Remove input barrier in the way of deep learning in signal processing
Temporal Pyramid Pooling
Jiduo Zhang1*, Robert Heinemann1, Otto Jan Bakker1, Menghui Zhu1
1 School of Engineering, The University of Manchester
Divide and pool
Constant-shaped output. As shown in the figure, three pool numbers (1, 2 and 4) are employed, which produces regions consisting of one vector with shape (N, L, C), two vectors with shape (N, L/2, C) and four vectors with shape (N, L/4, C). Then max pooling operations are conducted in these vectors, and the resulting maximum values are concentrated into a one-dimensioned vector shaped as (N, (1+2+4)*C) , which produces a constant shaped output for an input of any length.
A perfect match
Incredible speed and minimum size. With the utilisation of convolutional neural network (CNN), the TPP-CNN model could recognise both fine and macro structures of the signal with shared weights and minimum parameters, deployable for embleed devices and their integrated system.
TPP
CNN
Accuracy
99%
In identify incidence.
Accomodatable time
∞
Input in any size permitted.2
Up to
50
Identification times in 1 second.3
One for all. The proposed method allows deep learning model to process signal in any length, which could archieve both high accuracy and immediate response at one model.
Wide-spectrum and reusable. Compared with the traditional model, the proposed method could accurately identify process incidences for different combination of frequency and duration once it get well trained, thereby saving the time in rebuilding, retraining and retesting the models .
2: Subject to the hardware and software environment, excessive sized input will generate gaint feature maps and therefore exhaust computation resources.
3: The test was conducted on the environment of NVIDIA RTX 3080 with CUDA 11, and sampling duration is 0.1s with the frequency of 1KHz.
Given that the proposed method could be applied to any signal processing task, high-accuracy and immediate monitoring by deep learning approach could be achieved in a wide range of applications. Furthermore, we hope that this method could be a potential solution to the problem of general tool condition monitoring in the manufacturing industry.
The paper entitled "In-process tool incidence identification based on temporal pyramid pooling and convolutional neural network" has been accepted by 17th CIRP Conference on Intelligent Computation in Manufacturing Engineering.
Jiduo would like to acknowledge the continuing contribution to this research by China scholarship council.
If you have any questions or suggestions, please feel free to contact us.