Method for online quality detection of webbing machine vision

Abstract

The present application provides a woven tape machine vision online quality detection method, which is realized based on a visual online quality detection device arranged on one side of a polyester woven tape machine, and realizes flaw judgment by using a weight file loaded on a hardware development board. The weight file is obtained based on different categories of woven tapes in the following manner: S01, creating a data set; S02, training a model and exporting a weight file; and S03, placing the weight file on the hardware development board. The detection process is specifically as follows: the woven tape is collected by real-time photographing by a camera device of the detection device, and is classified by loading a classification weight file, and then whether the flaw degree exceeds a threshold value is judged based on a modified STPM algorithm. The present application can improve the woven tape flaw detection efficiency and detection effect, effectively reduce the hardware cost and maintenance cost, and has strong applicability to different enterprise environments by learning and training a model by using a flawless data set.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of woven tape flaw quality detection, and particularly relates to a woven tape machine vision online quality detection method. BACKGROUND

[0002] Woven tape is a kind of fabric with various yarns as raw materials, and there are various kinds of woven tape products, which are widely used in clothing, shoe materials, luggage, industry, agriculture, military supplies, transportation and other industrial departments. At present, the raw materials for woven tape are mostly nylon, vinylon, polyester, polypropylene, spandex and viscose, forming three major process technologies of weaving, braiding and knitting, and the fabric structure has plain weave, twill weave, satin weave, jacquard weave, double-layer weave, multi-layer weave, tubular weave and combined weave. The production of woven tape is large in the production process, so the production quality detection of woven tape has become a problem that many enterprises need to think about. The traditional detection is carried out by manual detection, which has low detection efficiency and high labor cost. However, in recent years, the method and technology of woven tape flaw detection have developed rapidly, which has solved the problem of woven tape quality flaw detection for many enterprises. The main detection methods are as follows: 1. using high-speed image scanning algorithm, using dynamic size, threshold and linear and nonlinear filter transformation to detect edges and boundaries for further analysis; 2. template matching to find the position of the flaw; 3. extracting image description rules from local areas.

[0003] However, in the process of implementing the technical scheme of the present application, the present inventors have found that the above-mentioned technology at least has the following technical problems: 1. the above-mentioned detection method uses image difference method for detection, and is mostly suitable for flaw detection problems that cannot prepare flaw models, such as scratch or superposition detection, texture image analysis, etc., and the detection effect of woven tape flaws with multiple categories and multiple structures is not good; 2. in the flaw detection, the camera needs to take images, which are then processed by hardware and software, and the above-mentioned method has high requirements for the resolution of the image, the performance of the camera and the hardware processor, and the overall hardware and software cost is high; 3. in the traditional detection method, a model is prepared by using deep learning, which often needs to provide a large number of flaw samples for training, but for some enterprises, it is difficult to provide flaw samples, so the detection method of deep learning is difficult to use. SUMMARY

[0004] The present application provides a woven tape machine vision online quality detection method, which can be used in most enterprises or factories by training with non-defect samples to determine whether the target sample contains flaws, and has low requirements for image resolution, camera and hardware processor, effectively reducing the hardware cost and maintenance cost of the project. Through intelligent deep learning and training of non-flaw model, the woven tape flaw detection efficiency and detection effect are improved to a certain extent.

[0005] To solve the above technical problems, the application discloses a woven tape machine vision online quality detection method, which is realized based on a visual online quality detection device arranged on one side of a polyester woven tape machine, and the detection object is a woven tape produced by the woven tape machine; the detection method realizes flaw judgment by using a weight file loaded on a hardware development board, and the weight file of different categories of woven tapes is obtained in the following manner: S01, creating a data set; collecting woven tape pictures with flaws and without flaws respectively, and obtaining a flawed data set and a flawless data set after conversion; S02, training a model; training an artificial intelligence model by using the flawless data set, and exporting a weight file; S03, placing the weight file; and placing the weight file trained in the above step on the hardware development board.

[0006] The detection process is as follows: the camera equipment of the detection device takes pictures of the woven tape in real time, the collected woven tape pictures are converted, the converted pictures are classified by loading the classification weight file in S03, and then it is judged whether the flaw degree exceeds the threshold value based on the modified STPM algorithm; when the flaw degree exceeds the threshold value, the level signal of the flaw exceeding the threshold value is sent through GPIO and an alarm shutdown is performed.

