Water Supply Pipeline Failure Evaluation Model Based on Particle Swarm Optimization Neural Network

Abstract

The degradation and failure of the urban water supply network may lead to serious safety hazards, including pipe breaks, water supply interruptions, water resource losses, and contaminant intrusions. The risk evaluation of water supply pipeline failure in a distribution network is a challenging task, because most of the available data cannot fully reflect pipeline failure events and many of the mechanisms still cannot be fully understood. Therefore, a predictive model is urgently needed to assess pipeline failure risk based on available data. In this paper, based on the traditional risk assessment theory, seven main factors affecting pipeline failure are selected and scored, and then a pipeline failure model is established by using the particle swarm optimization (PSO) neural network. The model uses the neural network training of historical data to evaluate the failure of the water supply pipeline, and the PSO is used to optimize the neural network to effectively improve the training time and accuracy. The model error and correlation coefficient are 0.003 and 0.987, respectively. The proposed model can be used as a powerful support tool to assist infrastructure managers and pipeline maintainers in their plans and decision-making.

1. Introduction

Pipe failure prediction is a prerequisite for preventive maintenance, a widely adopted industrial practice, particularly in water utility companies. The core function of a pipe failure model is to assess the failure risk for each pipe segment using various information, such as pipe material, pipe diameter, and data from past pipe events. This assessment enables effective maintenance planning and the calculation of maintenance budgets.

The pipe failure problem has long been an issue of concern as reported in studies [1,2,3,4]. With the acceleration of global urbanization, urban water distribution networks are continuously evolving. However, the drinking water infrastructure in many countries is generally deteriorating, leading to an increasing likelihood of pipe failures year by year [5]. The structural deterioration has posed challenges not only to daily life but also to sustainable societies since water is an essential and valuable resource for humanity. It has been estimated that more than USD 32 billion cubic meters of treated water physically leak annually through distributed networks worldwide [6]. The corresponding total annual cost is about USD 14 billion. The leakage of the urban water supply network does not only cause waste of water resources and economic loss, but also seriously affects social order and production. Therefore, it is essential to take measures to reduce the risk of pipeline leakage.

A traditional method for pipe risk assessment involves scoring by experienced experts who establish industry standards to evaluate pipe risk. However, this knowledge-intensive approach is often subjective and can lead to inaccurate assessments when environmental conditions change, such as alterations in geographical conditions [7]. Moreover, compared to the more established standards for oil pipelines [8], the standards for urban water distribution networks are less mature and require further exploration.

On the basis of traditional methods and risk manuals, a series of academic studies on pipeline risk have emerged, such as the Improved AHP–TOPSIS model [9], LSTM networks [10], and fuzzy fault tree analysis [11]. An enhanced AHP–TOPSIS model for determining positive and negative ideal solutions, along with an improved normalized equation for benefit and cost indices, makes TOPSIS more suitable for comprehensive risk evaluation of pipelines [9]. Shin proposes a pipeline risk detection method using a combination of LSTM networks and autoencoders to improve the accuracy and precision of identifying leaks in complex urban water supply systems [10]. The advantage of fuzzy fault tree analysis is that it can analyze the failure probability of pipelines without historical probability data [11]. With the continuous maturity of computer technology and artificial algorithms, the Monte Carlo method [12] and neural network [1] that rely on computer computing power are gradually applied in pipeline risk research. The leakage risk assessment model based on the Monte Carlo method is able to accurately achieve leakage detection, cause analysis and leakage volume forecast by avoiding deviation of data, model parameters and the uncertainties caused by the method [12]. Li proposes a method to predict pipeline failure pressure by optimizing the weights and thresholds of a BP neural network using a genetic algorithm [13].

