A real-time monitoring and fault diagnosis method for underground mine electrical automation equipment combined with edge computing

Take the three-phase bridge controllable rectifier circuit of the electrical automation equipment as an example to collect the original data of the electrical automation equipment. First of all, the fault is categorized, according to past experience, it can be seen that in the main circuit of the three-phase bridge controllable rectifier device in electrical automation, it is rare that three and more than three thyristors fail at the same time, and the protection circuit makes the circuit short-circuit faults into short-circuit faults, so the main study of the circuit there is a short-circuit fault of one or two thyristors. Three-phase bridge controllable rectifier circuit faults in electrical automation can be divided into five types. The first type is the normal working condition, the second type is only one thyristor failure, the third type is two thyristor failure, and these two thyristors belong to the same phase power supply, the fourth situation is also two thyristor failure, but these two thyristors in the same half of the bridge, and the fifth condition is the failure of two crossed thyristors. Next, set the simulation parameters, set the trigger pulse, the pulse duty cycle is about 26%, the trigger angle can be selected 0 °, 30 °, 60 °, 90 °, 120 °, 180 °. Set the thyristor parameters, set the resistance to 0.15 Ω, the inductance to 1.1 μH, the buffer voltage to 0.11 μH, set the three-phase AC power supply, set the frequency of each phase voltage to 51 Hz, the phase angle to 120°, the sampling period to 100 μs, the sampling time is controlled at 0~51 ms. Finally, the sampling point is selected, and the voltage at the output terminal and the current at the input terminal contain rich fault information, and this fault information is very easy to be detected, so the output voltage at the load terminal and the current at the input terminal are selected as the sampling point. The output voltage and input current contain a wealth of fault information, and this fault information is very easy to detect. Therefore, the load output voltage and input current are selected as sampling points.

Edge computing nodes perform core data analysis duties in electrical automation equipment monitoring and fault diagnosis systems. These nodes are located close to the data source and are responsible for performing complex data analysis tasks for real-time fault diagnosis. At edge computing nodes, the data analysis process includes feature extraction, pattern recognition, and fault determination, utilizing advanced algorithms and computational techniques to optimize performance and accuracy.

First, feature extraction is achieved through signal processing techniques, and commonly used methods include Fast Fourier Transform, which converts time-domain signals into frequency-domain signals to reveal key features of equipment performance. The formula is equation (1): (1)X(f)=∫−∞∞x(t)⋅e−i2πftdt$$X(f) = \int_{ - \infty }^\infty x (t) \cdot {e^{ - i2\pi ft}}dt$$

where X(f) is the output in the frequency domain and x(t) is the input signal in the time domain.

Next, in order to identify the failure modes from the extracted features, the edge computing nodes utilize support vector machines. The support vector machine distinguishes different classes of data by constructing an optimal hyperplane. The process can be expressed as equation (2): (2)minw,b,ζ12wTw+C∑i=1mζi subjecttoyi(wTϕ(xi)+b)≥1−ζi,ζi≥0$$\begin{array}{c} {\min _{w,b,\zeta }}\frac{1}{2}{w^T}w + C\sum\limits_{i = 1}^m {{\zeta _i}} \\ subjectto{y_i}\left( {{w^T}\phi \left( {{x_i}} \right) + b} \right) \geq 1 - {\zeta _i},{\zeta _i} \geq 0 \\ \end{array}$$

where w and b are model parameters, x_i is a nonlinear mapping, and C and ζ_i are used to control model complexity.

Finally, in order to accurately determine faults and distinguish false positives and false negatives, the confusion matrix is used to evaluate the performance of the classification model. The formula for each element in the confusion matrix is equation (3): (3)Pij= Numberofsampleswithactual categoryiandpredictedcategoryjTotalnumberofsamplesincategoryi$${P_{ij}} = \frac{\begin{array}{l} Number{\text{ }}of{\text{ }}samples{\text{ }}with{\text{ }}actual \\ category{\text{ }}i{\text{ }}and{\text{ }}predicted{\text{ }}category{\text{ }}j \\ \end{array} }{{Total{\text{ }}number{\text{ }}of{\text{ }}samples{\text{ }}in{\text{ }}category{\text{ }}i}}$$

where P_ij denotes the probability of predicting a category j when the actual category is i.