[0007] The technical scheme of the application has high expansibility and can provide independent neural network calculation units, reduces the overall cost under the premise of balancing performance and scalability; the model is trained and learned by using the flawless data set, and the flawed data set is used for detection and evaluation, so that the flaw quality detection of the target sample can be judged, and the detection efficiency and detection effect are improved to a certain extent.

[0008] As a preferred embodiment of the application, S01 specifically comprises the following sub-steps:

[0009] S011, picture size processing and background bright noise removal through preprocessing, the pictures are classified into flawed pictures and flawless pictures by the pattern type on the woven tape and whether the woven tape is a defective product.

[0010] S012, converting the pictures into tensors and performing normalization processing, the pictures are converted into tensors by using the PIL scheme, and the picture data is normalized by using the OpenCV function Normalize through the mean value = [0.485, 0.456, 0.406] and the standard deviation = [0.229, 0.224, 0.225] in the ImageNet database.

[0011] S013, adjusting and unifying the picture brightness, after the pictures are converted from the RGB mode to the HSV mode, the brightness of all the converted pictures is adjusted to be equal by using the PIL ImageEnhance.

[0012] The technical scheme of the present application classifies pictures in advance, which can improve the recognition of pictures under the same condition, thereby ensuring the further reduction of image resolution and operation amount under the same recognition rate to improve the recognition speed; in S013, the influence of environmental light changes on the recognition result in actual conditions can be avoided, and the flaw recognition and detection effect is improved.

[0013] As preferred in the present application, in S02, the following sub-steps are specifically included:

[0014] S021, modify the model output to the pyramid structure algorithm of STPM, remove the two-dimensional average pooling layer, the full connection layer, the third layer and the fourth layer on the basis of the Torchvision standard ResNet18 neural network, and then change the model output to the pyramid structure of STPM, output the results X1 of the first layer and X2 of the second layer with X1 as the input, i.e. the values of large flaws and small flaws respectively.

[0015] S022, declare the teacher model and the student model, based on the STPM algorithm in S021, the teacher model is pre-trained based on the ImageNet data set and has obtained the weight, and the student model is the data set in S01 to be trained.

[0016] S023, train the student model and export the weights of the teacher model and the student model, perform Online Learning on the flawless data set of the student model and adjust the learning rate through Step LR to make the student model tend to the teacher model optimization, and export and save the weight files of the student model and the teacher model: the student weight file a and the teacher weight file b.

[0017] S024, train the single-class ribbon model, execute S022 and S023 on each ribbon category data set in step S01 respectively, and correspondingly export the single-class student weight file a n and the single-class teacher weight file (b n ), wherein a n and b n are n kinds of ribbon patterns.

[0018] S025, train the classification accuracy and weight export, train the data set in S01 through the MobileNetV2 neural network embedded device for epoch=13, and export the classification weight file c after the classification accuracy of the ribbon is greater than 98%.

[0019] The technical scheme of the present application removes the two-dimensional average pooling layer, the full connection layer, the third layer and the fourth layer on the basis of the Torchvision standard ResNet18 neural network, sacrifices slight accuracy to obtain speed improvement to adapt to the embedded application environment, and simultaneously trains and learns and classifies the flawless data set through Online Learning and MobileNetV2, so that the weight file is more comprehensive, and the flaw detection and recognition effect of different types of braid patterns can be guaranteed.

[0020] As preferred in the present application, the following sub-steps are specifically included in S023:

[0021] S0231, determine the average loss between the models, set the data set sample quantity batch size=1 and use the Online Learning model training method, train the student model with the flawless data set, obtain the operation results between the student model and the teacher model, and normalize to obtain the loss L of each layer. n , and the loss of each layer (n is the number of layers) is combined to obtain the average loss L of each layer.

[0022] S0232, optimize the student model, use the gradient descent method SGD optimizer of the step learning rate adjustment mechanism Step LR to optimize in the direction of lower loss, that is, the student model tends to optimize the teacher model, and save the student model and teacher model weight files after training.

[0023] The technical scheme of the present application can reduce the model operation time by setting the data set sample quantity to 1, and can improve the accuracy of flaw detection and reduce detection errors by adjusting and optimizing the learning rate of the trained and normalized model.

