Cable Temperature Alarm Threshold Setting Method Based on Convolutional Neural Network

Power cable is used more and more in the power network, and its significance to the safety and stability of the power network is increasingly prominent. Especially in the urban power grid, the high voltage cable is related to the normal production and life of the city. Because of the particularity of the laying environment, it is very difficult to find and eliminate the fault points once the cable faults occur, which seriously affects the reliability of the power grid. Currently, 25% of cable faults are caused by elevated cable temperature, so it is important to set the cable temperature alarm threshold accurately. In this paper, a method of setting temperature alarm threshold using convolutional neural network is proposed. Experiments show that this method is 60% more accurate than other methods.


Introduction
The current cable is the use of optical fiber for temperature measurement. The technique, proposed by the University of Southampton in 1981, is based on the idea of sending a pulse of light into a fiber in which each individual point backscattering a small amount of light, including Stokes and anti-Stokes. Among them, Stokes light has nothing to do with temperature, while the intensity of anti-Stokes light changes with temperature. The temperature can be calculated from the quantitative relationship between the ratio of anti-Stokes light to Stokes light and temperature, so as to implement the temperature distribution measurement of each point along the optical fiber. This technology still uses temperature as the monitoring quantity, and the basic principle is the same as the cable online monitoring technology based on digital temperature sensor. The effect is not fundamentally improved, and the implementation cost of this technology is high.
This paper presents an on-line cable monitoring and fire prediction system based on Convolutional Neural Network[1,2].The sensor is directly installed on the surface of the cable, the cable surface temperature and current are collected online, the core temperature is calculated, the temperature upper limit is set to alarm, and the fault location is carried out.

Convolutional Neural Network Model for Temperature Threshold Recognition
A large number of facts have proved that convolutional neural network is very effective in temperature target detection and recognition, but there is a contradiction between network performance and network parameters in the process of detection and recognition. As a typical convolutional neural network, LeNET-5 has a high accuracy in the application of temperature recognition and is widely used in threshold judgment [3,4]. However, the physical meanings of temperature samples and experimental parameters in DEM are quite different, and the effect of completely applying LeNET-5  [5,6]. Therefore, combined with the temperature characteristics of the cable and the point group constraint relationship between the cable and other surrounding elements, the CNN network model was modified with LeNET-5 basic network as the framework [7].
As much as possible in order to fully reflect the characteristics of the cable temperature of local and global features, part of the sample layer the Lenet-5 network of transformation as shown in Fig.1. The temperature of the cable under the global features of sampling and sampling deep local features, and then to the extraction of the feature fusion, solve the problem of incomplete temperature element feature extraction. Finally, the temperature elements are identified according to the process in Fig.2.
1. Cable conductor current. The cable current refers to the current through which a cable line transmits electric energy. Under the condition of thermal stability, the cable current carrying capacity is called the cable long-term permissible current carrying capacity when the conductor reaches the long-term allowable operating temperature. The current left over the cable can be measured with a current clamp. The function of a conductor is to conduct current, so the more current it has, the more heat it gives. The amount of heat generated is proportional to the square of the current.
2. Historical mean temperature. The historical temperature can be measured by a temperature sensor. Temperature sensors are placed on the cable surface to measure the cable surface temperature in real time. The heat from the conductor of the cable is transferred to the surface of the cable through the cable insulation layer and is detected by a temperature sensor. Through the accumulation of historical data, the historical average value of cables at a certain time can be obtained.
3. Ambient temperature. The ambient temperature can be obtained by combining the local weather forecast and the temperature around the cable. Cable laying methods include direct buried laying, drainage laying and tunnel laying. The above laying methods have an effect on the surface temperature of the cable. Therefore, the ambient temperature is also an indispensable factor of the cable temperature threshold.
4. Relative humidity. Relative humidity denoted by Rh. Represents the ratio of the absolute humidity in the air to the absolute humidity at saturation at the same temperature and pressure, as a percentage. The ratio of the mass of water vapor contained in wet air to the mass of water vapor contained in saturated air at the same temperature and pressure, expressed as a percentage.
5. Infrared temperature measurement picture. It is gotten by infrared thermal imager. Infrared thermal imager is a kind of equipment that converts the temperature distribution image of the target object into visual image by infrared radiation detection of the target object, signal processing, photoelectric conversion and other means using infrared thermal imaging technology. The infrared thermal imager accurately quantifies the actual detected heat and images the whole subject in the form of a surface in real time, so it can accurately identify the suspected fault area that is heating.

Figure 1. Lenet-5 network of transformation
In the pre-training stage, LeNET-5 neural network structure is used as the basic framework to initialize its weight parameters, retain the weight and bias parameters except Softmax layer, and reverse fine-tune it with input sample data to calculate the difference between the detection  In the new CNN network model, the positive sample DEM data is used as the input of the first layer of the neural network. The convolutional layer adopts 5* 5 convolutional kernel to learn the data features of temperature threshold, and uses weight sharing to obtain global features, such as temperature and humidity, which reduces the parameters of network learning and reduces the complexity of network [8,9]. Each convolution kernels convolution operation with the upper input characteristic figure, after the activation function of nonlinear mapping to the next layer of network, the iterative mapping methods such as type (1), one l for layer, k is the convolution kernel, b is the bias, X is the characteristics of the current layer to calculate figure, M for the upper figure, f is Relu nonlinear activation function.
In the basic CNN network model, as the number of network layers deepens, the dimension of the feature graph keeps decreasing, which will dilute the structural features of the temperature threshold continuously, thus affecting the feature detection performance of the local temperature.
Therefore, continuous down-sampling should not be used only in the sampling layer. Instead, features of different scales should be fused by down-sampling the shallow layer features S2 extracted from the network, up-sampling the deep layer features S4, and maintaining the original size of the middle layer features C3, so as to obtain both local temperature features and global features.
The improved convolutional neural network structure is shown in Figure 1. Among them, the pooling treatment is carried out by maxpooling in the subsampling, which greatly reduces the possible overfitting phenomenon of the network and speeds up the convergence speed of the network. Its expression is shown in Equation (2), where S represents the size of the pooling template and M and N respectively represent the step size in the corresponding direction.
In the up-sampling stage, the inverse process of maxpooling is used to carry out "up-pooling", that is, the maximum position information is retained in the maxpooling stage, and then the information is used to expand the feature map in the up-pooling stage. Except for the maximum position, all the others are supplemented with 0.
The convolution layer and the sampling layer constitute a feature extractor, which extracts the discriminant feature information about the saddle and excavates the feature of the region where the saddle is located. Full-connection layer 5 will integrate the learned saddle features, and through fullconnection layer 6, the features will be compressed for the final Softmax classifier layer.
Target recognition can only be realized with the help of a certain classifier whether it is based on supervised strategy or unsupervised strategy. Softmax regression model, which is widely used in deep learning algorithms, is selected for the identification of a large number of temperature maps. Softmax is used in the classification process. It maps the output of multiple neurons to the interval (0, 1), and