as image anomaly detection based on patchcore and belt foreign object detection based on YOLO.

[0063] The data storage device is an SD card or a hard drive, used to store images identified as foreign objects.

[0064] The working principle of a conveyor belt foreign object recognition algorithm in this embodiment of the invention is as follows: it realizes the detection of all categories of foreign objects on the conveyor belt by a dual fusion of abnormal image detection and target image detection. It fuses and judges the detection results of multiple models related to the detection of foreign objects on the conveyor belt. Compared with the current method of using a single model to detect foreign objects on coal conveyor belts based on machine vision, it can effectively increase the number of detection categories, improve the detection accuracy, and the model has the ability to learn on the job and has high robustness. At the same time, it can calculate the size of foreign objects and use logical strategies to realize early warning of blockage at the coal drop outlet.

[0065] In the overall scheme, the recognition algorithm mainly includes the following steps: acquiring images of the belt operating normally and images of the belt with foreign objects, forming an anomaly detection dataset and a target detection dataset, and labeling them; dividing the anomaly detection dataset and training an image anomaly detection model, dividing the normal belt samples into training set samples and validation set samples, and then adding the labeled anomaly images to the validation set samples; the training set samples are used to train the anomaly detection model, and the validation set samples are used to evaluate the accuracy of the anomaly detection model; training the belt foreign object detection model using the labeled target detection dataset, and displaying the detection model results; learning and training the image anomaly detection model and the belt foreign object detection model, and then implementing online detection to obtain image anomaly detection results; drawing the detection area, calculating the size of the foreign object, and determining whether the size of the foreign object exceeds a set threshold; when the size of the foreign object exceeds the set threshold, recording the system time corresponding to this foreign object and incrementing the count, counting the number of times the foreign object appears within the set time interval, and if the number of times exceeds the set threshold, issuing an alarm message.

[0066] During the acquisition of images of the belt in operation, the field of view includes the belt and the rollers on both sides. The belt occupies 60% of the image area and is centered. The image resolution is not less than 1920*1080.

[0067] For the anomaly detection dataset, which includes images of empty conveyor belts, images of conveyor belts with loose coal but no foreign objects, and images of conveyor belts with foreign objects, the collected anomaly images are segmented and labeled to identify large pieces of coal, anchor bolts, and wire foreign objects. For the object detection dataset, which includes images of conveyor belts with foreign objects, the deep learning image labeling software CVAT is used for labeling. The label type is rectangular boxes, and the label category is large pieces of coal, anchor bolts, and other foreign objects.

[0068] The anomaly detection model uses the patchcore algorithm to detect foreign objects in the belt, which can be represented by a heatmap. Furthermore, different anomaly detection algorithms can be selected according to the actual application scenario.

[0069] Preferably, the belt foreign object detection model is YoloV5-m, which can maintain high detection accuracy and can run at real-time speed.

[0070] In practical applications, images are collected offline to create an image dataset, which is then used to train an image anomaly detection model and a belt foreign object detection model. The training of these two models is completed offline, allowing for sufficient learning and training.

[0071] During online detection, images of the belt are captured using an intrinsically safe camera. These images are then input into a patchcore network model to obtain feature vectors for training samples. By calculating the Mahalanobis distance between the real-time image and these feature vectors, anomaly detection results are obtained, including the location and confidence level of any foreign object. Based on this, foreign object detection is performed on the belt to obtain the foreign object detection results for the current image, including the category, confidence level, and location of the foreign object. The results are then fused and used for further analysis.

[0072] If both the anomaly detection model and the belt foreign object detection model identify foreign objects and the confidence level is greater than the set threshold, the size of the foreign object is calculated, an alarm image is uploaded, and the count is performed. If the anomaly detection model does not identify foreign objects or identifies foreign objects but the confidence level is less than the set threshold, then the image is considered to be free of foreign objects.

