Sensor temperature error compensation based on neu

  • Detail

Sensor temperature error compensation based on neural network fusion

Abstract: the magnetic flux leakage detection device of submarine oil and gas transmission pipeline works under high temperature and high pressure, in which the InSb Hall sensor is sensitive to temperature and needs to compensate temperature error. In this paper, a multi-sensor fusion model is constructed. The outputs of multiple Hall sensors and temperature sensors are fused with radial basis function (RBF) neural networks, and the networks are trained with genetic algorithm. The laboratory test data and the inverse defect shape show that the neural network fusion method is simple and convenient for error compensation, and the average temperature sensitivity coefficient of Hall sensor output is reduced by two orders of magnitude

key words: temperature error compensation; Nerve collaterals; Data fusion; Magnetic flux leakage detection

temperature error compensation for sensors based on neural netw ork fusion

chen Tianlu, QUE Peiwen

(Institute of Automatic Detection under Shanghai Jiaotong Univers ity,Shanghai 200030,China)

Abstract:The equipment to inspect submarine o il & gas transportation pipelines often works under high temperature and high pr e InSb Hall sensors in the equipment which are sensitive to temperature need to be compensated. A multisensor fusion model is e test data of m ultiple magnet sensors and a temperature sensor are processed by a radial basis function(RBF) neural netic algorithm is chosen to train the e labtested data and the simulation shapes of defects show that the temperatur e error compensation made by neural network fusion is simple and e mean temperature sensitive coefficient is reduced sharply to less than 1/100.

Key words:temperature error compensation; neural network; data fu sion; MFL detection

at present, the total length of long-distance oil and gas transmission pipelines in service in China is about 20000 kilometers. Oil and gas leakage accidents caused by corrosion occur from time to time, causing economic losses and energy waste. Therefore, the detection of oil and gas pipelines plays a very important role in the national economy. Magnetic flux leakage detection method is the main technology used in the National 863 high-tech project "pipeline detection Crawler". The key component InSb Hall sensor is sensitive to temperature due to the inherent characteristics of semiconductor materials and defects in manufacturing process, so certain temperature compensation measures need to be taken [1]. Aiming at the successfully developed detection equipment, this paper adopts multi-sensor data fusion to eliminate temperature error. A multi-sensor fusion model is constructed, and the radial basis function (RBF) network is selected to fuse the output of magnetic sensor and temperature sensor. The effectiveness of this method is verified by experiments. The accuracy and stability of the detection system have been significantly improved

1 magnetic flux leakage detection device and sensor temperature characteristics

1.1 magnetic flux leakage detection principle and device

magnetic flux leakage detection method is an effective method widely used in the detection of oil and gas pipelines in recent years. The principle is shown in Figure 1. If there is no defect on the tested pipe wall, the magnetic line of force is closed; If there is a defect, the magnetic line of force will pass through the pipe wall and produce a leakage magnetic field [2]. The magnetic sensor converts the magnitude of the leakage magnetic field into voltage data output. The amplitude of the output waveform has an approximate linear relationship with the depth of the defect, and the horizontal distance between the peaks of the waveform and the width of the defect. Therefore, the shape of defects can be inverted from the magnetic flux leakage signal waveform [3]

the developed pipeline detection device consists of 6 parts: driving robot, system controller, power supply part, magnetic flux leakage detection part, data preprocessing part and positioning device. During the detection, the system controller controls the driving robot to drive all parts to crawl in the pipeline. The magnetic flux leakage detection sensor group obtains the pipeline state information and sends the detection data to the preprocessing part for amplification, denoising, compression and storage for offline analysis and inversion. The positioning device is used to determine the current position of the detection system

1.2insb temperature characteristic

the magnetic flux leakage detection sensor group selects InSb Hall element as the sensing element. Compared with other commonly used magnetic sensors, Hall elements are small in size, low in power consumption, vibration resistant, not afraid of pollution or corrosion of oil, water vapor, etc., and have high sensitivity. However, the output voltage error of Hall sensor is large when the detection device works in an environment with high temperature and pressure and frequent changes. Figure 2 shows the relationship between the output voltage and temperature change of Hall sensor composed of stable and reliable materials of different materials [1]. InSb has serious nonlinearity

multi-sensor fusion to achieve error compensation

multi-sensor information fusion is a new discipline rising in the 1970s, and has been widely used in target recognition, state estimation, threat estimation and other fields. This technology processes the data from multiple sensors at multiple levels, aspects and levels, so as to generate new and meaningful information, which cannot be obtained by any single sensor [4]

2.1 compensation model

this paper attempts to apply multi-sensor information fusion to error compensation. A temperature sensor is added to the magnetic flux leakage sensor of the detection device to record the temperature of the working environment in real time. The diameter of domestic oil and gas transmission pipelines is generally small, and the equipment has been arranged for 7 circles, with a total of 70 magnetic flux leakage sensors, which makes the space tight. Moreover, the components are sealed, and the internal ambient temperature changes slowly, so only one temperature sensor is added. The 10 rows (7 in each row) of magnetic flux leakage sensors distributed axially are fused with the output of the temperature sensor respectively to obtain the characteristic parameters and inversion graphics of the defects in each part of the pipeline. The model is shown in Figure 3. The information fusion strategy adopts RBF neural network, which fuses the data from 8 sensors