Based on the traditional pipeline failure evaluation system, this study integrates the safety and stability of water pipeline structures to develop a particle swarm optimization-neural network model (PSO-BP) for predicting water pipeline failures. This model comprehensively considers various risk factors affecting the safety of water pipelines, such as pipe condition, environmental factors, and third-party damage, and refines these risk factors by setting quantitative scoring criteria. Data collection and scoring for each risk indicator are performed based on the characteristics of the pipeline cross-sections. The scores are used as input conditions for the neural network, and the PSO algorithm is employed to optimize the neural network, enhancing its effectiveness and generalization ability, and thereby improving the accuracy of pipeline leakage prediction. This predictive model can provide a basis for decision-making in the daily maintenance and renovation of water distribution systems.

2. Pipeline Failure Evaluation System

2.1. Structural of Pipeline Failure Evaluation System

Based on the theory of engineering risk analysis, the failure evaluation system of the water supply pipeline is constructed by using the multi-factor comprehensive evaluation method. Through a comparative analysis of the factors affecting the safety of water supply pipelines, important factors related to pipeline safety are screened. The failure evaluation system includes three main risk categories: pipeline state risk, environmental risk and third-party risk. In particular, pipeline state risk indicators include pipe diameter, pipe age, pipe material and pipe operating pressure; environmental risks include soil environment; and third-party risks include buried depth of pipeline and ground load. The pipeline failure evaluation architecture is shown in Figure 1.

2.2. Quantitative Evaluation Indicators

In the evaluation system, some risk indicators cannot be directly quantified, and the absolute values of other quantifiable risk indicators can be much larger than others, which is not conducive to the calculation of the models. Therefore, scoring standards are established for different risk indicators by means of mechanical calculation, numerical analysis or reference standard indicators, and then unified quantification is carried out in accordance with the scoring standards. For example, according to the mechanical properties of different pipe materials, specific pipe materials that are often used in water supply networks are scored. The strength of the metal pipe is closely related to its corrosion status, and the strength of the plastic pipe is strongly dependent on the initial laying conditions and its own aging. The buried depth of the pipeline is determined by referring to the Pipeline Risk Management Manual [8]. The operating pressure of the pipe network refers to the design specifications and standards of the water supply pipe network. In the case of pipeline leakage, the operating pressure of the pipe network is proportional to the leakage amount. The higher the operating pressure is, the higher the leakage rate of the pipeline will be. The road load is graded according to the Code for Design of Urban Road Engineering [14].

The score of different factors varied from 1 to 10 points according to the degree of correlation between the factors and pipeline accidents.

Table 1. Score of different pipe diameters.

Pipe Diameters/mm	≤400	600	800	1000	≥1200
Score	10	6	2	6	2

,

Table 2. Score of different pipe materials.

Pipe Materials	Nodular Cast Iron	Carbon Steel	Prestressed Cement	PCCP	Cast Iron
Score	10	8	6	4	2

,

Table 3. Score of different pipe pressures.

Pipe Pressures/MPa	≤0.15	0.15~0.25	0.25~0.35	0.35~0.45	≥0.45
Score	10	8	6	4	2

,

Table 4. Score of different pipe ages.

Pipe Ages/Year	<5	5–15	15–30	30–45	>45
Score	2	6	10	6	2

,

Table 5. Score of corrosion of soil.

pH Value	>8	4~8	<4
Score	5.5	10	1

,

Table 6. Score of pipes buried depth.

Pipes Buried Depth/m	≥2.5	2.5–1.5	1.5–1.0	1.0–0.5	0.5–0
Score	10	8	6	4	2

and

Table 7. Score of ground load.

Ground Load	Fourth Level	Third Level	Second Level	First Level
Score	10	8	6	4

show the scores of different factor indexes. The higher the score is, the less impact the index has on pipeline safety; the lower the score is, the greater impact the index has on pipeline safety and the higher the risk degree. The scoring criteria shown in Table 1, Table 2, Table 3, Table 4, Table 5, Table 6 and Table 7 apply to pipeline networks with variations in pipeline pressure.

