Fault prediction and maintenance of urban rail transit power supply system based on big data

With the continuous growth of urban population and the acceleration of urbanization, urban rail transit has become an indispensable and important means of transportation in modern cities. The normal operation of urban rail transit cannot be separated from the stable and reliable power system support [1-4]. Power system failure is one of the common problems in urban rail transit, and once the failure occurs, it will bring serious impact on urban transportation operation. Timely analysis and response to power system failure is important to ensure the safety and efficiency of urban rail transit [5-8].

Urban rail transit power system failure prediction is an important part of urban transportation operation and maintenance under the background of big data, through the real-time collection and analysis of rail transit system equipment operation data [9-10], potential faults can be found in advance, the possibility of predicting the occurrence of failures and the scope of influence, so as to provide a scientific basis for the maintenance, to achieve the failure of the active management and control [11-13]. Fault prediction not only helps to reduce the frequency of faults, reduce maintenance costs, but also effectively improve the reliability and safety of the rail transit system. Through the prediction to discover the pattern of faults and the reasons for their frequent occurrence [14-16], the possible occurrence of faults can be predicted, so that targeted measures can be taken to maintain the power system, in order to reduce the incidence of power system failures, and to ensure the safe and smooth operation of urban rail transit [17-19].

In this paper, we first study the equipment failure rate in different weather and use Monte Carlo simulation to diagnose faults in traction transformers in urban rail transit power supply systems. Ranging of faults in the traction power supply system is carried out, followed by predictive reasoning of the faults. Then, an example analysis is carried out to explore the prediction effect of this paper’s prediction method on contact network wear by comparing the relative error between the prediction results of this paper’s fault prediction method on the degree of contact network wear and the actual value. Three exogenous fault predictions are taken as examples to compare the performance of this paper’s method with other prediction models. Diagnose and predict reactor insulation faults using the method of this paper. Finally, feasible maintenance decisions are proposed.

2

Fault prediction system for urban rail transit power supply systems

2.1

Traction transformer fault diagnosis based on Monte Carlo simulation

2.1.1

Equipment failure rates in different weather conditions

Failures such as rejections and malfunctions of relays and circuit breakers lead to uncertainty in their relay protection history statistics. Under different weather conditions, the rejection and misoperation rates of relay protection devices and circuit breakers are different, and the overall situation shows that the higher the failure rate of the equipment under the worse weather is, the higher the failure rate. Therefore, this paper categorizes the weather states into: normal weather, bad weather and severe weather, and calculates the failure rate of each device under different weather states respectively, and reduces the uncertainty of information by introducing the failure rate into the Bayesian network parameter estimation. Relay protection devices and circuit breakers have action and inaction states (circuit breakers, i.e., trip and non-trip states), so their failure rates can be categorized into: false action rate and refusal rate.

1)

Calculation of the refusal rate of equipment in different weather conditions

Define λ_ref−x as the average refusal rate of the x device. λ_ref−x−i be the refusal rate of x equipment in i weather conditions. N is the number of refusals throughout the year. M is the number of correct actions throughout the year. N_i is the number of rejections in i weather conditions. M_i is the number of correct actions in i weather conditions. Therefore: (1) $λ_{r e f - x} = N / (M + N)$ \[{{\lambda }_{ref-x}}={N}/{\left( M+N \right)}\;\] (2) $λ_{r e f - x - i} = N_{i} / (M_{i} + N_{i})$ \[{{\lambda }_{ref-x-i}}={{{N}_{i}}}/{\left( {{M}_{i}}+{{N}_{i}} \right)}\;\] is obtained according to equation (2): (3) $N = \sum_{i = 1}^{3} N_{i} = \sum_{i = 1}^{3} λ_{r e f - x - i} (M_{i} + N_{i})$ \[N=\sum\limits_{i=1}^{3}{{{N}_{i}}}=\sum\limits_{i=1}^{3}{{{\lambda }_{ref-x-i}}}\left( {{M}_{i}}+{{N}_{i}} \right)\] is then obtained from Eqs. (1) and (3): (4) $λ_{r e f - x} = \frac{\sum_{i = 1}^{3} λ_{r e f - x - i} (M_{i} + N_{i})}{(M + N)} = \sum_{i = 1}^{3} λ_{r e f - x - i} P_{i}$ \[{{\lambda }_{ref-x}}=\frac{\sum\limits_{i=1}^{3}{{{\lambda }_{ref-x-i}}}\left( {{M}_{i}}+{{N}_{i}} \right)}{\left( M+N \right)}=\sum\limits_{i=1}^{3}{{{\lambda }_{ref-x-i}}}{{P}_{i}}\] Among them: (5) $P_{i} = \frac{M_{i} + N_{i}}{M + N}$ \[{{P}_{i}}=\frac{{{M}_{i}}+{{N}_{i}}}{M+N}\]

