With the development of electronic technology, neural network and artificial intelligence processing technology, new research methods of magnetic flux leakage signal processing are being carried out at home and abroad. Since the traditional method is seriously affected by human factors, it is easy to produce false detection and false detection, which greatly affects the detection accuracy. Therefore, an intelligent processing method for the defect signal is particularly needed. Multi-sensor data fusion technology is an integrated information processing technology that has been formed and developed in recent years. It makes full use of the complementarity of multi-source information and the high-speed computing power of computers to improve the quality of the resulting information. Multi-sensor data fusion enables multi-level, multi-faceted, multi-level processing of data from multiple sensors, resulting in new and meaningful information that is not available to any single sensor, and it is also The error caused by the failure of a single type of sensor can be effectively eliminated, and multi-sensor data fusion has received extensive attention in the field of signal processing [1]. In this paper, the wavelet threshold denoising algorithm is used to highlight the signal defect characteristics, and the processed signal is sent to the RBF neural network data fusion center. The information redundancy between the two types of sensors is used to improve the detection accuracy and eliminate the uncertainty in the measurement. , to obtain more accurate and reliable measurement results.
Data fusion sensor array
Multi-sensor data fusion requires a variety of sensors at various stages of the application process (eg, model building, feature extraction, target recognition, etc.). Since no sensor of any kind is absolutely better than other types of sensors, in a system, multiple types of sensors need to be used simultaneously to improve system detection, identification, classification and decision making. In this paper, two types of sensors are used according to the needs of the system: electromagnetic induction sensors and Hall sensors, which use the redundant information generated between them to detect the defect signal.
According to the characteristics of the defect signal and the environmental requirements, the magnetic flux leakage sensor array is used for data acquisition [2]. In order to improve the detection sensitivity and reduce the surface noise and temperature effects of the steel pipe, a sensor array consisting of 32 sensors is applied to the wear block that matches the surface of the steel pipe to form a probe. The sensors are divided into two groups of 16 each. One group consists of 16 electromagnetic induction sensors, and the other consists of 16 Hall sensors, which are alternately distributed on the probe surface. The two sets of 32 signals transmitted from the sensor array are preprocessed and sent to the fusion center for data fusion.
Electromagnetic induction sensor
The working principle of the electromagnetic induction sensor is that when it is scanned against the surface of the steel pipe, the leakage magnetic field generated by the defect of the steel pipe causes a change in the magnetic flux passing through the coil, thereby causing an induced electromotive force in the electromagnetic coil to form a defect signal. When the electromagnetic coil for detection is moved relative to the steel pipe, the induced electromotive force Uc generated by the coil detecting the leakage magnetic field is:
Where n is the number of turns of the coil, φ is the magnetic flux of the leakage magnetic field passing through the coil; B is the magnetic flux density of the leakage magnetic field; S is the cross-sectional area of ​​the coil, and t is the coil motion time. The electromagnetic induction sensor can be used in a wide temperature range, and has a long working life, strong resistance to dust, water and oil, and can withstand various environmental conditions and external noise.
Hall sensor
The Hall sensor detects the magnetic flux leakage signal. When the current I passes in the vertical direction with the magnetic field B, the Hall potential Hr is generated on both sides of the Hall sensor perpendicular to the current and the magnetic field:
In the formula, RH is the Hall coefficient; KH is the ratio of the Hall coefficient RH to the thickness t of the Hall sensor, and is called Hall element sensitivity. When the Hall coefficient RH and the current are constant, the Hall potential Hr depends only on the intensity of the magnetic field B and is independent of the moving speed of the leakage magnetic field. Therefore, the Hall sensor is not affected by the non-uniformity of the pipeline detection.
Signal preprocessing
For the magnetic flux leakage defect signal data measured by multi-sensor, in order to ensure the accuracy of the test, after the system obtains the signal, the signal must be pre-processed to filter out various external disturbances and various noises to obtain the correct measured coarse value. Generally, there are two methods: one is to perform smoothing, the actual algorithm can be realized by sliding the median smoother; the second is to eliminate the coarse error, and the data correlation culling method and the signal smoothing method can be simultaneously performed, and the correlation from the multi-sensor is related. Data fusion with complementary, redundant data. This method can make full use of the information of the measured object in time and space, and accurately describe the measured [4]. Therefore, the results of multi-sensor fusion are more accurate than the measurements of a single sensor.