[0024] As preferred in the present application, the following sub-steps are specifically included in the real-time photographing, collecting and converting process of the camera equipment of the detection device:

[0025] S041, braid pattern classification, the camera in the detection device collects real-time pictures of the braid and converts them into float32 floating point arrays, loads the classification weight file c in S025 and obtains the braid pattern classification result d.

[0026] S042, calculate the flaw distribution map; load the single-class student weight file an and the single-class teacher weight file bn in S041 corresponding to the braid pattern classification result d, input the picture array based on the modified STPM pyramid structure algorithm in S02 and execute the following steps:

[0027] S0421, generate w x h feature vector tuples F s and Ft , execute the OpenCV function Normalize in S01 to obtain: F sN and F tN .

[0028] S0422, according to the calculation, the defect distribution map is as follows:

[0029] In the formula, w and h are the sizes of the picture array respectively;

[0030] Through matrix calculation, the distribution map is reduced to w1xh1 by Bilinear Interpolation, and the results of each layer of the pyramid are superimposed, and after Gaussian Filter noise reduction, whether the defect degree of the point with the largest defect degree exceeds the threshold value is judged.

[0031] In S0422, based on the function score of the over-limit degree, if the cumulative score exceeding the threshold value reaches a specific defect determination value within n times, execute S05 to send the level of the defect exceeding the threshold value through GPIO and perform alarm shutdown; in S0422, if the defect degree does not exceed the threshold value, execute S05 to send the level data of the defect not exceeding the threshold value to the cloud platform.

[0032] The whole process of the technical scheme of the application is completed on a local embedded device, without uploading data to a server or a central computer room, which can be applied to embedded devices with weak computing power without affecting the efficiency of the belt defect detection, and has a wider application range.

[0033] As a preferred embodiment of the application, if the level signal is a defect degree exceeding the threshold value, the following steps are executed:

[0034] S051, send the level signal to the Internet of Things platform, send a hardware shutdown signal to the belt machine through the Internet of Things platform to stop working and perform flashing warning.

[0035] Further comprising the following sub-steps:

[0036] S052, the Internet of Things platform sends the defect situation to the server end to complete recording, and after receiving, the server end sends it to the front end large screen and mobile terminal in real time, and after processing at the front end, the defect situation is presented to the on-site staff in real time.

[0037] The technical scheme of the application combines the Internet of Things platform and the detection device, directly sends the detection signal and displays it on the front end and the terminal, which can effectively reduce the cost of traditional manual labor, and realize automatic and intelligent operation of detection. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1A flow chart of the online quality detection method of the weaving tape machine vision of the present application;

[0039] Figure 2 A classification pattern schematic diagram of the online quality detection method of the weaving tape machine vision of the present application; DETAILED DESCRIPTION

[0040] The specific embodiments of the present application will be further described in detail below in combination with the drawings and examples. The following examples are used to illustrate the present application, but are not used to limit the scope of the present application. The following contents not described in detail all belong to the technical common sense in the art.

[0041] Reference Figure 1 and Figure 2 In specific embodiments, the method is realized based on the visual online quality detection device arranged on one side of the polyester tape machine, and the detection object is the weaving tape produced by the weaving tape machine. The weaving tape machine detection device includes a terminal processing device for data processing, sending and recording, a camera processing device for taking pictures of the polyester tape and transmitting image data to the terminal processing device, and an alarm device for sound and light alarm of the defective polyester tape identified by the camera processing device.

[0042] The detection method needs to load the weight file on the hardware development board to realize the defect judgment, and the specific ways of saving and exporting the weight file based on different categories of weaving tapes are as follows:

[0043] S01, create a data set, respectively collect weaving tape pictures with defects and without defects, and obtain a defective data set and a non-defective data set after conversion.

[0044] S012, convert the picture to a tensor and perform normalization processing, use the PIL scheme to convert the picture to a tensor, and use the mean value = [0.485, 0.456, 0.406] and the standard deviation = [0.229, 0.224, 0.225] in the ImageNet database to normalize the picture data using the OpenCV function Normalize; S013, adjust and unify the picture brightness, after converting the picture from RGB mode to HSV mode, adjust the brightness of all converted pictures to be equal through PIL ImageEnhance.