Then Eq. (4) is the refusal rate correction model under the weather three-state model.

The purpose of the weather three-state model is to calculate the refusal rate of the equipment in different weather states, while usually we can only get the average refusal rate through statistics. Definition F_i is the percentage of the number of rejections of the equipment in the i weather state to the number of rejections in the whole year: (6) $F_{i} = N_{i} / N$ \[{{F}_{i}}={{{N}_{i}}}/{N}\;\] can be obtained from Eqs. (1), (4) and (6): (7) $λ_{r e f - x - i} = λ_{r e f - x} F_{i} / P_{i}$ \[{{\lambda }_{ref-x-i}}={{\lambda }_{ref-x}}{{{F}_{i}}}/{{{P}_{i}}}\;\]

2)

Calculation of equipment false alarm rate under different weather conditions

Define λ_inc–x as the average false alarm rate of the x device. λ_inc−x–i be the misfire rate of the x device in i weather conditions. S is the number of false alarms for the whole year. T is the time of counting. S_i is the number of false alarms in the i weather state for the whole year. T_i is the duration of the i weather state: (8) $λ_{i n c - s} = S / T$ \[{{\lambda }_{inc-s}}={S}/{T}\;\] (9) $λ_{i n c - x - i} = S_{i} / T_{i}$ \[{{\lambda }_{inc-x-i}}={{{S}_{i}}}/{{{T}_{i}}}\;\] is obtained from Eq. (8) and Eq. (9): (10) $λ_{i n c - x} = \frac{\sum_{i = 1}^{3} λ_{1 n c - x - i} T_{i}}{T} = \sum_{i = 1}^{3} λ_{i n c - x - i} P_{t i}$ \[{{\lambda }_{inc-x}}=\frac{\sum\limits_{i=1}^{3}{{{\lambda }_{1nc-x-i}}}{{T}_{i}}}{T}=\sum\limits_{i=1}^{3}{{{\lambda }_{inc-x-i}}}{{P}_{ti}}\]

Among them: (11) $P_{i j} = T_{i} / T$ \[{{P}_{ij}}={{{T}_{i}}}/{T}\;\]

Then Eq. (10) is the error rate correction model under the weather three-state model.

As with the calculation of the refusal rate, we can usually only obtain the average false alarm rate. Therefore, define C_i as the percentage of the number of equipment malfunctions in the i weather state to the total number of malfunctions for the year: (12) $C_{i} = S_{i} / S$ \[{{C}_{i}}={{{S}_{i}}}/{S}\;\] is obtained from Eq. (8), Eq. (10) and Eq. (12): (13) $λ_{i n c - x - i} = λ_{i n c - x} C_{i} / P_{t i}$ \[{{\lambda }_{inc-x-i}}={{\lambda }_{inc-x}}{{{C}_{i}}}/{{{P}_{ti}}}\;\]

2.1.2

Monte Carlo-based fault simulation model for traction power supply

In this section, a traction power supply fault simulation model based on Monte Carlo simulation is developed in conjunction with a directed acyclic diagram based on relay protection configurations [20].

The traction transformer T fault simulation scenario is shown in Fig. 1. Take the traction transformer T in Fig. 1 as an example. Assuming that a fault occurs in the traction transformer T, the action state of its main protection Tm is randomly sampled according to the refusal rate of the protection device to generate a state value of 1 or 0, which represents the action and inaction of the protection device, respectively. If the main protection action, and then according to the circuit breaker refusal rate, through the same random sampling method, to generate the corresponding circuit breaker status value 1 or 0, respectively, on behalf of the corresponding circuit breaker tripped and not tripped.