In this paper, the wavelet denoising algorithm is used for the magnetic flux leakage defect signal. This method first denoises the measured value of each sensor with the wavelet threshold to reduce the influence of noise on the measured value of the sensor. In order to better reconstruct the sensor signal, each sensor measurement can be normalized and then sent to the data fusion center of the RBF neural network for fusion. Wavelet analysis of the defect signal can highlight the defect point. After wavelet analysis, the local modulus maximum value of the defect signal and its position and the waveform characteristics of the magnetic flux leakage signal can be used as feature information to distinguish different defects.
Signal model
In the process of magnetic flux leakage signal acquisition, it is assumed that there are N sensors measuring different positions of the same defect, and the measured value of the magnetic flux leakage signal obtained by each sensor is recorded as Xj (j=1, 2, 3. . . . N). There are internal and external noise effects during the measurement. The measured values ​​can be expressed as:
Where S(n) is the true measured value, ej(n)(j=1,2,3..N) is the additive noise of the jth sensor at time n, and Xj(n) is the first The actual measured value of j sensors at time n. Since each sensor is subject to different levels of noise interference, the extent to which the actual measured value deviates from the actual measured value is also different.
Wavelet threshold denoising
The wavelet denoising method is mainly realized by setting the threshold value. For the discrete wavelet transform of the magnetic flux leakage signal, all wavelet coefficients are calculated, the wavelet coefficients considered to be related to noise are removed, and then the signal is obtained by the inverse transform of the wavelet transform. There are many ways to choose a threshold for a given signal. This article uses a method based on Stein's Unbiased Risk Estimation (SURE):
The threshold t in the equation is obtained as a likelihood function, and then the likelihood function is minimized to obtain the desired threshold. In the VISU method, the threshold selection is fixed, while in the SURE method, the threshold is adaptively changed, which can better reduce the influence of noise on the defect signal [4]. According to the waveform of the magnetic flux leakage signal after wavelet processing, the feature vector of the defect can be extracted and used as the input of the neural network fusion center. Data fusion can be divided into pixel layer fusion, feature layer fusion and decision layer fusion according to the level of fusion and actual content. In this paper, the wavelet-reduced noise signal is subjected to feature level fusion through the RBF neural network fusion center to quantitatively analyze the signal.
Neural network fusion algorithm
Common data fusion methods include neural networks, clustering algorithms, or template methods. Among them, the artificial neural network has excellent functions such as learning, memory, association, fault tolerance and parallel processing. It has flexible application in topology and weight adaptive, and has been widely used in calibration and fault diagnosis of measurement and test instruments. In multi-sensor measurement systems, the use of data fusion technology can bring many benefits to the system [5], such as enhancing system stability, increasing system credibility and improving system detection capabilities. Because BP neural network has slow convergence speed, long network training time and local minimum value, RBF neural network has faster learning characteristics and stronger approximation ability than BP neural network. Therefore, the system uses RBF neural network as the fusion center of the feature layer fuse for fusion training.
Neural network structure
The neural network structure diagram is shown in Figure 2. The first layer is the input layer, which is used for feature information fusion information collection to form the input sample space X. The second layer is the hidden layer, which is used to map the input sample space to high-dimensional. Radial basis function space, that is, feature extraction of the input information space X. The hidden layer node parameter vector includes the central value Ci and the standard deviation δi; the third layer is the output layer, ωi is the connection weight of the i-th basis function and the output node, the output is Y, and the radial basis function selects the Gaussian function, as follows :
Where m is the number of hidden layer nodes, ‖? It is the Euclidean norm.
The neural network learning process is divided into two phases: in the first phase, the radial basis function and its parameters are determined according to all input samples, that is, the central value and standard deviation of the Gaussian function of each node of the hidden layer are determined; the second phase, After determining the parameters of the hidden layer, the weight of the output layer is obtained by using the gradient descent algorithm according to the sample.