[0045] S02, train the model, use the non-defective data set to train the artificial intelligence model, and export the weight file.

[0046] S021, modify the model output to the pyramid structure algorithm of STPM, on the basis of the Torchvision standard ResNet18 neural network, remove the two-dimensional average pooling layer, the full connection layer, the third layer and the fourth layer, and then change the model output to the pyramid structure of STPM. The results X1 of the first layer and the results X2 of the second layer inputting X1 are output, that is, the numbers of large and small defects respectively.

[0047] S022, declare the teacher model and the student model, based on the STPM algorithm in S021, the teacher model is pre-trained on the ImageNet dataset and has obtained the weight, and the student model is the dataset in S01 to be trained.

[0048] S023, train the student model and export the teacher model and student model weights, perform Online Learning online learning on the flawless dataset of the student model and optimize the student model to the teacher model through Step LR learning rate adjustment, and export and save the weight files of the two: student weight file a and teacher weight file b; S023 specifically includes the following sub-steps:

[0049] S0231, determine the average loss between the models, set the sample size of the dataset batch size = 1 and use the Online Learning model training method, train the student model with the flawless dataset, obtain the operation results between the student model and the teacher model, and normalize to obtain the loss L of each layer. n , combine the losses of each layer (n is the number of layers) to obtain the average loss L of each layer.

[0050] S0232, optimize the student model, use the gradient descent method SGD optimizer of the step learning rate adjustment mechanism Step LR to optimize in the direction of lower loss, that is, the student model tends to optimize the teacher model, save the student model and teacher model weight files after training.

[0051] S024, train the single-class ribbon model, execute S022 and S023 for each ribbon category dataset in step S01, and correspondingly export the single-class student weight file a n and the single-class teacher weight file (b n ), wherein a n and b n are n types of ribbon patterns.

[0052] S025, train the classification accuracy and weight export, train the dataset in S01 through the MobileNetV2 neural network embedded device for epoch = 13, and export the classification weight file c after the classification accuracy of the ribbon is greater than 98%.

[0053] By removing the two-dimensional average pooling layer, the full connection layer, the third layer and the fourth layer on the basis of the Torchvision standard ResNet18 neural network, sacrificing slight accuracy to exchange for speed improvement to adapt to embedded application environment, and through Online Learning and MobileNetV2 to train and learn and classify the flawless data set, the weight file can be more comprehensive, and the flaw detection and recognition effect of different types of webbing patterns can be guaranteed.

[0054] S03, placing the weight file, placing the weight file trained in the above steps on the hardware development board.

[0055] The hardware development board in the embodiment is Sipeed Maix II Dock hardware, and is based on ArmV7l platform and Linux system, so that it has high expansibility and can provide independent NPU neural network calculation unit, reduces the overall cost under the premise of balancing performance and scalability; the size format of the picture in S01 is saved as 144x144 and 244x244 respectively, and the picture data can be conveniently processed through Normalize normalization processing; in addition, the picture is converted and adjusted to the same brightness by PILImageEnhance, so as to avoid the influence of the environment.

[0056] The webbing machine vision online quality detection method is realized by loading the above weight file, and specifically as follows: real-time detection is performed on the webbing picture, the webbing picture collected by the camera of the detection device is converted and then loaded into the classification weight file in S03 to perform classification, and then it is judged whether the flaw degree exceeds the threshold value based on the modified STPM algorithm; the level of the flaw exceeding the threshold value is sent through GPIO and an alarm shutdown is performed, or the level data of the flaw not exceeding the threshold value is sent to the cloud platform and the new webbing is continuously detected.

[0057] The detection process of the real-time picture of the webbing through the camera equipment of the detection device for real-time detection of the webbing picture specifically includes the following sub-steps:

[0058] S041, webbing pattern classification, the camera in the detection device collects the real-time picture of the webbing and converts it into a float32 floating point array, loads the classification weight file c in S025 and obtains the webbing pattern classification result d.