[0073] If the anomaly detection model identifies a foreign object with a confidence level greater than the set threshold, but the target detection model does not identify the foreign object, it means that this type of foreign object has never appeared before. Save the image, manually annotate it, and add it to the anomaly detection dataset and target detection dataset in step one. Optimize the model so that it has the ability to detect this type of foreign object.

[0074] If the anomaly detection model identifies a foreign object with a confidence level greater than the set threshold, while the target detection model identifies a foreign object with a confidence level less than the set threshold, the product of the two detection confidence levels is calculated. If the product is greater than the set threshold, the size of the foreign object is calculated, an alarm image is uploaded, and the count is performed. Otherwise, the detection continues in the next frame.

[0075] The system counts the number of times foreign objects of different sizes appear in the 10 minutes preceding the current moment. If the number exceeds a set threshold, an audio alarm is triggered, indicating a risk of blockage at the coal chute.

[0076] The calculation of foreign object size mainly includes the following steps: S5.1, draw the detection area and obtain the coordinates of the two intersection points with the belt. and ;

[0077] S5.2, Define the coordinates of the upper left corner of the detection area. and the coordinates of the upper right corner The coordinates of the upper left corner of the detection frame are obtained based on the belt foreign object detection model. and the coordinates of the bottom right corner ;

[0078] S5.3, based on the long side EH of the detection frame and the actual width of the belt. The size of the foreign object can then be determined, i.e.:

[0079] Foreign object size = W pd ×max{(x r -x l ), (y r -y l )} / EH;

[0080] It should be noted that the foreign object recognition device of the foreign object recognition algorithm includes an image acquisition device, a communication transmission device, a data processing device, and a data storage device; the image acquisition device is an intrinsically safe underground camera capable of capturing high-definition images during conveyor belt operation; the camera surface uses dust-proof material to prevent underground dust from adhering and affecting image quality; the communication transmission device includes a 5G transmitter, a Wi-Fi transmitter, a USB transmitter, and an Ethernet transmitter; the data processing device is an AI edge computing server with sufficient computing power and memory to complete patchcore-based image anomaly detection and YOLO-based conveyor belt foreign object detection tasks; the data storage device is an SD card or hard drive used to store images identified as foreign objects.

[0081] In summary, the conveyor belt foreign object recognition algorithm in this embodiment of the invention achieves full-category detection of conveyor belt foreign objects by fusing abnormal image detection and target image detection. It integrates and judges the detection results of multiple models related to conveyor belt foreign object detection. Compared with the current machine vision-based coal conveyor belt foreign object detection using a single model, it can effectively increase the number of detection categories, improve detection accuracy, and the model has the ability to learn on the job and has high robustness. At the same time, it can calculate the size of foreign objects and use logical strategies to provide early warning of blockage at the coal drop outlet.

[0082] The above specific embodiments should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, any alternative improvements or modifications made to the embodiments of the present invention shall fall within the scope of protection of the present invention.

[0083] Any aspects of this invention not described in detail are well-known to those skilled in the art.