2.2rbf neural network principle

rbf network is a typical local approximation neural network. Unlike the global approximation neural network, it needs to adjust each input-output data pair and each weight, but adjust a few weights that affect the output, so that the local approximation network has obvious advantages in approximation ability and learning speed [5]

the RBF network structure is in the form of 8-20-1. The 8 nodes of the input layer only transmit input signals to the hidden layer, and the 20 elements of the hidden layer are transformed by radial basis function and output to the output layer. The output to avoid specimen slippage and fracture in the fixture inner layer node is just a simple linear function. The most commonly used radial basis function is Gaussian kernel function, as shown in equation (1)

where UJ is the output of J hidden layer nodes, x = (x1, X2,..., xn) t is the input sample, TJ is the central value of Gaussian function, σ J is the normalized constant, that is, the radial basis width, and M is the number of hidden layer nodes. The output range of the node is between 0 and 1, and the closer the input sample is to the center of the node, the larger the output value is

The output Yi of the

complex is a linear combination of the output UJ of the hidden layer node, as shown in equation (2)

2.3 training method

it can be seen from formula (1) that the parameters to be learned in this network include three categories: the center, width and connection weight of RBF. You can train separately or at the same time. When the number of hidden nodes is determined, genetic algorithm is used to train the TJ and width of the center at the same time σ J and the connection weight Wij between the hidden layer and the output layer

genetic algorithm is a computational model that simulates the process of biological evolution. It repeatedly uses selection, crossover and mutation operations for groups containing possible solutions, and constantly generates new groups, so that the population continues to evolve. When there are many input nodes, the algorithm has better global optimality and faster speed than the traditional BP algorithm [6]. The fitness function of the algorithm is shown in equation (3)

where n is the number of samples, M is the number of hidden layer nodes, B is the undetermined coefficient (generally take a larger value to ensure that the fitness is greater than zero), D is the expected output, and Y is the actual output of the network

the selection probability s (J) of individual J is shown in equation (4). Where FJ represents the fitness of individual J. S is the group size

in this paper, a single point crossover is used to combine the values of the corresponding crossover positions of the two gene strings to generate a new gene string

repeat the selection of crossover and mutation operations until the network meets the accuracy requirements

3 experiment

using a simplified undersea pipeline detection device (only a row of 7 magnetic flux leakage sensors and a temperature sensor are installed in the magnetic flux leakage detection component), under laboratory conditions, take 10 temperature points (- 10 ~ 80 ℃, one point per 10 ℃) to detect a half section pipeline respectively. The pipeline has the same material and pipe diameter (195 mm) as the actual submarine pipeline, and according to the requirements of American nondestructive testing standards, using EDM 2. The key points of the use of pressure shear testing machine, many defects of different sizes, shapes and types have been machined on its inner surface. At each temperature point, each MFL sensor takes 55 data to form 55 groups of data, 44 groups are taken as training sample data, and 11 groups are taken as test data. The above neural network method is used for training and testing to realize the fusion compensation of temperature error. Due to multi-dimensional input, the training speed is slow. When the network error is set to 0.001, the Gauss function training generally needs about 2300 steps to meet the requirements. It only takes 1700 steps to train with genetic algorithm

a standard rectangular defect with a length of 10 mm and a depth of 5 mm is detected. The network output at two typical temperature points is shown in Figure 4. The solid line in the figure is the expected output, the two dotted lines are the output without fusion at 70 ℃ and - 10 ℃ respectively, and the "○" line and "+" line are the output after fusion at these two temperature points. It can be seen from the figure that the two output waveforms before fusion deviate from the target curve, and there is a certain gap between the horizontal and vertical spacing of the peaks and troughs representing the defect characteristics and the target value. The output after fusion almost coincides with the expected value, and the compensation effect is remarkable. Figure 5 is the defect diagram after inversion. The meaning of linetype is the same as that in Figure 4 The defects of data fusion at 10 ℃ and 70 ℃ basically coincide with the actual defects. There is a deviation between the defect of unmelted data inversion and the actual defect depth

define the temperature sensitivity coefficient of the sensor α S is the average value of the maximum relative change in output voltage caused by temperature change of 1 K within the operating temperature range

where s ∈ (1, 2,..., 44) is the serial number of the measuring point, which represents different detection positions as the device crawls in the pipeline. T1 and T2 are the upper and lower limits of the working temperature, and U (T1) and U (T2) are the output values of the sensor when the temperature at s is T1 and T2 respectively. The test data is used to verify the sensitivity and adaptability of the network. The average temperature sensitivity coefficients of the seven sensors before and after fusion are 3.1 respectively × K-1 and 2.3 × K-1。 It can be seen that the temperature sensitivity coefficient of the sensor is reduced by two orders of magnitude after temperature compensation by multi-sensor fusion

4 conclusion

this paper applies the data fusion theory and neural network method to the error compensation of magnetic flux leakage sensor, which greatly improves the stability and accuracy of magnetic flux leakage detection sensor. After fusing the detection data of multiple MFL sensors and temperature sensors, the temperature sensitivity coefficient of the output value is reduced by two orders of magnitude, which provides a guarantee for the whole detection system to obtain high accuracy results in high temperature environment. When there are many input nodes, the speed of training network with genetic algorithm is faster than that with Gaussian function


[1] yuan Xiguang Sensor technical manual [M] Beijing: National Defense Industry Press.

[2] Jin Tao, que Peiwen Wavelet analysis of noise elimination experiment of magnetic flux leakage detection [J provides a creative and constructive environment for its engineers] Sensing technology

Copyright © 2011 JIN SHI