[0059] S042, calculating the flaw distribution map; loading the single-class student weight file an and the single-class teacher weight file bn in S041 corresponding to the webbing pattern classification result d, inputting the picture array based on the modified STPM pyramid structure algorithm in S02 and executing the following steps:

[0060] S0421, generate w x h feature vector tuples F for each layer of the pyramid through the student model and the teacher model respectively s and F t , execute the OpenCV function Normalize in S01 to obtain F sN and F tN ;

[0061] S0422, according to the calculation, the flaw distribution map is as follows:

[0062] where w and h are the length and width dimensions of the picture array respectively;

[0063] Through matrix calculation, the Bilinear Interpolation is performed and the distribution map is reduced to w1 x h1. The results of each layer of the pyramid are superimposed, and after Gaussian Filter noise reduction, it is determined whether the flaw degree of the point with the largest flaw degree exceeds the threshold value.

[0064] In S0422, based on the function score of the over-limit degree, if the cumulative score exceeding the threshold value reaches a specific flaw determination value within n times, execute S05 to send the level of the flaw exceeding the threshold value through GPIO and perform alarm shutdown; in S0422, if the flaw degree does not exceed the threshold value, execute S05 to send the level data of the flaw not exceeding the threshold value to the cloud platform and.

[0065] The woven tape pattern is first classified through the classification weight file, then the weight file corresponding to its pattern category is loaded, and the flaw distribution map is obtained by calculating and processing based on the modified STPM pyramid structure algorithm, wherein one of the better solutions balancing time and quality is 18 x 18.

[0066] In specific embodiments, if the flaw degree does not exceed the threshold value, the level signal is sent to the central server or the cloud platform for data storage processing, and if the level signal is the flaw degree exceeding the threshold value, the following steps are executed:

[0067] S051, send the level signal to the Internet of Things platform, send the hardware shutdown signal through the Internet of Things platform to make the woven tape machine stop working and perform flashing warning.

[0068] S052, the Internet of Things platform sends the flaw condition to the server end to complete recording, after receiving, the server end sends it to the front end large screen and mobile terminal in real time, after processing at the front end, the flaw condition is presented to the on-site staff in real time.

[0069] The embodiment realizes judging whether the target sample contains defects by training with a defect-free sample, can be applied to most enterprises or factories, has low requirements on image resolution, a camera and a hardware processor, effectively reduces project hardware cost and maintenance cost, and has a certain degree of improvement in the efficiency of the braid defect detection through intelligent deep learning and training of a defect-free model; and in combination with the detection device, the braid can realize automatic and intelligent defect quality detection and alarm processing, reduces labor cost and guarantees detection effect.

[0070] The above is only the preferred embodiment of the present application, and it should be pointed out that, for those skilled in the art, without departing from the technical principles of the present application, a number of improvements and refinements can be made, which should also be considered as the protection scope of the present application.

Claims

1. A machine vision-based online quality inspection method for webbing, characterized in that, This detection method is based on an online visual quality inspection device installed on one side of the polyester webbing machine. The object of the inspection is the webbing produced by the webbing machine. This detection method uses a weight file loaded on the hardware development board to determine defects. The weight files for different types of webbing are obtained as follows: S01. Create datasets; collect images of flawed and flawless webbing respectively, and convert them to obtain flawed datasets and flawless datasets; S02, Training the model; Train an artificial intelligence model using a flawless dataset and export the weight file; S03, Place the weight file; Place the weight file obtained from the training in the above steps onto the hardware development board; The detection process is as follows: The camera device of the detection device takes real-time pictures of the webbing. After the captured webbing pictures are converted, they are first classified by loading the classification weight file in S03. Then, the defect level is judged based on the modified STPM algorithm to determine whether the defect level exceeds the threshold. When the defect level exceeds the threshold, a defect level signal exceeding the threshold is sent through GPIO and an alarm is triggered to stop the machine. S01 specifically includes the following sub-steps: S011. Image size processing and preprocessing to remove bright noise from the background. The images are classified into flawed images and flawless images according to the type of pattern on the webbing and whether the webbing is defective. S012. Convert the image into a tensor and perform normalization. Use the PIL scheme to convert the image into a tensor and use the mean = [0.485, 0.456, 0.406] and std standard deviation = [0.229, 0.224, 0.225] from the ImageNet database to normalize the image data using the OpenCV function Normalize. S013. Adjust and unify image brightness. After converting the image from RGB mode to HSV mode, use PIL ImageEnhance to adjust the brightness of all converted images to be equal. S02 specifically includes the following steps: S021. Modify the model output to the STPM pyramid structure algorithm. Based on the Torchvision standard ResNet18 neural network, remove the two-dimensional average pooling layer, the fully connected layer, the third layer and the fourth layer, and then change the model output to the STPM pyramid structure. Output the result of the first layer X1 and the result of the second layer X2 with X1 as input, which are the values ​​of large flaws and small flaws, respectively. S022. Declare the teacher model and student model, based on the STPM algorithm in S021. The teacher model is pre-trained on the ImageNet dataset and has already obtained weights, while the student model is the dataset in S01 to be trained. S023. Train the student model and export the weights of the teacher model and student model. Perform online learning on the flawless dataset of the student model and adjust the learning rate through Step LR to make the student model tend to optimize the teacher model. Export and save the weight files of both: student weight file (a) and teacher weight file (b). S024. Train the single-class webbing model. For each webbing type dataset in step S01, execute S022 and S023 respectively, and export the corresponding single-class student weight file (a). n ) and single-category teacher weighting documents (b n ), where a n and b n There are n types of ribbon patterns; S025, Training classification accuracy and weight export: The dataset in S01 is trained with epoch=13 using a MobileNetV2 neural network embedded device, and the classification weight file (c) is exported after the webbing classification accuracy is greater than 98%.