The grading of pipe diameter and pipe material is based on the probability of historical pipeline accidents in Luoyang, Henan Province, China and the opinions of experts in related fields. Pipeline pressure is a major factor in pipeline failure. The increase in pipeline pressure will directly increase the pipeline stress, which will aggravate the risk of pipeline corrosion and pipeline cracking. Once a leak occurs, an increased pipe pressure will increase the amount of the leakage. The effect of pipe age on pipe failure can be characterized by the famous “bathtub” curve [15]. At the initial stage of pipeline construction, there is a high initial accident rate, which is generally called the “initial failure stage”. With the elimination of these defects, the curve in the second stage tends to be flat, reflecting that random events maintain a fairly stable accident rate in this area. When the pipeline age exceeds a certain number of years, the pipeline wear can become more serious, which can increase the pipeline failure rate.

The concentration of ions in soil will significantly affect the corrosion rate, which is generally measured by pH value. Corrosion may be enhanced at a pH less than or above 8 (compared with the neutral pH range of 4 to 8). In general, soils that are more acidic (pH < 4) are more corrosive than soils that are more alkaline (pH > 8). The shallower the buried pipeline is, the weaker the ground load capacity is, the more vulnerable it is to third-party risks such as natural damage and man-made damage, and the greater the risk of pipeline failure.

3. Pipeline Failure Model

3.1. Back Propagation Neural Network Parameter Settings

The neural network reflects the structure and function of the human brain neural network, which abstracts the basic characteristics of the real brain according to theory and simplifies it into an information processing system. The back propagation neural network (BPNN) is a kind of multilayer feedforward neural network trained according to the error back propagation algorithm, and is one of the most widely used neural networks. Such neural networks typically consist of one input layer, one or more hidden layers, and one output layer, and its basic structure is shown in Figure 2. The gradient search technique of the BP neural network is the gradient descent method, which takes the minimum mean square error (MSE) as the training basis [16].

The input layer and output layer formulas of the BP neural network are expressed in Equations (1)–(4) [17]:

$$s_j=\sum _{j=1}^L{\omega }_{ij}{x}_i+b_j$$

(1)

$$a_j=g\left(s_j\right)$$

(2)

$$p_k=\sum _{k=1}^L{\omega }_{jk}{a}_i+b_k$$

(3)

$$t_k=g\left(p_k\right)$$

(4)

where $s_j$ is the input value of the $j$ node of the hidden layer, $x_i(i=1, 2,\cdots ,m)$ is the input value of BP neural network; ${\omega }_{ij}$ is the weight of the input layer to the hidden layer; $a_j$ is the output value of the $j$ node in the hidden layer; $b_j$ is the hidden layer threshold; $p_k$ is the input value of the $k$ node in the output layer; $t_k$ is the input value of the $k$ node in the output layer; ${\omega }_{jk}$ is the weight of the hidden layer to the output layer; $b_k$ is the output layer threshold; $g\left(x\right)$ is the transfer function of BP neural network. Error function $f\left(x\right)$ and threshold correction coefficient are shown in Equations (5)–(9):

$$f\left(x\right)={\displaystyle \frac{1}{2}}\sum _{k=1}^N{\left(y_k-t_k\right)}^2$$

(5)

$${∆\omega }_{jk}=-\eta {\displaystyle \frac{\left[f\left(x\right)\right]}{\partial {\omega }_{jk}}}$$

(6)

$${∆\omega }_{ij}=-\eta {\displaystyle \frac{\left[f\left(x\right)\right]}{\partial {\omega }_{ij}}}$$

(7)

$${∆b}_k=-\eta {\displaystyle \frac{\left[f\left(x\right)\right]}{\partial b_k}}$$

(8)

$${∆b}_j=-\eta {\displaystyle \frac{\left[f\left(x\right)\right]}{\partial b_j}}$$

(9)

where $N$ is the total number of samples; $y_k$ is the measured value; $t_k$ is the actual output value; $\eta$ is the coefficient of the learning rate; ${∆\omega }_{jk}$ is the weight correction coefficient of the hidden layer to the output layer; ${∆\omega }_{ij}$ is the weight correction coefficient from the input layer to the hidden layer; ${∆b}_k$ is the threshold correction coefficient of the output layer; and ${∆b}_j$ is the threshold correction coefficient of the hidden layer.