Through these steps, the edge computing nodes are able to process and analyze large amounts of data in real time, providing fast and accurate fault diagnosis, which significantly improves the monitoring and maintenance efficiency of electrical automation equipment.

The limitation of unidirectional propagation makes the application of recurrent neural network not very wide. And in equipment fault diagnosis, the situation of the fault is different at different time periods, which requires the analysis of the state of each time node, and the antecedents and consequences of the equipment fault are analyzed in depth. With the rise of deep learning research and applications, researchers have proposed a method that applies LSTM neural networks to classification. The method automatically learns features and effectively models remotely relevant information. However, the model is more complex, with flaws in model training and long prediction times. The GRU neural network was created to solve the problem. Although GRU is not widely used today, the method inherits the advantages of the LSTM model, which can automatically learn features and efficiently model remotely relevant information. GRU has a significant speed advantage based on the fact that it has comparable segmentation performance with LSTM. In this chapter, a text classification method based on GRU neural network is proposed to classify the grid monitoring remote information according to the degree of its impact on the grid, and the classification results can be used as a reference for the grid staff to deal with different alarm information differently through the classification results, thus eliminating the waste of resources caused by manual classification.

GRU neurons contain only two gate structures: update gate and reset gate, and the structure of GRU network is shown in Fig. 2. For update gate z_t, a larger value indicates more information to be retained from the current neuron and less information to be retained from the previous neuron. For reset gate r_t, when the value of the equation is 0, it means that the data passed from the previous neuron is useless and can be discarded, i.e., only the input of the current neuron occurs as an input, so that some useless information passed from the previous neuron is considered to be discarded.

Figure 2.

GRU neuron structure

Eqs. (4)-(7) are the mathematical expression formulas for the neurons of the GRU neural network. (4)zt=σ(Wt⋅|ht−1,xt|)$${z_t} = \sigma \left( {{W_t} \cdot \left| {{h_{t - 1}},{x_t}} \right|} \right)$$ (5)ri=σ(Wt⋅|ht−1,xt|)$${r_i} = \sigma \left( {{W_t} \cdot \left| {{h_{t - 1}},{x_t}} \right|} \right)$$ (6)h=tanh(W⋅[rt*ht−1,xt])$$h = \tanh \left( {W \cdot \left[ {{r_t}^*{h_{t - 1}},{x_t}} \right]} \right)$$ (7)ht=(1−zt)*ht−1+zt*h$${h_t} = \left( {1 - {z_t}} \right)^*{h_{t - 1}} + {z_t}^*h$$

Where Z_t is the update gate, r_t is the reset gate, h_t−1 is the output of the previous neuron, x_t is the input of the current neuron, W_z is the weight of the update gate, W_r is the weight of the reset gate, σ is the sigmoid function, h_t is the output value of the current neuron, and h is the output value of the current neuron to be determined in the current neuron.

The inclusion of GRU neurons in the intelligent monitoring system designed in this paper makes it possible to analyze a sequence as it comes in, not only for the current time t state, but also in combination with the previous and subsequent states. By using the GRU recurrent neural network, the causes before and after the faults are correlated, not only limited to a single point, but the front is stretched out over a period of time, and thus it can be very competent in the duty of fault diagnosis. In this paper, the following equipment fault diagnosis model framework is constructed, as shown in Fig. 3.

Figure 3.

Fault diagnosis framework

As can be seen in the figure, the first input layer, its role is mainly responsible for processing the sequence data and performs the duties of preprocessing the data, labeling the failure modes, and transforming the variables. The second layer is the hidden layer, which contains the GRU neurons mentioned in the previous section and performs the linking back and forth, making it possible to accomplish the analysis and learning of the time series data. The third layer is the output layer, which outputs the learning output to the classifier and finally performs classification labeling and processing. The last layer is the decision-making layer, which serves to accept the final data after classification. And combined with the relevant specifications for result diagnosis and decision making.

