Artificial intelligence based optical character recognition system and interference suppression in industrial internet of things network

Abstract

In factories, many measuring instruments are used to display, for instance, pressure, voltage, temperature, and humidity. Human errors are the main problem often occurring in many different processes mostly done manually, such as data acquisition. With manual data acquisition, the collected data cannot be used immediately in real-time. Thus, it is very important to think how we can obtain such data automatically and correctly in real-time. Most meters of measuring instruments use the seven-segment displays (SSDs) for digital meters and a needle for analog meters. We propose an automatic data acquisition system (DAS) using a low-cost camera for both digital and analog meters. In the proposed system, the key component behind is a numeral recognition system (NRS), which functions based on an optical character recognition (OCR) approach. The NRS embedded industrial IoT (IIoT) is used to serve a real-time service. The OCR applies the multi-layer perceptron (MLP) to efficiently recognize the numeral data. The proposed system also uses the Hough transform (HT) technique to detect the needle in the analog meters. Then, the longest straight line is found in the round analog meter. Along the lines of SSDs, digit numbers are recognized using the MLP. It is very common that the instruments' screens may confront the rotation problem. This is solved using the histogram of oriented gradients (HOG) and HT techniques. Salt and pepper (SP) noise, Gaussian noise, and Speckle noise are considered as they are typical noise types on images. The proposed system performs excellently in all situations achieving 95 percent accuracy with very low computation time up to only 1.86 second, which is suitable for a real-time service. When the situation of many such devices in a factory are considered, coexistence problems due to co - and adjacent channel interference are indispensable since they cause unnecessary contention as the devices are forced to defer transmissions until the medium is clear causing a loss of throughput. Adjacent channel interference is even more crucial causing corrupted data, which makes necessary retransmissions. We propose two interference recognition methods based on a long short-term memory network (LSTM) and a bi-directional LSTM (Bi-LSTM). They can be integrated into the IIoT network to mitigate the effect of interferences. We compare our proposed method to the traditional minimum mean square error (MMSE) and the MLP. In terms of the cumulative distribution functions (CDFs) of bit error rate (BER) the proposed method with the input data from a fast Fourier transform (FFT) algorithm outperforms the others since it is based on an LSTM and a Bi-LSTM which are suitable for the sequence type of the transmitting and receiving data. The Bi-LSTM outperforms the regular unidirectional LSTM due to the additional training capability passing the input data twice into the model, i.e., left-to-right and right-to-left. Finally, the coexistence problems are always indispensable, especially in such a typical factory, which contains a lot of machines and measuring devices.