In the BP neural network model of pipeline failure, the risk subitems of each pipeline are taken as the input of the model, and leakage is taken as the output of the network model. The number of hidden layer nodes $m$ is the core part of the neural network structure, which can be calculated by Equation (10) [18]:

$$m=\sqrt{l+n}+a$$

(10)

where $l, m, n$ are the number of nodes of the input layer, hidden layer and output layer of the neural network, respectively [19]; $a$ is the regulating constant, $a=1, 2, 3,\dots , 10$; There are seven risk subitems in the model, and thus the number of input nodes $l=7$, and the output is leakage, represented by logical values $01$ and $10$, then $n=2$. In the data selected in this paper, the error is minimum when $m=11$, that is, $a=8$. The number of nodes in the hidden layer has a great influence on the prediction results of the model [20]. If the number of nodes is too small, the network will not obtain sufficient information for the training, which will lead to a poor training effect and thus greater errors in the prediction results. When the number of nodes of the hidden layer is increased, the training ability can be improved, and so does the prediction ability. However, when the number of nodes is excessive, the prediction ability will decline.

3.2. Back Propagation Neural Network Optimized by Particle Swarm Optimization

Particle swarm optimization (PSO) is a random global search algorithm based on swarm intelligence theory [21,22], which is characterized by information sharing and coevolution among individuals in a colony by simulating the migration and aggregation behavior of birds in the process of foraging and abstract modeling the rule of colony activity. It is widely used in various fields because of its simple concept, fast convergence speed and easy implementation. Each particle in the algorithm is a solution to the problem. Through calculation and comparison of fitness function values, the position and velocity of particles are constantly updated, and the whole space is searched to find the global optimal solution.

This study proposes a PSO-BP model to address the issues of traditional BP neural networks, such as their tendency to become trapped in local minima, slow convergence, and weak generalization ability. Compared with BP neural networks, the PSO algorithm can search in a larger space and prevent BP neural network from falling into the limitation of local optimal solution. The traditional PSO algorithm optimized BP neural network method is to take the weight and bias of the network as the position of the particle and update the weight and bias of the network, so as to realize the optimization of the BP neural network. However, due to the problem of premature convergence of particle swarm optimization, the algorithm falls into the local minimum value and does not take into account the error back propagation characteristic of the BP neural network itself. Therefore, this paper introduces the PSO algorithm into the BP neural network, and uses the way of error back propagation to constantly update the position and velocity of each particle in the particle swarm based on the optimal of each particle in the particle swarm and its global optimal. By calculating the fitness value of the whole space search, and keeping the back propagation (BP) neural network to search the global optimum with the PSO algorithm, the above problems can be avoided.

In the particle population, each particle should have a position vector, a velocity vector and an adaptive value determined by the established objective function [23]. The position vector of particle i in n-dimensional space is expressed as $X_i=\left(X_{1 },X_2,\dots ,X_n\right)$, and the flight velocity vector is expressed as $V_i=\left(V_1,V_2,\dots ,V_n\right)$, each dimension has a maximum speed limit $V_{max}$. If the speed of one dimension exceeds $V_{max}$, the speed of this dimension is limited to $V_{max}$. The velocity and position of the particle are updated by:

$$\left\{\begin{array}{c}{V}_{id}^{(k+1)}=\omega V_{id}^{\left(k\right)}+c_1{r}_1\left(P_{best}-X_{id}^k\right)+c_2{r}_2(G_{best}-X_{id}^k)\\ X_{id}^{(k+1)}=X_{id}^k+V_{id}^{(k+1)}\end{array}\right.$$

(11)

where, $V_{id}^k$ is the velocity of the $d$-dimension of particle $i$ in the $k$th iteration; $X_{id}^k$ is the position of the $d$-dimension of particle $i$ in the $k$th iteration; $\omega$ is inertial weight; $c_1$ and $c_2$ are learning factors; and $r_1$ and $r_2$ are random numbers.