If the main protection Tm refuses to operate, its state 1 or 0 is generated by random sampling based on the failure rate of the backup protection Tp. If its near backup protection operates, the corresponding breaker state 1 or 0 is generated again by random sampling based on the breaker refusal rate, and the process is shown in Fig. 2.

If both its main protection Tm and backup protection Tp refuse to act, or circuit breaker CB2 refuses to act, its status value is randomly generated according to the failure rate of the far backup protection Bm. If the far backup protection operates, the corresponding circuit breaker status value is then randomly generated according to the refusal rate of circuit breaker CB3, and the process is shown in Fig. 3.

The Bayesian nets of nodes or other components not involved in the simulation scenarios are similarly generated by random sampling of their state values. Component nodes have the relative failure probability of each component as their random sampling probability, and other nodes have the corresponding misbehavior rate as their random sampling probability.

2.2

Fault prediction and reasoning for traction power supply in urban rail transit

2.2.1

General fault ranging method for traction power supply system

1)

Comparison of current draw of the kiosk

Comparison of substation and zoning suction on the current, substation canonical in the 1AT section, zoning canonical in the 2AT section. In practice, the tripping near the AT because of the interval difference, the substation suction on the current and zoning suction on the current size does not have a significant regularity, resulting in the distance measurement device interval judgment error, which in turn causes a great distance measurement error. Therefore, the predicted distance of the ideal condition and the predicted distance in the actual condition need to be effectively corrected in order to reduce the error distance, first of all, we need to understand the principle of the realization of the method in the ideal condition, under the premise of the ideal condition, the suction-up current ratio method schematic diagram is shown in Figure 4.

The fault distance is calculated as Eq: (14) $D = L_{1} + X = L_{1} + \frac{I_{A T 2}}{I_{A T 1} + I_{A T 2}} L_{2}$ \[D={{L}_{1}}+X={{L}_{1}}+\frac{{{I}_{AT2}}}{{{I}_{AT1}}+{{I}_{AT2}}}{{L}_{2}}\]

Where, $\frac{I_{A T_{2}}}{I_{A T_{1}} + I_{A T 2}}$ $\frac{{{I}_{A{{T}_{2}}}}}{{{I}_{A{{T}_{1}}}}+{{I}_{AT2}}}$ is the leakage reactance.

2)

Up and down current size comparison method

Similar to the method of comparing the size of the current on the Ting suction, determine the difference between the size of the upstream and downstream currents on the T or F lines. If the difference between the up and down lines is small, the fault occurs in the 2AT section. As with the Sotinami current size comparison method, it is also possible to judge the wrong interval, which can lead to significant ranging errors. By comparing the current size of the up and down lines to determine the location of the fault point, the fault distance is calculated as: $X = \frac{I_{t 2}}{I_{t 1} + I_{t 2}} \times 2 (L_{1} + L_{2})$ \[X=\frac{{{I}_{t2}}}{{{I}_{t1}}+{{I}_{t2}}}\times 2({{L}_{1}}+{{L}_{2}})\].

3)

Reactance ranging method

Reactance ranging method is widely used in direct supply under the calculation of fault distance ranging method, the use of direct supply line reactance increasing principle, by calculating the size of the reactance during the fault, to determine the location of the fault, reactance ranging principle diagram shown in Figure 5.

4)

Cross-line current ratio method

Similar to the upstream and downstream current ratio method, the cross-connect current ratio is utilized to determine the fault point location. Its fault distance calculation formula is: (15) $L = \sum_{i = 0}^{n - 1} D_{i} + \frac{I_{H L (n + 1)}}{I_{H L n} + I_{H L (n + 1)}} \times D_{n}$ $L=\sum\limits_{i=0}^{n-1}{{{D}_{i}}}+\frac{{{I}_{HL(n+1)}}}{{{I}_{HLn}}+{{I}_{HL(n+1)}}}\times {{D}_{n}}$ (16) $I_{H L n} = | i_{T F n} |$ ${{I}_{HLn}}=\left| {{i}_{TFn}} \right|$ (17) $\frac{X}{D} = \frac{i_{F H 2} - i_{T H 2}}{i_{F H 1} - i_{T H 1} + i_{F H 2} - i_{T H 2}}$ $\frac{X}{D}=\frac{{{i}_{FH2}}-{{i}_{TH2}}}{{{i}_{FH1}}-{{i}_{TH1}}+{{i}_{FH2}}-{{i}_{TH2}}}$ (18 $X = | \frac{I_{F H 2} - I_{T H 2}}{I_{F H 1} - I_{T H 1} + I_{F H 2} - I_{T H 2}} | \times D$ $X=\left| \frac{{{I}_{FH2}}-{{I}_{TH2}}}{{{I}_{FH1}}-{{I}_{TH1}}+{{I}_{FH2}}-{{I}_{TH2}}} \right|\times D$ (19) $D = L_{1} + L_{2} + \dots + L_{n}$ $D={{L}_{1}}+{{L}_{2}}+\cdots +{{L}_{n}}$