The neural network algorithm mainly adjusts the connection weights so that the output layer and the expected output gradually become more consistent. According to the principle of minimum mean square error (MSE), when the error indicator is within a certain range, the operation can be stopped, indicating that the network training is successful.
Neural RBF network parameter selection and weight update
In the fusion center, the structural parameters of the RBF neural network have the number of hidden layer nodes, the center value and the standard deviation of the radial basis function. The more nodes, the stronger the learning ability, and the reasonable number of nodes can be optimally trained. The central value and standard deviation can be determined by a simple and effective clustering algorithm K-means clustering algorithm. The method has the advantages of simple implementation, small calculation amount, strong anti-noise ability and high recognition rate, and can well solve the problem of unreasonable distribution of modeling samples. The size of the standard deviation affects the response of the radial basis function to the input. If the standard deviation is too small, the basis function can only respond to a small area near the input data. If the standard deviation is too large, the inherent local information may be lost, and the model accuracy is also Poor, so the standard deviation should be selected within a stable interval.
In this paper, the gradient descent algorithm is used to determine the connection weight. Assume that the total error is:
Where p(xj) is the expected output of the jth training sample; y(xj) is the actual output of the network, and n is the total number of training samples.
In the formula, Yi(xj) is the output of the i-th basis function of the hidden layer; it is the updated value of the connection weight ωi; η is the learning step size, and is generally selected between 0.2 and 0.9 [7].
Experimental simulation analysis
In this paper, MATLAB software is used to carry out simulation experiments. Through the provided 40 sets of artificial crack samples, the RBF neural network system is trained and pattern recognition. The input data is preprocessed in the early stage of learning, and the best is generated through sample training. RBF network [8]. Another 10 sets of test sample data were taken as input, simulated by MATLAB software, and evaluated by RBF neural network and compared with conventional BP neural network.
The simulation results show that the RBF neural network fusion can detect the depth of signal defects more accurately. The RBF neural network is superior to the conventional BP neural network in terms of learning ability and detection accuracy. The average absolute error of this method is generally 2.69%. The average absolute error of the conventional BP neural network is 5.47%. It can be seen that the former has a significantly better detection effect than the latter.
In the magnetic flux leakage defect signal, the wavelet noise reduction preprocessing can effectively reduce the influence of noise while retaining the feature information of the defect to the greatest extent. The magnetic flux leakage sensor array overcomes the detection of the original single sensor system to some extent. Error; data fusion processing based on RBF neural network for magnetic flux leakage signal can be used for quantitative analysis of crack depth detection. The experimental results show that the RBF neural network not only has a fast learning speed, but also can effectively improve the accuracy and accuracy of detection, eliminate the uncertainty of information, and improve the reliability of the sensor.
Power cables are used to transmit and distribute electric energy. Power cables are often used in urban underground power grids, internal power supply of enterprises, outgoing lines of power stations, internal power supply of industrial and mining enterprises, and underwater power transmission lines across rivers and seas.
In power lines, the proportion of cables is gradually increasing. Power cable is a cable product used to transmit and distribute high-power electric energy in the trunk line of power system.
The structure of power cable is relatively simple, which can be divided into conductor, internal and external insulation layer, sheath, armor layer and protective layer. The conductor is mainly made of copper and aluminum, which is used to conduct current. The insulating layer is usually wrapped around the conductor to isolate the conductor and so on. There are various types of power cables. If they are classified according to voltage, they can be divided into low-voltage, medium voltage, high-voltage, ultra-high voltage and ultra-high voltage cables.
Advantages
1. Small land occupation. Generally, it is buried in the soil or laid indoors, in ditches and tunnels. The insulation distance between lines is small, without poles and towers. It occupies less land and basically does not occupy space on the ground.2. High reliability. It is less affected by climatic conditions and surrounding environment, with stable transmission performance and high reliability.
3.It has more favorable conditions for the development of ultra-high voltage and large capacity, such as low temperature and superconducting power cables.
4.Less maintenance work.
5 The possibility of electric shock is low.
Power Cable,High Voltage Cu Power Cable,Low-Votage Power Cable,Pvc Power Cable
Ruitian Cable CO.,LTD. , https://www.hbruitiancable.com