According to the objective function, the particle can calculate the adaptive value of the optimal solution at the current position, and know the best position discovered so far ($P_{best}$) and the current position X_i, that is, the particle’s own flight experience. In addition, each particle also knows the ($G_{best}$) position (the flight experience of the swarm) that is found for all particles in the entire swarm. The particle uses its own experience and the best experience of the group to determine how to adjust the direction and speed of its flight for the next iteration. In this way, the whole population of particles will gradually approach the optimal solution.

3.3. PSO-BP Pipeline Failure Model

The PSO-BP Pipeline failure model flow chart is shown in Figure 3.

The main steps of the PSO-based parameter optimization process are summarized below:

(1) Determine the structure of the neural network, select the number of nodes in the input and output layers of the network according to the sample data, and determine the number of nodes in the hidden layer.

(2) According to the established BP neural network initialization parameters of the PSO algorithm, build a single particle network and particle swarm network, set the particle swarm size, set the number of iterations, inertia weight $\omega$, $c_1$ and $c_2$, etc.

(3) Input training data and testing data sets.

(4) Iterative optimization, traverse all particles in the particle swarm, and calculate the fitness value of each particle.

(5) In the iteration process, compare fitness values with the corresponding terms in three parts.

In part 1, the current fitness value of particles is compared with the optimal position of individual particles in the past dynasties. If the current fitness value is better than that of the individual particles in the past dynasties, $P_{best}$ is set as the current particle position; otherwise, the single particle network is updated according to the original historical optimal back propagation.

In part 2, the current fitness value of the particles is compared with the global optimal position of all particles in the population. If the current fitness value is better than the global optimal position of all particles in the population, the current particle position is assigned to $G_{best}$; otherwise, the PSO network is updated by the propagation according to the global historical optimal inverse position.

In part 3, the current $P_{best}$ and $G_{best}$ of particles is compared. If $P_{best}$ is better than G_best, assign the current $P_{best}$ of particles to $G_{best}$, otherwise update the particle swarm network according to the original $G_{best}$ back propagation.

(6) Update $P_{best}$ and $G_{best}$, and update the positions and velocities of all particles in the particle swarm by using the characteristics of adjusting weights and thresholds by BP neural network back propagation.

(7) Judge whether the overall error reaches the target convergence precision range or the number of iterations reaches the maximum number of iterations. If the conditions are met, the calculation is terminated; otherwise, the number of iterations +1 returns to step (4) to continue the iterative calculation of the fitness value of each particle.

(8) Verify the trained neural network with the test set.

4. Experiment and Analysis

To demonstrate the impact of pipeline pressure on pipeline failure, the project team conducted joint experiments with the Luoyang Metrology and Testing Center. The experimental setup is illustrated in Figure 4 and includes a 1.1 m long test pipe with an outer diameter of 107 mm and an inner diameter of 100 mm. The test pipe is produced by Guorun Pipe Industry Co., Ltd., Luoyang, Henan Province, China. One end of the test pipe is connected to a standard testing bench via a DN100 flange, which is manufactured by YINSTEEL Company in Ningbo, China. While the other end is sealed with a DN100 blind flange to maintain stable internal pressure. The DN100 blind flange is manufactured by YINSTEEL Company in Ningbo, China. Four circular leak holes with diameters of 2 mm, 4 mm, 6 mm, and 8 mm were set at the bottom of the test pipe.

The test pipe is connected to a standard testing bench, which is illustrated in Figure 5. The standard testing bench is a comprehensive testing system that includes major equipment such as a variable frequency pump, a pressure stabilizing tank, a pressure transmitter, a flow meter, exhaust valves, test connection pipes, and a water reservoir. The standard test bench and related instruments are manufactured by Dandong best automatic engineering & meter Co., Ltd. (Dandong, China).

