Predicting pipe failures with CMMS/GIS/Modeling
Yazdan T. Emrani RJN Group, Inc. 1003 Wirt Road, Suite 204 Houston, Texas 77055 The intent of this paper is to describe an important planning tool for the City of Houston’s underground infrastructure maintenance program. The development of a GIS based predictive modeling tool is part of a general trend toward reducing corrective maintenance by increasing preventive and predictive maintenance activities. There are benefits of a design based on the use of a GIS as the central technology incorporating predictive modeling tools. Experience has shown that a significant number of infrastructure repairs are performed on an unscheduled basis. In this time of budget cuts and limited resources, the ability to optimize the use of maintenance dollars by employing predictive models in the planning stages is rapidly becoming a reality of underground infrastructure management. Planned maintenance for a facility in need of repair can yield significant savings over unscheduled or emergency repairs. The key is to enable planners to predict accurately which components are in the most urgent need of repair, and when others will need repair. This differs from corrective maintenance where repairs are made to fix an existing problem, or preventive maintenance where repairs are performed on a scheduled basis. Knowledge of which facilities are most in need of repair, or most likely to fail offers planners several advantages.
Commercial use of predictive modeling has been limited by a number of obstacles. One reason is models were primarily designed for research. Another is the need to draw upon skills from multiple disciplines (engineering, statistics, and computer science) to develop a robust predictive modeling tool. In addition, accessible, comprehensive, historical databases are necessary to provide the models with the data they require. Typically, considerable computational power was necessary to perform the analyses. The ability to develop a predictive modeling system now is due in part to the convergence and advancement of several key technologies. Computer hardware is plentifbl and relatively inexpensive in terms of computation power, disk capacity, and graphical display. Software packages such as geographic information systems, relational databases, statistical packages and artificial intelligence tools have reached a significant level of maturity. Further, these technologies are interacting through interfaces to each other. Statistical packages can operate on data in a GIS or RDBMS and links exist between GIS and facility management systems. In addition research activities are providing improved methods for prediction and display. Houston GIMS and predictive modeling The city of Houston is the fourth largest city in the United States with a population of 1.7 million covering an area of over 1500 square kilometers (over 600 square miles). The metropolitan area is comprised of over 19,000 square kilometers (over 7,500 square miles) with a population of 3.7 million. Department of public works and engineering DPW&Es budget for 1997 was over $600 M. Of this total, approximately $200 M was the capital expenditure program for wastewater facilities. Some of this money will be applied to emergency repair and unscheduled maintenance. The Department of Public Works maintains extensive facilities, including water distribution, wastewater, and storm water runoff systems. Table 1 shows the relative sizes of these three systems. In an effort to help streamline and optimize the management process, RJN Group, Inc. was contracted by the City of Houston to develop a Geographic Information Management System (GIMS). This system will serve multiple purposes, including map maintenance and publishing and maintenance management. In addition, it will contain the predictive modeling component for use by the DPW&E. Predictive modeling has proven to be effective in a variety of fields related to infrastructure management including water pump, station maintenance, water main breakage, and pavement assessment. GIMS’ predictive modeling component will be used to determine failure likelihoods for Houston’s underground infrastructure system components. TABLE 1. Estimated element lengths and counts for GIMS facility database components.
City of Houston has a dynamic growth rate. Houston continues to annex areas that are unincorporated, thereby increasing its size and population on an on-going basis. Data requirements for predictive modeling The development of a decision support system that provides predictive modeling information requires several basic data services. A database of factors related to the failure of each system component must be designed, developed and maintained. The database management system must support database development (data collection and maintenance), access to the data via several methods, and interactive data query and display mechanisms. A critical factor in developing an accurate model is having historical maintenance data. That is why it is imperative to develop an inventory of the existing underground infrastructure and track the maintenance history for these structures. The most efficient way to accomplish this is by means of a Computerized Maintenance Management System (CMMS). CMMS database development Data items describing each facility’s components must be continually captured, stored and maintained. Each of the facility components to be modeled will require access to numerous attributes for each actual component in the system. These attributes fall into several data categories: tabular, spatial, topological, and temporal. Tabular attributes include age, warranty date, and pipe type, and others. Spatial, topological and temporal attributes are important when predicting failure potential and failure rates. Spatial attributes include information concerning positional coincidence with other entities such as soil properties, and intersections with fault lines. Topological information includes properties of those features that are adjacent to the feature being modeled. Finally, temporal data will include historical maintenance records with time stamps, allowing repair dates and trends to be modeled. In GIMS; attribute management is provided by a combination of R.TN’sCMMS called CASS WORKS and ESRI’S Arc/Info GIS. The facility management system is built on top of the Oracle relational database management system (RDBMS), and maintains the tabular factors such as length, diameter, and pipe material. The GIS supports direct connections to the RDBMS, and will supply the spatial and topological model parameters including soil type, and number of breaks within a certain proximity to a specified feature. The temporal attributes can be viewed as calculated values based on tabular data. Data server tools The predictive model tools will require access to the database via several different methods. SQL based queries can be performed in either the GIS, the facility management system, or using the RDBMS tools directly. The facility management system is designed to allow easy development of complex SQL queries, which can isolate sets of records for modeling based on tabular data criteria. Using a GIS allows for the creation of spatial and topological oriented data access methods. The results of the model runs must be stored back into the