For the fully parallel AT power supply mode, the cross-connection line current ratio method can basically adapt to all short-circuit faults and carry out ranging, and its principle does not involve the AT autotransformer leakage reactance and short-circuit transition resistance, so the ranging accuracy is extremely high, but the disadvantage is that it does not support customized correction.

2.2.2

Traction power supply fault reasoning

The traction power supply fault inference model is the core content of the traction power supply fault analysis and inference mechanism, and the inference model mainly consists of a knowledge base, inference machine, database, human-computer interaction interface, traction power supply fault knowledge base management system, and knowledge acquisition system, etc. [21-22]. After effective data fusion, the current fault situation of the traction power supply equipment is obtained, and the dynamic database saves the different fault data generated by the traction power supply system to the library in real time, and at the same time, the corresponding reminder warning state when these faults appear is used as a sample. In the process of reasoning about the traction power supply faults, the working condition of the traction power supply equipment will be reasoned and recognized first, to detect whether the traction power supply equipment is in a faulty state or not, and through the knowledge acquisition system of the traction power supply faults, the fault conditions of the equipment will be saved into the knowledge base as the knowledge, and according to the characteristics of the data of the faults of the traction power supply equipment, to construct a complete set of rule bases. The traction power supply fault inference model is shown in Figure 6.

During the internal operation of the traction power supply fault reasoning model, the reasoning principles and processes of the model are provided through an interpretation mechanism and are directly oriented to the human-computer interaction interface, which facilitates the experts to improve the knowledge base. Considering the real-time dynamics of the data in the dynamic database, the reasoner selects the knowledge base according to the reasoning strategy and finally displays the diagnosis results to the user. The human-computer interaction interface is embodied through the interaction interface, which is the window for users and experts to link the system. Through the human-computer interaction interface, the fault data of the traction power supply system is deposited into the library through the traction power supply fault knowledge acquisition system, the faults are diagnosed according to the knowledge base and rule base of the traction power supply faults, and at the same time, the feedback information will be transmitted to the knowledge base through the management system of the knowledge base of the traction power supply faults, and then the experts can amend and improve the system through the window. The expert can correct and improve the system through the window, and the user can understand the system fault problem through the window.

The traction power supply fault inference model can diagnose and analyze the faults and defects of the traction power supply system equipment, in order to guarantee the normal operation of the traction power supply system, it is necessary to repair the defects in advance, and to provide timely repair strategies for the faults that occur suddenly, through the real-time monitoring system to ensure that the faults can be found in the first time and to determine the location of the faults in order to ensure that timely remedial measures can be completed, through the real-time monitoring system. Remedial measures, through the traction power supply fault inference model for traction power supply system faults efficient solution to play a good preliminary groundwork, to ensure the effectiveness of the subsequent traction power supply system equipment fault detection.

3

Failure prediction analysis

3.1

Contact network wear prediction

This subsection takes high-speed railroad as an example, collects relevant data through research, takes the contact pressure, train speed, wire height, pull-out value, voltage, train frequency to obtain the mean and variance as inputs to the prediction model, and the average wear of contact network as outputs through MATLAB programming, thus carrying out the prediction of contact network wear. The gatbx toolbox is used in the MATLAB programming process. According to the determined neural network structure and the related function of genetic algorithm in the gatbx table, the relevant parameters are set, and the number of iterations is set to be 50, the number of populations is set to be 40, the crossover probability is set to be 0.6, and the probability of variance is set to be 0.01.

