Shulou(Shulou.com)06/02 Report--

Hu Feng, Senior algorithm expert of Kangaroo Cloud

Head of kangaroo cloud artificial intelligence laboratory. Ten years of research and development work in image processing, pattern recognition, machine vision, etc., the main research fields are: intelligent transportation, industrial vision analysis, OCR, video intelligent quality diagnosis, intelligent ball machine tracking, human face, etc., published more than ten first author invention patents and one international invention patent.

China's industrial Internet has formed "three ways": the first is to build an intelligent factory to improve the internal production efficiency of the enterprise, through Internet technology and some related machine vision, image algorithms and pattern recognition, improve the production efficiency of the enterprise; the second is to build intelligent products to extend the external value chain of the enterprise. The third is to gather the resources of the industrial chain, realize the transformation to the platform operation, and create the ecological operation ability driven by data.

Based on the Internet, combined with the industrial scenario, a solution to add cloud is proposed, as shown in the following figure:

The complete solution of end + cloud

The deployment of the solution is divided into local and cloud. The local side is divided into two parts: one is the data layer, the real-time data generated by the data layer in some production equipment and corresponding business systems. The other is the local running layer, the generated data passes through the local running layer, and the running layer acquires the data in real time. After obtaining the data, on the one hand, it stores the data, on the other hand, it carries on the model calculation to the acquired data. After the calculation, the decision is made, and the next round of operation can be carried out after the decision.

Receive offline sequence data in the cloud, carry out a model analysis and model training based on big data platform in the cloud, and then send the updated model to the local for update after training, which in turn can continuously improve the system ability and algorithm ability.

Application scene Analysis based on Machine Vision

Human's understanding of the world is very simple. We can tell at a glance whether a fruit is an apple or an orange. But with a computer, how can you write a program for the machine to distinguish people from cats, apples and oranges? In fact, in the process of thinking, we can think based on the following aspects. first of all, we can look at it in the most intuitive way, such as color, shape, texture, edge outline, etc., and convert it into a computer. we extract the lowest three component features through the color histogram, the apple is biased towards red and the orange is biased towards orange, so it can be analyzed. Then through the shape, for example, the outline of a provincial map is extracted as a feature, and the distance from the center of the image to the outline is used as a distance feature; at a higher level, structural features can be carried out based on the underlying features. some edges and shapes make a structural analysis.

The structural feature on the far left is actually a psychological graph, and different people see different dimensions. some people see a white wine glass, but others look at it instead, these are two face-to-face faces. this kind of image analysis can put forward the structure, and then the diamond structure on the right, as well as the texture structure of some knowledge maps, are structural features.

Feature extraction by deep learning

In front of it, the traditional image processing methods are commonly used, which adopt the feature extraction method based on deep learning. The method of deep learning is relatively simple, which is to input the image we input into the network, and each network carries out an iterative training for a different process. In the early stage of our network, like some low-level features, such as simple edge texture features, in the middle part, we can learn a relatively advanced structural feature, such as when we are human face recognition. We can extract some of the eyes, mouth, ears and so on, and then go higher, and finally we can train the model of the object. For example, the rough outline of human face is trained, which is the feature extracted by deep learning method.

Anomaly detection and analysis

After feature extraction, anomaly detection and analysis is carried out. The goal in the picture is for us to find out which one is different from the others. Because there will be some abnormal target anomaly analysis in the industrial scene, the anomaly analysis is very similar to the above figure, and it is also the same in the industry, we just want to automatically locate and classify the abnormal targets.

For abnormal target analysis, the first step is to carry out a model training. the process of model training is like this. First, feature extraction is performed on the acquired image, such as some colors, textures, structures, and even some deep learning methods are used to automatically extract features. After feature extraction, a model is established, and a model expression is obtained. For example, our model can be represented by the distribution function in the graph. The distribution form of the model has something to do with the method we established, we may use some information methods, or it may be a two-dimensional surface, or a multi-dimensional surface.

Take the two-dimensional plane as an example, assuming that the feature extracted by the model is x1century x2, the model we have learned is a classification surface, and there is a normal sample in the classification plane, and assume that x1 is positively correlated with x2, that is, the larger x1 is, the larger x2 is. The red abnormal points can be detected automatically by model calculation. Now when dealing with this kind of anomaly, we can automatically distinguish it outside the classification plane.

The model must be used after training, first deploy the trained model to the production line, and then obtain the image in real time after starting on the production line, then extract the features, and output the final decision after the model is output. In the analysis process of some production lines, we monitor some key curves in real time, and visualize the real-time output results of these curves, such as these curves in the figure, some of the key factors in the process we are doing are printed out by the way of curves.

Through the curve, we can judge the maintenance, for example, which production line we are, or which machine has gone wrong, and then there will be some anomalies corresponding to the history of this curve, and we can use the curve to make real-time decisions. and we can judge when these machines began to break down according to some historical information.

Image scene analysis

In addition, in some practical cases, we will encounter some relatively harsh environments, for example, in some environments, some of them are toxic on the production line, and there are also scenes of high temperature, high pressure and high radiation. This situation is difficult to observe with human eyes or some visible light cameras. We use the infrared thermal camera to analyze, obtain the infrared thermal induction image result picture, and then segment the result image, and then we can get the corresponding container position after segmentation. then we carry on an abnormal analysis to the inside of the container position, we can get the abnormal region, and then classify the abnormal area, because there are a variety of impurities, liquid and solid. In the end, you can get a decision result, which can be pushed back to what caused the result.

We also have the thermal radiation pipeline monitoring of the production line, and there are some fuel transmission pipes in the production line, as well as some special liquid transmission pipelines. We need to monitor in real time. The liquid passes through a temperature. If the temperature is too high, it is easy to cause the rapid aging of the production line pipeline. After aging, some will burst or crack and leak out. We use infrared to obtain a real-time image, and then segment the image. After that, a label is made, such as a pipe in blue that represents the direction of the number, and green is a horizontal pipe. After marking, the curve of the labeled pipeline is monitored to get the working monitoring chart on the right, through which you can know the real-time temperature of the pipeline and the relevant monitoring values. This can continue to help manufacturers to see the relevant situation from the historical system diagram.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.