Claims

1. A foreign object detection algorithm for a belt, characterized in that, The recognition algorithm Includes the following steps: S1. Collect images of the belt operating normally and images of the belt with foreign objects, and construct an anomaly detection dataset and a target detection dataset, and then label them; S2, the anomaly detection dataset is divided and an image anomaly detection model is trained. Normal belt samples are divided into training set samples and validation set samples, and then the labeled anomaly images are added to the validation set samples. The training set samples are used to train the anomaly detection model, and the validation set samples are used to evaluate the accuracy of the anomaly detection model. S3 uses a labeled target detection dataset to train a belt foreign object detection model and displays the detection model results; S4, the image anomaly detection model and the belt foreign object detection model are trained and then online detection is performed to obtain the image anomaly detection results; S5, draw the detection area, calculate the size of the foreign object, and determine whether the size of the foreign object exceeds the set threshold; S6. When a foreign object is detected whose size exceeds the set threshold, the system time corresponding to the foreign object is recorded and the count is incremented by one. The number of times the foreign object appears within the set time interval is counted. If the number of times exceeds the set threshold, an alarm message is issued. The steps involved in drawing the detection area, calculating the size of the foreign object, and determining whether the size of the foreign object exceeds a set threshold are as follows: S5.1, Draw the detection area and obtain the coordinates of the two intersection points with the belt. and ; S5.2, Define the coordinates of the upper left corner of the detection area. and the coordinates of the upper right corner The coordinates of the upper left corner of the detection frame are obtained based on the belt foreign object detection model. and the coordinates of the bottom right corner ; S5.3, based on the belt pixel width EH at the corresponding height of the detection box and the actual width of the belt. The size of the foreign object can then be determined, i.e.: Foreign object size = W pd ×max{(x r -x l ), (y r -y l )} / EH; During online detection, images of the belt are captured using an intrinsically safe camera. These images are then input into a patchcore network model to obtain feature vectors for training samples. By calculating the Mahalanobis distance between the real-time image and these feature vectors, anomaly detection results are obtained, including the location and confidence level of the foreign object. Based on this, foreign object detection is performed on the belt to obtain the foreign object detection results for the current image, including the category, confidence level, and location of the foreign object. Finally, the obtained results are fused and judged. If both the anomaly detection model and the belt foreign object detection model identify foreign objects and the confidence level is greater than the set threshold, the size of the foreign object is calculated, an alarm image is uploaded, and the count is performed. If the anomaly detection model does not identify foreign objects or identifies foreign objects but the confidence level is less than the set threshold, the image is considered to be free of foreign objects. If the anomaly detection model identifies a foreign object and the confidence level is greater than the set threshold, but the target detection model does not identify the foreign object, it means that this type of foreign object has never appeared before. Save this image, manually annotate it, and add it to the anomaly detection dataset and target detection dataset in step one to optimize the model so that the model has the ability to detect this type of foreign object. If the anomaly detection model identifies a foreign object with a confidence level greater than the set threshold, while the target detection model identifies a foreign object with a confidence level less than the set threshold, the product of the two detection confidence levels is calculated. If the product is greater than the set threshold, the size of the foreign object is calculated, an alarm image is uploaded, and the count is performed. Otherwise, the detection continues in the next frame.

2. The belt foreign object identification algorithm according to claim 1, characterized in that: The image field of view includes the belt and the rollers on both sides. The belt occupies 60% of the image area and is centered. The image resolution is not less than 1920*1080.

3. The belt foreign object identification algorithm according to claim 1, characterized in that: The anomaly detection dataset includes images of empty conveyor belts, images of conveyor belts with loose coal but no foreign objects, and images of conveyor belts with foreign objects. The collected anomaly images are segmented and labeled to identify large pieces of coal, anchor bolts, wires, and other foreign objects. The target detection dataset contains images of foreign objects on belts, which are labeled using the deep learning image labeling software CVAT. The label type is a rectangular box, and the label category is large coal pieces, anchor bolts, and other foreign objects.

4. The belt foreign object identification algorithm according to claim 1, characterized in that: The anomaly detection model uses the patchcore algorithm to detect foreign objects in the belt.

5. The belt foreign object identification algorithm according to claim 1, characterized in that: The belt foreign object detection model is YoloV5-m.

6. The belt foreign object identification algorithm according to claim 1, characterized in that: The foreign object identification device of the foreign object identification algorithm includes an image acquisition device, a communication transmission device, a data processing device, and a data storage device; The image acquisition device is an intrinsically safe underground camera that can capture high-definition images during the conveyor belt operation. Dust-proof material is used on the surface of the camera to prevent the adhesion of underground dust and its impact on image quality. The communication transmission equipment includes a 5G transmitter, a Wi-Fi transmitter, a USB transmitter, and an Ethernet transmitter; The data processing device is an AI edge computing server, which has a certain computing power and memory to complete tasks such as image anomaly detection based on patchcore and belt foreign object detection based on YOLO. The data storage device is an SD card or a hard drive, used to store images identified as foreign objects.

Images