The contact network wear prediction is carried out using 15 sets of test samples, and the prediction results are shown in Fig. 7, with Fig. (a) and Fig. (b) showing the training set results and test set results, respectively. From Fig. 7, it can be seen that the training model R=0.94868 and 0.90857 obtained from the training set and test set are close to 1, which indicates that the model has a good fit and it is feasible to predict the contact network abrasion.

The predicted value of the prediction model is compared with the real value of the contact network wear, and the comparison results are shown in Table 1. By analyzing the error between the real value of contact network wear and the predicted value in Table 1, it can be obtained that the relative error is within 5%, which has a high degree of feasibility. It shows that the method of predicting contact network wear by the fault prediction model of urban rail transit power supply system in this paper is effective, and its accuracy will be improved with the improvement of data volume.

Table 1.

Error comparison analysis

Sample point	Prediction value	Real value	Relative error (%)
1	2.0678	2.0837	0.76
2	1.0866	1.0907	0.38
3	2.0401	2.0661	1.26
4	1.5654	1.5152	3.31
5	1.2444	1.2174	2.22
6	1.9858	2.0198	1.68
7	2.6681	2.6494	0.71
8	1.1717	1.1586	1.13
9	2.9034	2.8231	2.84
10	1.6201	1.6435	1.42
11	2.7024	2.7277	0.93
12	2.1865	2.2072	0.94
13	2.1349	2.2009	3.00
14	1.2467	1.1942	4.40
15	2.9372	2.9752	1.28

3.2

Exogenous Failure Prediction Performance

This subsection still takes the high-speed rail in rail transportation as an example, by analyzing the prediction effect of this paper’s fault prediction system on the faults caused by different external factors. Precision, Recall, F1 and Accuracy are used as evaluation indexes. Three types of faults caused by external causes, namely contact network lightning tripping faults, floating object intrusion contact network faults and contact network porcelain insulator haze fouling flash faults, are selected for three prediction performance experiments to compare this paper’s model with other prediction models.

3.2.1

Prediction of lightning tripping faults on contact networks

In Experiment 1, this paper’s method is compared and analyzed with other four prediction classification methods (Long and Short-term Memory Network, Support Vector Machine, Convolutional Neural Network, and Elman Neural Network) for the prediction of lightning trip faults in contact networks, and the experimental results are shown in Table 2.

Table 2.

Comparison of experiment results

Method	Precision	Recall	F1	Accuracy
LSTM	90.25%	99.55%	96.73%	91.66%
SVM	92.52%	97.98%	96.94%	92.32%
CNN	90.25%	99.55%	96.73%	91.66%
Elman	93.41%	97.46%	97.14%	92.44%
Ours	96.74%	96.48%	97.64%	94.59%

As shown in Table 2, the prediction methods in this paper have the highest scores in Precision, F1 and Accuracy indicators, which are 96.74%, 97.64% and 94.59%, respectively. A good predictive classification algorithm should be able to accurately predict the risk of contact network lightning strike tripping with high Precision, Recall and F1 values, i.e., a good predictive classification algorithm prevents the prediction of normal samples into faulty samples, and reduces unnecessary losses, such as spare parts acquisition and scheduling. From the comparative experimental results, it is clear that the method proposed in this paper is a better prediction method for contact network lightning strike tripping.

The Accuracy index of Long Short Term Memory Network and Convolutional Neural Network is 91.66%, while the Accuracy index of Support Vector Regression Machine and Elman Neural Network is slightly higher than 91.66%. The support vector regression machine and Elman neural network can improve the prediction performance of contact network lightning strike tripping, but the prediction performance improvement is limited, and only improves the Accuracy index to 92.32% and 92.44%, respectively. It can be seen that the long and short-term memory network, support vector regression machine, convolutional neural network and Elman neural network perform relatively conservatively in this study. The method proposed in this paper has better robustness on unbalanced datasets and has better predictive performance, ensuring Accuracy along with other predictive performance metrics.

3.2.2

Prediction of contact network faults due to floating object intrusion

In order to verify the effectiveness of this paper’s method on the prediction of faults of base float infested contact networks, a comparison experiment (Experiment 2) is carried out, and the comparison models include: random forest, neural network, support vector regression machine, Bayesian network, and long and short-term memory network.