Figure 6 shows some scenarios from the experiment:

• Figure 6a displays the pressure stabilizing tank in the standard testing bench. The tank has a maximum pressure capacity of 0.6 MPa and a volume of 30,000 L, ensuring stable internal pressure in the pipe.
• Figure 6b primarily showcases the inspection devices in the testing system, including the pressure transmitter and flow meter.
• Figure 6c illustrates the test pipe and its state during a leak.
• Figure 6d shows the interior of the control room, where experimenters can control the pipe’s status and measure the leak volume.

Figure 6. Parts of the experimental scenario (a) buffer tank (b) outside of the control room and the standard test system (c) experimental pipeline (d) inside of the control room.

Figure 6. Parts of the experimental scenario (a) buffer tank (b) outside of the control room and the standard test system (c) experimental pipeline (d) inside of the control room.

The experiment begins after the pipeline installation is completed and the sealing is checked. The pipe pressure was raised to 0.35 Mpa, 0.30 Mpa, 0.25 Mpa and 0.20 Mpa, respectively. The four pipe pressures, respectively, represent the pressures in different sections of the urban water supply pipeline [24].

Based on actual water supply operations, four internal pressure conditions were set for the pipeline: 0.35 MPa, 0.30 MPa, 0.25 MPa, and 0.20 MPa [23]. The experiment measured the leak rates for different orifice diameters under these various internal pressure conditions. The experimental results are shown in

Table 8. Leakage volume under different working conditions.

Leakage Leakage Hole Volume (L) (mm) Pressure (MPa)	2	4	6	8
0.35	0.194	0.942	1.950	3.055
0.30	0.189	0.885	1.791	2.811
0.25	0.173	0.794	1.612	2.587
0.20	0.134	0.713	1.413	2.301

. Leakage diagrams under different operating conditions are drawn according to Table 8. Figure 7 shows the leakage amounts under different working conditions. When the leakage aperture remains unchanged, the leakage area is also unchanged. The increase in the leakage amount with the pressure in the pipe indicates that the stress at the corresponding position increases under the critical leakage condition. When the leakage occurs, the pipeline pressure is proportional to the amount of the leakage, that is, the loss is greater. Therefore, the risk of pipeline pressure should be fully considered when evaluating pipeline failure.

In practice, the shape of the leak hole is often not round. In order to study the influence of the shape of the leak hole on the leakage amount, a finite element simulation of the leak hole with different shapes was carried out, and the parameters of the three-dimensional pipeline model were consistent with the experimental parameters. The simulation results are shown in Figure 8. Figure 8a shows the flow field near the circular leak hole, while Figure 8b shows the flow field near the rectangular leak hole. The pipeline pressure, leakage hole surface area and flow field state of the two are consistent. Through calculation, it can be concluded that the average surface velocity of the circular leakage hole is 4.58 m/s, and the average surface velocity of the rectangular leakage hole is 4.87 m/s, with a difference of 6.3%. This shows that under the same pipeline pressure conditions, the influence of the shape of the leakage aperture on the leakage amount is not significant. Combined with the experiment, it is further shown that the pipeline pressure is the main factor affecting the leakage amount.

5. Case Study

Luoyang is an industrial city with many industrial enterprises, and thus industrial electricity needs to be guaranteed. Located in the northwest of Luoyang, the GH is a hydroelectric power plant. It uses water supply pipes to transfer water from the city’s water plants to ensure power generation. The pipes are 37 km long: 8.6 km in the urban area and 28.4 km in the suburbs. The pipeline GIS map is shown in Figure 9. Point A is located in the west of the city as the starting point of the pipeline, and point B is the GH power plant at the terminal point of the pipeline. The pipeline was built in 2009 and completed in 2010. Due to different geological and environmental conditions, different pipes are used in the construction according to the construction requirements. There are many hillsides along the pipeline, and the terrain is undulating and environmental conditions are different, which is suitable for a pipeline failure model study. As the pipeline is too long, area C is selected as the key research area in this paper. Area C is the area of the first-level pressurized pipe section, where the terrain is sinuous and undulating. The pipeline needs to climb over hills and hillsides according to the terrain, the highest of which is about 90 m above sea level, which means that the primary pressure must reach enough pipe pressure to allow the water to flow over the hill. However, the continuous high pressure in the pipeline will increase the risk of the pipeline failure.