2. The online quality inspection method for webbing using machine vision as described in claim 1, characterized in that, S023 specifically includes the following steps: S0231. Determine the average loss between models. Set the dataset sample size to 1 and use the Online Learning model training method. Train the student model using a flawless dataset, obtain the computational results between the student model and the teacher model, and normalize them to obtain the loss of each layer, which is L. n The average loss of each layer is obtained by merging the losses of each layer, which is L. S0232. Optimize the student model by using the Step Learning Rate (Step LR) gradient descent (SGD) optimizer to optimize in the direction of lower loss, i.e., the student model tends to optimize the teacher model. After training, save the weight files of the student model and the teacher model respectively.

3. The online quality inspection method for webbing using machine vision as described in claim 1, characterized in that, The real-time detection process for webbing images includes the following steps: S041. Webbing pattern classification: The camera in the detection device captures real-time images of the webbing and converts them into a float32 floating-point array. The classification weight file (c) in S025 is loaded and the webbing pattern classification result (d) is obtained.

4. The online quality inspection method for webbing using machine vision as described in claim 3, characterized in that, It also includes the following steps: S042. Calculate and obtain the defect distribution map; load the single-class student weight file (an) and single-class teacher weight file (bn) from S041 corresponding to the webbing pattern classification result (d), input the image array based on the modified STPM pyramid structure algorithm in S02, and perform the following steps: S0421, each layer of the pyramid is generated using both the student model and the teacher model. Feature vector tuple F s and F t， The OpenCV function Normalize in S01 is used for normalization to obtain: F sN and F tN ; S0422, the defect distribution diagram obtained from the calculation is as follows: In the formula, w and h are the dimensions of the image array, respectively; Bilinear interpolation is performed using matrix calculations and then reduced to [the desired value]. The distribution map is obtained, and the results of each layer of the pyramid are superimposed. After noise reduction by Gaussian Filter, the result is judged based on whether the degree of defect of the point with the greatest defect exceeds the threshold.

5. The online quality inspection method for webbing using machine vision as described in claim 4, characterized in that, In S0422, a function scoring based on the degree of exceeding the limit is used. If the cumulative score of exceeding the threshold within n times reaches a specific defect judgment value, the level of the defect exceeding the threshold is sent via GPIO and an alarm shutdown is triggered in S05.

6. The online quality inspection method for webbing using machine vision as described in claim 4, characterized in that, In S0422, if the defect level does not exceed the threshold, then in S05, the level data of the defect level not exceeding the threshold is sent to the cloud platform.

7. The online quality inspection method for webbing using machine vision as described in claim 1, characterized in that, If the level signal has a defect level exceeding the threshold, then perform the following steps: S051. Send the level signal to the IoT platform, and send a hardware shutdown signal through the IoT platform to stop the weaving machine and issue a flashing warning.

8. The online quality inspection method for webbing using machine vision as described in claim 7, characterized in that, It also includes the following steps: S052. The IoT platform sends the defect information to the server to complete the record. After receiving the record, the server sends it to the front-end large screen and mobile terminal in real time. After processing at the front end, the defect information is presented to the on-site staff in real time.

Images