The prediction method of this paper is compared and analyzed with other five prediction classification methods, and the experimental results are shown in Table 3. As can be seen from Table 3, compared with the other five prediction comparison models, the prediction method of this paper shows more superior prediction performance in the prediction of contact network faults infested by floating objects, with Precision, Recall, F1 and Accuracy values of 95.84%, 94.98%, 95.46% and 94.98%, respectively. Therefore, the prediction method in this paper has better robustness and adaptability to unbalanced data.

Table 3.

Comparison of experiment results

Method	Precision	Recall	F1	Accuracy
RF	91.15%	90.74%	90.23%	90.74%
Neural network	84.76%	85.27%	85.03%	85.27%
SVM	56.94%	72.11%	62.33%	72.11%
Bayesian network	90.03%	88.93%	89.84%	88.93%
LSTM	56.94%	72.11%	62.33%	72.11%
Ours	95.84%	94.98%	95.46%	94.98%

On the other hand, Support Vector Machines and Long and Short Term Memory Networks have smaller Precision values and larger Recall values, causing these two methods to easily predict medium and high risk fault test samples into low risk fault categories. Similarly, neural networks have the same problem of not being able to differentiate between medium and high floating objects intruding into the fault boundary of the contact network, resulting in neural networks predicting a large number of high-risk fault test samples as medium-risk fault categories. Although Bayesian networks and random forests show better prediction results, they are still inadequate compared to the prediction methods in this paper.

3.2.3

Haze Fouling Flash Fault Prediction

In order to verify the effectiveness of this paper’s method for the prediction of haze fouling flash on contact network porcelain insulators, Experiment 3: Prediction Performance Comparison Experiment was conducted. The comparison models used include: random forest, neural network, support vector regression machine and long and short-term memory network. The results of the comparison experiments between the prediction method of this paper and the other four prediction classification methods are shown in Table 4.

Table 4.

Comparison of experiment results

Method	Precision	Recall	F1	Accuracy
RF	99.52%	96.24%	98.83%	97.08%
Neural network	96.23%	99.96%	98.05%	96.43%
SVM	96.02%	99.96%	97.68%	95.77%
LSTM	96.02%	99.96%	97.68%	95.77%
Ours	99.87%	99.85%	99.76%	99.69%

As can be seen from Table 4, compared with the remaining four prediction models, the prediction method of this paper shows better prediction performance in the prediction of haze fouling flash of contact network porcelain insulators, with Precision, Recall, F1 and Accuracy values of 99.87%, 99.85%, 99.76% and 99.69%, respectively. Compared to the Holding Vector Regression Machine and Long and Short Term Memory Networks, Neural Networks and Random Forests perform better in the prediction of haze fouling flash of porcelain insulators in contact networks, with Accuracy values of 96.43% and 97.08%, respectively. But compared with the method in this paper, the prediction effect is still insufficient.

In summary, compared with the other four prediction methods, the prediction method in this paper has better prediction performance, Precision, F1 and Accuracy indexes are optimal, can better distinguish the boundary of the occurrence of porcelain insulator haze tainted flash, and has better robustness and adaptability to the unbalanced dataset used.

3.3

Reactor insulation fault diagnosis and prediction results

Based on the circuit parameters of a certain type of train power supply system, the algorithm proposed in this paper is tested and verified by simulation of reactor insulation faults. Assuming that its equivalent grounded insulation resistance value is 104Ω at normal time, its insulation degradation fault is simulated at different grounding positions. At t of 5.1s, 5.2s and 5.3s, its actual insulation resistance value is adjusted to 3000Ω, 1500Ω and 800Ω in turn.

