MT method and deep learning
MT methods suitable for manufacturing sites
It is said that MT method is good at tasks related to manufacturing, such as surveillance and inspection problems, while deep learning is good at tasks such as image and language processing.
The reasons why MT method is suitable for manufacturing are "high sensitivity in detecting anomalies" and "ease of understanding for humans. As the feature of Deep Learning and MT method becomes clearer, the importance of using them for different purposes is becoming more recognized.
The following is an explanation of the differences between the two in terms of structure and characteristics.
The MT method has a network structure as shown in the figure, where the circles are items and the lines mean correlations between items. And the MT method uses this structure to identify differences in patterns. Correlation is the magnitude of the relationship between two items. For example, there is a large correlation between the degree of opening of a water tap and the amount of water. Correlation can be found between all items, large or small.
This correlation network is obtained from only normal data. Determining the network structure only from normal data is a major feature of the MT method. This network is called the UNIT SPACE.
Next, it is calculated that the distance from this normal correlation network to the target. Then, if the distance is close, we can determine that it is normal, and if it is far, it is abnormal.
Mahalanobis distance, a statistical mathematics, is used as a means of calculating the distance. If a distance is greater than 4, it is considered abnormal in general. In other words, if the distance is more than 4, the probability of being a member of the normal group becomes quite small.
Structure of MT method
Distance to the target
Structure of Deep Learning
( in the case of auto encoder )
Deep learning has a structure consisting of an input layer, an output layer, and hidden layers. The figure is an example of a deep learning structure, called an autoencoder. There are 14 circles at the left and right ends, respectively, and three hidden layers. The circles represent human brain cells and the lines represent the strength of the connections between the cells. The input and output layers are fed with known data called "teaching data", and after thousands of thousands of iterations of learning, the state of the cells and the strength of the connections are determined.
In the figure, the number of cells in the input and output layers is the same, and the hidden layer is symmetrical. In this structure, only normal data is trained as in the MT method. This structure is called an autoencoder because the input data is compressed once in the hidden layer.
When a pattern that is similar to the teaching data is input, the error that appears in the output layer is small. On the other hand, when a different pattern is input, the error becomes larger. By setting a threshold value for the magnitude of the error, the system can determine whether the pattern is normal or abnormal.
In the deep learning, the number of hidden layers and the number of cells can be set arbitrarily by the user. In other words, the number of hidden layers can be set to 7 or 9, and the number of cells is also arbitrary. Therefore, a variety of structures are possible.
Differences in characteristics
The MT method learns only one type of state (normal state).
In the field of manufacturing, i.e., plant monitoring and product inspection, most of them are normal. But, abnormalities vary and can be unknown. Therefore, it is impossible to completely cover all abnormal conditions. Therefore, the MT method is very suitable for learning only the normal state and issuing an alarm only when something "not normal" occurs.
The MT method can also diagnose the cause of an abnormal result. It clearly shows which item is different from the usual one, or if the balance of multiple items was different. This is why it is called white-box AI. Based on the diagnosis results, engineers can take the following actions.
Comparison of characteristics between MT method and deep learning
MT method has high sensitivity to anomalous data and small error in teaching data
Comparison of cause diagnosis results (horizontal axis: item number)
The diagnostic results of the MT method are easy to understand, but difficult in deep learning.
Deep learning generally learns a large number of patterns in a single network. Therefore, it is powerful for classification problems such as images, text, and language.
However, in deep learning, it is possible to train only normal data as in the MT method. This is the structure called "autoencoder" mentioned above. In this structure, when normal data is input, the error in the output value is small, and when the data is not normal, the error becomes large. Therefore, it is possible to judge whether the data is normal or abnormal based on the magnitude of the error.
Also, the error shown by each cell in the output layer is equivalent to the cause of the abnormality. However, it is often difficult for humans to understand the results. This is because the results of deep learning are a black box. We can't understand the reason why the deep learning came to that conclusion.
The characteristics of both can be summarized as follows.The characteristics of the MT method are; .The results of recognition are clearer and also the causes of abnormalities are easier for humans to understand.It can be sensitive to unknown abnormalities because it learns only from normal data.High speed and low memory load, suitable for edge computing The characteristics of deep learning are; The results are almost identical to MT methods, but the reasons for recognition are difficult for humans to understand. It is possible to learn multiple patterns in a single network. It can process huge amount of information if there are no problems with memory capacity and computation time.