For 2017, the American Society of Civil Engineers has graded the US drinking water and wastewater infrastructure with D and D+ grades, respectively. The Environmental Protection Agency prescribes a simple solution: a $271 billion investment in waste water infrastructure. Despite the critical role of water and wastewater infrastructure in our daily life, the issue does not interest or motivate voters. Absent a major disaster, the likelihood of large scale long-term capital investment in critical water and wastewater infrastructure is remote.
Less than 5% of US federal funds spent on infrastructure is allocated to water services. On an annual basis, local governments spend $20 billion on sewer capital (CapEx) and further $30 billion on Operation and Maintenance (O&M). This article explores the potential impact of Machine Learning for Asset Maintenance on critical water and wastewater expenditures. Specifically, we will address how the use of Big Data can optimize O&M spending.
We do not suggest that critical infrastructure investments should be delayed or minimized. Machine Learning for Asset Maintenance is complimentary. For the purposes of this article, we relate to the case of aging US infrastructure. However, the principles we address can be applied to infrastructure in other regions including Western Europe and Asia/Pacific.
The Background: Industry 4.0
The fourth industrial revolution has been named Industry 4.0 or the Smart Factory. At its core, Industry 4.0 operationalizes vast quantities of data that is generated by production machines’ sensors. McKinsey & Company has broken down Industry 4.0 into 25 levers including Smart Energy Consumption, Automation, Concurrent Engineering and Predictive Maintenance.
Although Industry 4.0 has wider applications to the industry, our focus is on the unrealized opportunity of Machine Learning for Predictive Maintenance or Industrial Analytics.
How did Machine Learning become part of Asset Maintenance? Until recently Machine Learning was a topic mostly confined to academia. However, in the past few years, there has been a significant move to commercialize the discipline. In the past, there were significant hurdles with the capture, storage and transmission of sensor data captured by production assets. These costs have all fallen substantially, along with the cost of computational power. Specifically, advances in cloud technologies has made access to Industrial Analytics solutions more affordable and accessible.
The confluence of lower costs and greater accessibility gained the attention of government officials, senior executives from leading industrial entities and forward-looking technology companies (both established and startups).
Until now, most of the data generated by sensors embedded in machine equipment was unused. The exception is SCADA-based monitoring of high priority sensors (e.g., temperature or vibrations). Plant maintenance staff monitor this data to track the health of a machine. The principle is based on determining whether human-set control thresholds have been breached.
Let’s use a specific example – SCADA monitoring of the temperature of a specific machine. If the limit is set at 40 degrees and the machine temperature exceeds 40 degrees, this over-heating will trigger further actions.
Machine Learning for Industrial Analytics use different principles and are not (human) rule-based. An algorithm is trained to detect abnormal data patterns (or correlations of patterns) regardless of whether the control thresholds have been breached. In the case of SCADA, only breaches above 40 degrees and below 20 degrees are monitored. With Machine Learning, algorithms are looking for behavior patterns even within the thresholds.
Manually configured rule-based alertsSource: Presenso
Advanced Artificial Intelligence, enable the measurement of Mean Time to Failure (MTTF) and the application of Root Cause Failure Analysis (RCFA).
Machine Learning is a dynamic field and there are multiple methodologies used. Below is a high-level review of two of the most common approaches:
1) Supervised Machine Learning. An algorithm is “trained” on the underlying production asset by using data labels or classification. Once it is trained, the algorithm can apply the classification to new data. For instance, if it is trained on machine failure, it applies this training on new data.
2) Unsupervised Machine Learning. Data labels are not provided to the algorithm. Instead, vast amounts of data are analyzed and the algorithm itself generates the label.
In general terms, Supervised Machine Learning is more resource intensive as data labelling requires intensive human input which can slow the deployment of an Industrial Analytics solution.
How Industrial Analytics uses Big Data to Optimize O&M Spending
Approximately half of the annual expenditures in the wastewater sector are allocated to O&M and this share is rising.
Within the industry, there is an over-reliance on run-till-failure maintenance. Equipment is used until there are indications that failure is imminent or has already occurred. With run-till-failure, the plant uses a Reactive Maintenance approach and repair crews need to identify the cause of the problem with limited guidance.
Reactive Maintenance is the most expensive maintenance approach. Repairs crews may be selected from other less urgent jobs and may need to travel to the worksite. Furthermore, with Reactive Maintenance the incidence of work-related injuries rises.
Injury Accidents versus Maintenance Planning
Source: Wim Vancauwenberghe, Director BEMAS (Belgian Maintenance Association)
The chart above depicts the impact of Reactive Maintenance on employee safety. When 75% of maintenance is Reactive, the incidence of Maintenance Technician lost-time injuries is 12 times higher than when less than 25% of maintenance is Reactive.
Reducing asset downtime is the most important objective of Industrial Analytics. This occurs when the Machine Learning algorithm identifies anomalous sensor behavior and provides early warning of asset degradation or failure. Furthermore, by providing facility technicians with Root Cause Failure Analysis, less time and resources are wasted with trial -and-error and Wrench Time is optimized.
With few exceptions, it is difficult to calculate downtime for most industries. In the Water and Wastewater industry, the cost of downtime cannot simply be measured by lost production and wasted resources. Given the complexity of the machinery at industrial plants the culture of run-till-failure, and human errors add an additional dimension to downtime and O&M overhead.
The following is one example of the unintended consequence of current O&M practices within the industry. In February 2017, flooding at Seattle’s West Point Treatment Plant in Magnolia’s damaged an electrical circuit that led to a plant shutdown. According to the Seattle Times, the local county “has dumped an estimated 235 million gallons of untreated wastewater — including 30 million gallons of raw sewage and hundreds of tons of partially treated solids into Puget Sound.” An investigation has attributed this disaster to “errors in judgment, poor communication, a lack of training, equipment failures and faulty maintenance.”
The experience from these types of environment disaster demonstrates the importance of preventing human error by using Machine Learning for diagnostics and analytics.
Concluding Thoughts: Can Machine Learning Extend the Asset Life of Plant?
With aging infrastructure, there are different ways to extend asset life. With Machine Learning, industrial water and wastewater plants can repair equipment and machinery before it fails, thereby avoiding run-till-failure scenarios where assets can be further damaged.
However, Machine Learning alone is not the solution or excuse delaying investment decisions that impact the health and safety of local populations. Effective O&M practices are only one component of an overall infrastructure plan.