To construct the test, the test object is selected as an electrical automation equipment, and the SIEEEVE-175 node test system is utilized to test the designed fault intelligent diagnosis technology. The electrical automation equipment fault parameters are used as the test basis for near-neighbor classification. The specific parameters of the electrical automation equipment node selected for this test are shown in Table 1.

Combined with the information in Table 1, 60 faults of electrical automation equipment are set up, and simulation tests are carried out using the traditional fault intelligent diagnosis technology as well as the fault intelligent diagnosis technology designed in this paper respectively, and the traditional fault intelligent diagnosis technology is set up as the test control group. In order to ensure the accuracy of the test results, the overall test is carried out in the same environment. The number of faults diagnosed at the 12 test points is compared with the actual number of faults, and the fault intelligent diagnosis technology with a smaller number of differences has a higher fault intelligent diagnosis accuracy.

According to the design of the test, the collection of 12 nodes in the control group and the test group in the fault data is shown in Table 2. it can be seen that this paper’s fault diagnosis model diagnosis of the number of faults are greater than 30, compared with the control group and the actual number of faults 60 is closer to the diagnostic accuracy of its higher.

Hyperparameters have an important impact on the performance, convergence speed, and generalization ability of the model in this paper. In practical applications, it is necessary to find the optimal combination of hyperparameters through experimental methods to improve the performance and accuracy of the model. In this study, best practices in the industry and previous engineering experience were followed to set the hyperparameters such as learning rate, convolution kernel size, and the number of iterations were selected through experiments. Figure 4 shows the variation in prediction accuracy between the training and sample sets, and Table 3 shows the hyperparameter combinations of the model.

Figure 4.

The effect of iteration number on model accuracy

Figure 4 shows how the prediction accuracy of the training and test sets changes as the number of model iterations increases. The results show that the accuracy of the model initially increases progressively with the increase in the number of iterations, which reflects the model’s effective learning of data features. After 225 cycles of iterations, the accuracy of the model on the training set tends to stabilize and there is no obvious sign of overfitting, at which time the accuracy of the training set reaches 93.11%, showing the model’s ability to fit the training data well. At the same time, the accuracy of the test set is about 85.62%, which is slightly lower than that of the training set, but still within the acceptable range, verifying the model’s good generalization ability on the unknown data. As can be seen from Table 3, the optimal input data pixel of the model is 96*96, the optimal convolution kernel is 5*5, the optimal number of convolution kernels is 256, and the optimal learning rate value is 0.001.

The dataset collected for this paper contains 400 samples, and the labels corresponding to these faults are shown in Figure 5. 35 partial discharge faults, which are early signs of deterioration or damage to the transformer’s internal insulation system; 75 low-energy discharge faults, which are usually associated with minor deterioration of the insulation material or localized defects; 130 high-energy discharge faults, which are often accompanied by strong energy release and may cause serious physical damage to the transformer; 90 low-heat faults, which are associated with abnormally high internal temperatures and may be caused by a variety of factors such as overloading and cooling system failure; 50 medium-heat faults; and 50 medium-heat faults. release, which may cause serious physical damage to the transformer; 90 samples of low-heat faults, which are related to the abnormal increase of internal temperature of the transformer, and may be caused by a variety of factors, such as overloading, cooling system failure, etc.; 50 samples of medium-heat faults and 20 samples of high-heat faults, medium-heat and high-heat faults are the common temperature anomalies in the operation of the transformer, and the degree of severity is increasing, which poses a threat to the long-term stable operation of the transformer. The gradual increase in severity is causing a threat to the long-term stable operation of transformers.

Figure 5.

Fault corresponding label

Figure 6 shows the values of the three evaluation metrics of this paper’s model. Combined with Fig. 6, it can be seen that the precision, recall, and accuracy of the GRU-based fault diagnosis model proposed in this paper to recognize transformer faults have the values of 0.899, 0.913, and 0.935, respectively, which indicates that the model in this paper has excellent performance in the field of transformer fault recognition, which can not only efficiently and accurately recognize the working state of transformers, but also make precise classification among the complex and changeable types of faults. It can also accurately classify complex and changing fault types, which can provide strong technical support for the stable operation of the power system.

Figure 6.

Textual Model prediction accuracy