database for subsequent inspection and analysis. Data presentation The modeling results must be displayed using effective methods. Those results which are tabular in nature (e.g. sorted lists of facilities in need of repair) are handled in CASS WORKS. The GIS provides graphic mechanisms to view the results. Thematic mapping of a facility’s failure likelihood values is easily performed. This allows the user to display different failure levels graphically. Visual inspection of the spatial distribution of the thematic display may provide insight into patterns related to facility failure. Predictive modeling tools Predictive models in GIMS will supply a variety of tools to the planner. These include descriptive statistical methods, a variety of queries, potential use of expert systems and neural networks, predictive statistical analysis, and spatial analysis, all of which will be described in this paper. Accuracy verification is critical to the successful development of any of these tools. Verification is dependent on the continued development of a large historical database, which is in the process of being populated. Tabular data screens have been developed to manage facility maintenance data and work order processing, including historical repair data. Descriptive statistical methods Statistical methods can enhance the information in the database by creating variable transformations, descriptive values and basic relationships between component attributes and failure. For instance, summary statistics (mean, standard deviation, median, etc.) can be computed for interval and ratio data types. In addition, linear regressions can be run to determine the correlations between factors and failures. For example, age and pipe diameter has been correlated with failure rates of water mains, even though this is not necessarily a consistently reliable indicator. These values can be added to the database and become available to the planner for subsequent modeling efforts. These new data, combined with existing tabular data will support the execution of powerfid exploratory queries. Queries Given a rich set of facility attributes, summary and descriptive statistics, powerful queries can be run to provide insight into attribute/failure relationships. Queries considered to be standard will be built directly into the interface to allow push button execution. It would also be desirable to allow the user to develop customized queries. This can readily be accomplished for SQL type queries. However, queries retrieving data that is not tabular are more difficult due to the lack of a query language that supports both spatial and tabular attributes. By providing tools, a data dictionary and access to the GIS programming language, users will be able to develop new queries. Following are some examples of the type of queries that could be posed:
Rule based expert systems are a natural extension of the use of selection rules. Using an expert system, complex selection rules could be formulated that would lead to the selection of high risk facilities. Expert systems offer a mechanism to capture knowledge of experts in the field in an attempt to retain and share their knowledge. Expert systems are not currently part of the GIMS effort. Expert systems have been used for sewer network maintenance in Paris with a limited degree of success and for pavement repair management with some signs of success by the New Mexico State Highway and Transportation Department. Predictive statistical methods Statistical predictive models are based on historical data. Dates of events with attendant descriptive information need to be available to the model. In time, as available information increases, the accuracy of this type of approach should improve. There will be more data available to the models for calibration. Survival analysis using Proportional Hazards Models has been used successfidly to determine the probability of failure for water mains. Development is currently underway based on the use of the Andersen-Gill Proportional Hazards Model. This model can be used to predict multiple future failures. Predicted failure rates can then be used in a cost benefits analysis to determine whether to ignore, repair, or replace the facility. One disadvantage of the use of statistical predictive models is their inherent complexity. Although attempts will be made to isolate the user from all the complexities of this type of procedure, the user will still be required to have a significant level of understanding of the statistical process in order to produce valid results. Neural networks Another method under investigation for predictive modeling involves the use of neural networks. Neural networks are a branch of artificial intelligence that is very good at solving pattern recognition problems. The patterns of attributes and failure events in the database should be detectable by neural networks. The major benefit to using neural networks is their ability to find the patterns in the data without a complete understanding of the predictive processes, in addition they can easily operate on all data types including nominal data. Neural networks offer a solution that requires less knowledge on the part of the user. This can be considered a black box approach. Spatial analysis The use of GIS as an integral component of a facility management decision support tool brings an additional set of analytical tools to the system: spatial analysis tools. Operations including buffering and overlay are found in many geographic information systems. Research has shown that breaks often occur in clusters. Spatial analysis tools like Quadrat analysis, Thiessen polygon generation, spatial autocorrelation and fiture analytical techniques may offer new methods for predictive modeling. System design The GIS serves as the integrating technology for the other components. The relational database management system is available for tabular and spatial attribute storage. The database is accessible via the facility management front end, and through the GIS tools. Statistical and neural network tools that require more than just tabular data retrieve these data via GIS access methods. Attributes are linked to the spatial entity/item. Geographic information systems typically provide a wide range of display capabilities. These display functions are closely related to the database contents. Thematic mapping (shading) can be performed based on different variables. Different layers can be overlaid. Geographical measurements can be performed on the screen. Interactive exploratory selection and query is supported to inspect the attributes of particular features. Conclusions The development of a decision support tool for facility management planning that incorporates predictive modeling tools will be of significant value to managers and planners. The functionality outlined in this paper will allow managers and planners to analyze the effects of many different factors on failure and failure rate. Developing a comprehensive CMMS is the foundation for giving the end user the ability to build on and develop an accurate Predictive Modeling system. Using and integrating GIS as the central component provides a structure that can facilitate the operation of the different tools necessary to implement a Predictive Modeling system. | |||||||||||||||||||||||||||||||||||
|
|