A Data-Driven Model for the Prediction of Water Mains Condition

Background
Water distribution networks are critical infrastructures in cities all around the world responsible for delivering safe drinking water to their various costumers. Many water pipes in these networks are close to the end of their service life. As such, they are susceptible to frequent failures. Their failure may result in huge financial, environmental, and social losses [1]. According to the 2018 study from Utah State University on water main break rates in United States and Canada [2], the overall water main break rate in North America increased to 14/(100 miles)/year from 11/(100 miles)/year, reported in a similar survey in 2012. This indicates an increase of 27% in only 6 years. Based on estimations from the American Water Works Association (AWWA), the cost of restoring buried water infrastructure will, most likely, be more than $1 trillion over the next 25 years if the current level of water service is to be kept [3]. Therefore, well-established plans are crucial to rehabilitate and replace the critical water pipes that are in the final stage of their life. Predictive maintenance is a technique that uses data analysis tools to determine the condition of equipment in-use and predict when rehabilitation/repair is required in order to prevent unexpected failures and breakdowns. To assess equipment condition, predictive maintenance uses historical and real-time data from nondestructive testing technologies such as electromagnetic, ultrasonic and acoustic. The main advantages of implementing predictive maintenance are: - Lengthening asset lifespan - Lowering the risk of failures - Decreasing unplanned shutdowns - Saving money For water networks, conducting field measurements on all pipes is not reasonable, due to the constraints imposed by cost/budget and time. Therefore, there is a need for developing a reliable predictive model that can estimate the condition of water pipes in a network. Such a model can provide valuable information about the current and future states of the network and enable municipalities to take efficient predictive actions. In the recent years, data driven approaches such as machine learning techniques have been widely used to develop predictive models and digital twins in different industries.
Objective In this study, based on the condition assessment data available for a small portion of a water network, a model was developed to predict the condition of the remainder of the water network.
Methodology
A regression Artificial Neural Network (ANN) model was developed to predict the remaining wall thickness of water pipes (target), based on a series of input data (feature set). The dataset used in this study was provided by Echologics. Echologics uses the ePulse technology to assess the condition of water pipes by measuring the remaining wall thickness without the need for large excavations or service disruptions. With this technology, acoustic sensors are placed on available access points such as fire hydrants or valves. A sound wave is induced in the pipeline and travels along the pipe. By measuring the speed of the wave in a pipe segment and relating it to the actual strength, the actual condition and the overall remaining thickness of the pipe segment can be determined. The dataset contains 860 segments with a total network length of 117 km. For each pipe in the dataset, the following variables were readily available: - Original wall thickness - Pipe diameter - Pipe material - Segment length - Measured wall thickness - Laid and inspection dates - Pipe location The studies conducted on the degradation of water network pipes implied that geographical features, such as proximity of pipes to major roads or water bodies, and the type of soil the pipe is buried in may have significant impacts on the condition of pipes [4,5]. Therefore, using the location of the pipes and the available GIS databases, the following features were extracted: - Proximity to major/minor highways - Soil type - Environment's corrosivity index The dataset underwent preprocessing steps (imputing missing values and one-hot encoding of categorical variables) and normalization. Since the feature set contained both categorical and numerical variables, the inputs of those variables were fed to the model separately. A thorough set of tests were performed to optimize the model hyperparameters such as number of layers, number of neurons, learning rate and activation functions parameters. The general architecture of the neural network model is displayed in Figure 1. To develop the model, 70% of the records in the dataset was used for training and the remaining 30% was used for the evaluation of the performance of the model.
Results
Figure 2 shows the predicted remaining thickness versus the measured remaining thickness for the pipes in the testing set. As it can be observed from the graph, a good correlation between predicted and measured values was obtained. Mean Absolute percentage error (MAPE) and Root Mean Squared Error (RMSE) were used as metrics to evaluate the model prediction performance. RMSE and MAPE values of 0.054 and 0.16 were achieved for the model on the testing set. This means that the model can predict the average condition of the remaining pipes with an accuracy of 84%. To further assess the performance of the model, the measured and predicted remaining thicknesses of the pipes in the testing set were plotted against their degradation levels (thickness reduction relative to the original wall thickness) in Figure 3. As can be seen, the predictions of the model are in good agreement with the field measurements. The performance of the model deteriorates for the pipes at higher degradation levels. This is believed to be due to the limited data available at higher degradation levels, which compromises the ability of the model to train in that particular range. The predictive performance of the model can be significantly improved by testing more samples and increasing the size of the dataset.
Conclusions
The neural network model developed based on the ePulse measurements for the remaining thickness of the water pipes achieved an accuracy of 84%. The accuracy of the model can be improved by increasing the samples size, collecting more related and reliable data regarding the corrosivity of the environment, or other related parameters to include as input to the model. Such a predictive model, as a digital twin for the condition assessment technology, can be utilized to assess the condition of the other pipes in the network for further inspection/monitoring and or planning preventive actions.