Application of multi-sensor information fusion technology in improving the measurement accuracy of orifice plates

Preface

Multi- sensor information fusion technology is an important method in the current field of intelligent information processing. The so-called multi-sensor information fusion is an information processing process that automatically analyzes and synthesizes the complementary and redundant observation information in space or time obtained by multiple sensors according to a certain optimization principle. The information collection of a single sensor is insufficient and is easily affected by interference factors such as the surrounding environment. Therefore, it is difficult to ensure the accuracy and reliability of the detection information, which affects the correctness of the system decision. Therefore, the multi-sensor information fusion technology is adopted to make up for the defects of a single sensor by using the differences and complementarity of various sensors in performance, so as to obtain a more consistent explanation of the description system. Orifice plates are widely used in industrial production processes such as oil refining, chemical industry, storage and transportation, and natural gas due to their own advantages such as low price, simple principle, good reliability and easy maintenance. They are the main measuring instruments for detecting various gases and liquids in current and future industrial production. However, compared with other types of measuring instruments, orifice plates have larger measurement errors. The fundamental reason is that the actual characteristics of the measured medium at work do not match the characteristics set when the orifice plate is designed, and the temperature, pressure and density of the measured medium have changed to a certain extent. To this end, this paper proposes to use multi-sensor information fusion technology and a neural network fusion method to eliminate the influence of these factors on the orifice plate measurement accuracy. 1 System fusion structure When using an orifice plate for flow measurement, the actual flow to be measured is set to Q, and the output differential pressure of the orifice plate is p. During operation, the measurement results of the orifice plate are also affected by the deviations Δt, Δp, and Δρ between the temperature, pressure, and density of the measured medium and the set medium temperature, pressure, and density. The measured actual flow Q is actually a quaternion function, that is, Q = f(Δt, Δp, Δρ, p). Therefore, while using the orifice plate, temperature sensors, pressure sensors, and online density analyzers are used to obtain real-time characteristic information of the medium, thereby obtaining the deviation from the working characteristic information set when the orifice plate is designed; then, the various information obtained is pre-processed (including shaping, filtering, denoising, normalization, etc.) and sent to the fusion center, which uses a neural network method to fuse the data. The fused data concentrates the information of the four sensors, greatly improving the measurement accuracy of the orifice plate. The configuration structure of the system is shown in Figure 1.

Figure 1 System configuration structure

2 Algorithms and models of neural network fusion

There are many algorithms for data fusion, and the commonly used ones are Bayes decision theory, Kalman filtering, fuzzy fusion, neural network fusion, etc. Among them, artificial neural network is a large-scale information processing system formed by a large number of simple processing units called nodes or neurons connected to each other. It mainly imitates the human brain in terms of overall structure and function, rather than realistically reproducing details. It pays more attention to the information flow and movement mode in neural activities. Each neuron is an independent information processing unit, which performs independent calculations on the information it receives (rather than directly taking it out from memory), and then transmits the results. This distributed storage allows the system to restore the original information when part of it is damaged, so it has strong fault tolerance and associative memory characteristics; at the same time, because the neural network has the ability to process a large amount of data in real time, and the information processing is non-programmed, it can learn according to a certain external criterion, so the neural network has the characteristics of self-organization, self-learning, and self-adaptation, which makes the neural network widely used in information fusion.

2.1 Neural network fusion algorithm

For BP neural network, the most commonly used training algorithm is BP algorithm, which is actually a simple fast descent static optimization algorithm. When Ak is calculated, it is only corrected according to the negative gradient direction at time k, without taking into account the previously accumulated experience, that is, the gradient direction at previous moments, so it often causes the training learning process to oscillate and converge slowly. Here, the strategy of adaptive adjustment of the learning rate is adopted, and the formula of the improved algorithm is:

Ak＋1＝Ak＋CkXk
Ck＝2γCk－1
γ＝sign[XkXk－1]

In the formula, Ak＋1 is the network weight at the k＋1th iteration; Ak is the network weight at the kth iteration: Ck is the step size of the kth iteration; Ck－1 is the step size of the k－1th iteration; Xk is the negative gradient of the kth iteration; Xk－1 is the negative gradient of the k－1th iteration; γ is the step size adjustment coefficient.

When the gradient direction is the same for two consecutive iterations, it indicates that the descent is too slow, and the step size can be doubled; when the gradient direction is opposite for two consecutive iterations, it indicates that the descent is too fast, and the step size can be halved. The process of the algorithm is shown in Figure 2.

Figure 2 Improved neural network algorithm flow

2.2 Neural network model in orifice metering

Neural network consists of input layer, hidden layer and output layer. The characteristics of the whole network are determined by the connection weights of neurons between adjacent layers and the thresholds of neurons in the hidden layer. During the training process, the output differential pressure p of the orifice plate, the deviations of medium temperature, pressure and density from the set medium temperature, pressure and density Δt, Δp, Δρ are used as the input of the neural network; the output is the medium flow Q′, and its value will eventually approach the actual flow Q of the measured medium with a certain allowable deviation. The network structure is shown in Figure 3.

In Figure 3, the input layer has 4 input quantities, 20 neurons are set, the hidden layer has 40 neurons, the output layer has 1 output quantity, 10 neurons are set, and the network error E＝0.1. The learning rate adaptive adjustment algorithm is adopted, and through trial training, the mean square error between the output of the neural network, i.e. the measured flow fusion value Q′ and the actual flow Q of the measured medium, reaches the minimum value as soon as possible.

Figure 3 Model structure of neural network fusion in orifice plate measurement

3 Simulation experiment

During the production process of beer, due to the changes in its actual temperature, working pressure and bacterial concentration, the error caused by using the orifice plate to measure its flow rate is relatively large. The following is the historical data of the beer flow of a brewery to train the neural network. Some of the historical data of the brewery are shown in Table 1.

Table 1 Some historical data used for neural network training

After normalizing the data, a standard sample library of input and output of the neural network was established, and then the neural network was trained. The training simulation results show that the neural network reached stability after 2013 trainings. The trained neural network was then used to calculate other untrained experimental data. Table 2 shows some measurement results.

Table 2 Comparison of some orifice plate measured data and fused data 104·N·m3/h

It can be seen from Table 1 and Table 2 that due to the influence of temperature, pressure and density changes, the medium flow measured by the orifice plate has a large deviation from the actual flow. After the neural network multi-sensor fusion technology is adopted, the accuracy of the measurement results is greatly improved, which strongly proves the feasibility and practicality of the orifice plate measurement neural network model and its learning rate adaptive adjustment algorithm established in this paper.

4 Conclusion

When the temperature, working pressure and density of the measured medium change with the set values ​​when the orifice plate is designed, it has a considerable impact on the measurement results of the orifice plate, resulting in a large measurement error. The results of the simulation experiment show that the improved algorithm of the neural network adaptive adjustment of the learning rate is used to fuse the input information of multiple sensors, which is very effective in improving the measurement accuracy of the orifice plate. Therefore, this measurement technology and method has important practical significance.

References
[1] Li Renhou. Intelligent Control Theory and Methods [M]. Xi'an: Xi'an Jiaotong University Press, 1994, pp. 134-165
[2] Liu Junhua. Intelligent Sensor System [M]. Xi'an: Xi'an University of Electronic Science and Technology Press, 1999, pp. 377-406
[3] Liu Tongming. Data Fusion Technology and Its Application [M]. Beijing: National Defense Industry Press, 1998, pp. 36-22
[4] Wang Yongji, Tu Jian. Neural Network Control [M]. Beijing: Machinery Industry Press, 1998, pp. 63-102.