The simulation effect of the reactor front-end grounded insulation fault diagnosis and prediction is shown in Fig. 8, and Figs. (a)~(c) show the relevant sensor sampling values, the real-time calculation results of the fault characteristic quantity, and the diagnosis and prediction results, respectively. From Fig. 8(a), it can be seen that when grounding occurs at the front end of the reactor, its grounding detection voltage value (U_jd–_{F_DCP}) starts to superimpose a pulsating AC quantity with a frequency of 100 Hz, and its pulse amplitude gradually increases as the value of the grounding insulating resistance decreases. From Fig. 8(b), it can be seen that feature J_{E_x1}(k) ~ J_{E_x4}(k) gradually deviates from the normal value, and J_{E_x2}(k) has a positive deviation, while the remaining 3 have a negative deviation. Among feature J_{D_x1}(k) ~ J_{D_x4}(k), J_{D_x2}(k) does not differ much from the normal value, while J_{D_x1}(k), J_{D_x3}(k), and J_{D_x4}(k) gradually become larger with the decrease of insulation resistance. From Fig. 8(c), it can be seen that when the insulation resistance resistance drops to 1800 Ω at the moment of 5.1s, the system detects the abnormality after about 5ms and sets the abnormality flag bit F to one. Subsequently, the exact grounding point is located after about 20 ms and the grounding flag F_DCP on the reactor front is set to 1. The reconstructed value of its grounded insulation resistance R_{jd, E} achieves an accurate tracking of the true value R_{jd, R} for 1 calculation cycle (5.14s moments) after the fault type is localized. Throughout the process of grounded insulation degradation, the health factor Z_HI changes dynamically with the reconstructed value of grounded insulation resistance, which realizes the real-time tracking of the degradation degree of the system, and thus achieves the effect of early warning of ground faults in the front end of the reactor.

4

Maintenance decisions

The maintenance decision-making algorithm adopts the relevant algorithms researched by other members of the group, inputs the health status of the equipment and the cost of each maintenance, and outputs the recommended maintenance method and the number of days between the next maintenance cycle, and its calculation steps are as follows:

1) Set the health state of the traction transformer (s) and the corresponding value of the maintenance method (m).

2) Set the state transfer matrix P(m,s₁,s₂), where s₁ and s₂ indicate the health state of the equipment before and after a certain maintenance, respectively.

3) Set the next maintenance cycle interval optional days matrix O, equipment in different health states using different maintenance methods of maintenance cost matrix C, various types of health state of routine maintenance management cost matrix D, equipment testing costs T_e, failure downtime loss S_t.

4) A maintenance decision model based on Markov process [23] is used. Based on the input parameters P(m,s₁,s₂), O, C, D, T_e, S_t, the recommended maintenance mode M_m and the number of days between the next maintenance cycles M_d are calculated by the strategy iteration algorithm.

5

Conclusion

The article constructs a fault prediction system for urban rail transit power supply system, simulates traction transformer fault diagnosis through Monte Carlo, ranges the faults, and deduces the traction power supply faults. The prediction performance of the fault prediction method in this paper is analyzed through examples, and maintenance decisions are proposed.

The fitting values of the fault prediction model in this paper on the training set and test set are 0.94868 and 0.90857, respectively, and the model has a good fit, and the relative error between the real value of the contact network wear and the predicted value of this paper’s model is no more than 5%, which has a high prediction accuracy. The Precision, Recall, F1 and Accuracy values of this paper’s prediction method in contact network lightning strike tripping fault prediction are 96.74%, 96.48%, 97.64% and 94.59%, respectively, and 95.84%, 94.98%, 95.46% and 94.98% in floating object intrusion contact network fault prediction, respectively, and 99.87%, 99.87%, 99.88%, and 94.98%, respectively, in the haze dirty flash fault prediction are 99.87%, 99.85%, 99.76% and 99.69%, respectively, and the prediction model of this paper has the best prediction performance among all prediction models. In the reactor insulation fault prediction, this paper’s method can accurately determine the reactor front-end grounding and accurately predict the degree of grounding insulation degradation.

Sprache:: Englisch

Zeitrahmen der Veröffentlichung:: 1 Hefte pro Jahr
Fachgebiete der Zeitschrift:: Biologie, Biologie, andere, Mathematik, Angewandte Mathematik, Mathematik, Allgemeines, Physik, Physik, andere

Zeitschrift RSS Feed

Fault prediction and maintenance of urban rail transit power supply system based on big data

Wenfei Zhao

Jiao Deng

Online veröffentlicht: 17. März 2025

Eingereicht: 04. Nov. 2024

Akzeptiert: 02. Feb. 2025

DOI: https://doi.org/10.2478/amns-2025-0225

SchlüsselwörterMonte Carlo simulation, Traction transformer, Fault ranging, Fault prediction

© 2025 Wenfei Zhao et al., published by Sciendo

This work is licensed under the Creative Commons Attribution 4.0 International License.

Schlüsselwörter
Monte Carlo simulation, Traction transformer, Fault ranging, Fault prediction