The research team followed relevant departments of water supply in Luoyang to observe the pipeline leakage situation on site, as shown in Figure 10. Based on the data collection at the leak site and the historical data of Luoyang water supply departments, a true and reliable pipeline leak data set was obtained.

The pipeline data in the data set were quantitatively scored according to the evaluation indexes in Section 2.2. Some data in the data set are shown in

Table 9. Pipeline data.

Number	Pipeline Diameters	Pipeline Material	Pipeline Pressure	Pipeline Age	Corrosion of Soil	Pipelines Buried Depth	Ground Load	Leakage
1	2	4	2	6	10	10	8	Yes
2	2	6	6	6	5.5	4	4	Yes
3	6	10	4	6	10	4	10	No
4	6	6	8	6	10	6	8	No
5	10	6	6	6	10	8	6	No

.

The dataset contains 100 data, including 20 leaking data and 80 non-leaking data. Shuffling the data order, 80% of the data are used as a training set and 20% of the data are used as a test set. The pipeline risk item score is used as the input and leakage is used as output. The PSO-BP neural network is trained by the training set, and the reliability of the model is evaluated by the test set.

The results of the PSO-BP neural network were compared with those of a traditional BP neural network [25]. The error between the predicted value and the actual value of the BP neural network was 0.032, and the error of the PSO-BP neural network was 0.003, which was smaller than the former. Figure 11 shows the correlation coefficient of the two methods. The correlation coefficient of the BP neural network is 0.913, and that of the PSO-BP neural network is 0.983. Therefore, the proposed method has a smaller error and stronger correlation with the actual values compared with the traditional method. These results show that the nonlinear relationship established by the BP neural network optimized by the PSO algorithm is accurate, the training effect is better, and the prediction result is reliable.

In order to further understand the influence degree of seven influencing factors on pipeline leakage, correlation analysis is necessary. The analysis results are shown in Figure 12. It can be seen that the correlation coefficient of pipeline pressure is the largest, indicating that pipeline pressure is the main influencing factor of pipeline failure, the second influencing factor is ground load, and the third influencing factor is pipeline burial depth, which is followed by pipeline diameters, pipe material and corrosion of soil. The least influencing factor is pipeline age.

6. Suggestions and Improvement Measures

The model is used to analyze each area of the pipeline, and it is found that area C has a higher risk compared with the other areas. The main reason is that the pipeline pressure is continuously high, and the buried depth is shallow due to the topography fluctuation. Therefore, we suggest setting up a secondary pressurizing station in the steep hillside area of area C and changing this part of the pipeline from the original primary pressurizing mode to the secondary pressurizing mode, to minimize the risk caused by the continuous high pressure of the pipeline. Further, warning signs should be set up in shallow places and defects on the road surface to remind construction parties and large vehicles to pay attention to reduce third-party risks.

7. Conclusions

In this study, a pipeline failure risk assessment system was constructed, and risk sub-items were scored to unify all indicators and facilitate neural network processing. The important influence of pipeline pressure on pipeline risk was verified by experiments. A PSO algorithm was used to optimize the BP neural network, and a PSO-BP pipeline failure model was established, which was verified experimentally. The experimental results demonstrated that our PSO-BP Pipeline failure model has less error and higher correlation than the traditional method. Finally, the feasibility and universality of the method were verified by a practical case and improvement suggestions were also given for the pipeline cases.