Predictive maintenance for asset-heavy portfolio companies
Tldr;
Economic headwinds are forcing portfolio companies to tighten cost controls. For organizations with capital intensive physical assets, this is a difficult challenge because equipment failures often need to be dealt with immediately
Predictive maintenance solutions built upon unsupervised anomaly detection models can help operators take a proactive approach to asset investment rather than being reactive to failures and malfunctions
These solutions are based on anomaly detection algorithms and help to drive enterprise value growth from both a cost and revenue perspective
Reduce costs: These solutions can identify direct cost reduction opportunities, they enable proactive maintenance (reducing the number of expensive and catastrophic equipment failures), and they enable better capital expenditure forecasting
Stabilize revenues: They can help reduce fluctuations in asset throughput, maximizing the throughput of the system, and ultimately the revenue that the asset generates
Want to learn more about applying predictive maintenance? Get in touch
In today's macroeconomic environment, portfolio companies across industries face mounting pressure to maintain tight cost controls and sustain revenues. While some business models benefit from straightforward and predictable cost structures (e.g. software), others present complexities that require careful consideration and significant time to understand. In particular, portfolio companies operating capital-intensive physical assets encounter capital expenditure that is critical to their operations, unpredictable at times, often urgent, and entangled. Industries such as manufacturing, energy, transportation & logistics, mining, telecommunications, construction, utilities, aerospace, and chemical & pharmaceuticals are among those that commonly face challenges in this regard.
In these sectors, when critical equipment fails, such as a pump in an oil refinery or the electronic system in a cargo plane, companies must act swiftly to repair or replace the assets. The downtime of physical assets directly impacts their throughput and consequently, their revenue. Thus, it becomes imperative for these companies to minimize asset downtime. To do so, they often invest substantially in monitoring systems that provide real-time updates on their assets and in many cases, they also enlist the expertise of dedicated consultants to design and implement business continuity plans to mitigate the damage from operational disruptions.
While these investments are crucial and provide value in times of crisis, they are reactive approaches. By leveraging machine learning (ML) and the same data that is already collected from their assets, companies can proactively manage their capital expenditures and optimize asset performance. This Perspective explores the value of applying ML technologies to increase enterprise value for asset-heavy portfolio companies. We explore this topic from the lens of a private equity firm, hence the common reference to portfolio companies. We will delve into how models, built on the foundation of sensor data, can empower companies to make proactive decisions and how this translates into greater enterprise value.
Predictive Maintenance Using Machine Learning
By collecting data from various asset sensors and combining it with contextual information such as weather conditions, ML enables companies to identify early warning signs for equipment failures and malfunctions. In turn this gives operators more time to respond to issues and in many instances it lessens the severity of the equipment malfunction. From an ML perspective, predictive maintenance can technically follow an unsupervised or supervised approach. However, from our experience building these solutions, we find that companies seldom have the structured records of previous equipment malfunctions needed. At best, they have some structured records but greatly overestimate the volume and usefulness of them. With large datasets of sensor readings, unsupervised models, namely anomaly detection1, tend to produce the best result and be the most feasible approach.
The output of these models are highlighted instances in sensor readings that deviate from their typical pattern and are likely to indicate that some kind of malfunction has occurred or that an equipment failure is likely to take place in the future. For example, in an oil refinery, anomaly detection can help identify unusual temperature or pressure readings in critical equipment, such as pumps or compressors. If the sensor data indicates a deviation from the expected operating range under its current conditions, technicians can be alerted to investigate the potential malfunction and take preventive measures before a catastrophic failure occurs. Several ML algorithms like clustering2, density-based methods3, and autoencoders4 can all form the basis of anomaly detection solutions. In other instances, anomalies can indicate abnormal and costly operating conditions. While working with a commercial real estate operator, we built a solution to analyze sensor data coming from their HVAC systems across their properties. After applying anomaly detection, we identified a number of units with an abnormally low amount of downtime. After having a technician go investigate the uncommon behaviour, the company realized that faulty equipment was preventing the HVAC system from turning off. Not only was this unnecessarily wearing down the equipment, but it was also costing them millions of dollars each year.
How predictive maintenance improves the enterprise value of asset-heavy businesses
Like all technology implementations, portfolio company managers should be asking themselves how investments in a predictive maintenance solution leads to an uplift in a company’s enterprise value. In the case of asset-heavy businesses, predictive maintenance can help to reduce their capital intensity, enabling them to demand higher valuation multiples compared to industry peers.
Firstly, predictive maintenance enables this because it allows operators to better manage their costs. In some cases, like the HVAC example above, companies directly reduce the cost to operate their assets simply by better understanding the operating conditions. In other cases companies can proactively resolve and repair small malfunctions before they manifest into larger and more expensive failures. By detecting early warning signs of equipment deterioration, companies buy themselves time to make informed repair or replace decisions rather than being pressured to replace critical equipment as soon as possible at a premium price point. Finally, it also leads to better cost forecasting capabilities. These models can be used to estimate the health of their assets on an ongoing basis and help them to decide whether it is more efficient to repair an asset or replace it early along with other equipment to negotiate better pricing from vendors.
Predictive maintenance also helps reduce asset downtime which in turn maximizes an asset’s production capacity. By identifying potential equipment failures in advance, companies can take preventive measures (e.g. part repairs and servicing) to ensure consistent throughput and uninterrupted production. A more stable production schedule allows companies to smooth out their revenues and de-risk their operations.
Predictive maintenance solutions offer a strategic advantage to companies managing expensive physical assets. They reduce capital expenditure, enable maintenance schedules to be optimized, and ensure throughput stability, culminating in stronger financial performance and a higher enterprise value. Ultimately, they give management more flexibility to navigate their complex cost structure and in a turbulent economic environment, time is almost as important as the assets that these companies manage.
Notes
1 Anomaly detection aims to identify abnormal states of equipment operation by detecting deviations from normal patterns in sensor data.
2 Clustering: Clustering algorithms, such as k-means or DBSCAN (Density-Based Spatial Clustering of Applications with Noise), group similar data points together based on their features. Deviations from the established clusters can indicate anomalies in the sensor data. For example, if a cluster represents normal equipment behavior and a data point falls outside any cluster, it could signal a potential anomaly.
3 Density-Based Methods: Density-based anomaly detection algorithms, like Local Outlier Factor (LOF) or Isolation Forest, measure the local density of data points to identify those with significantly lower density, indicating potential anomalies. These algorithms can capture anomalies in data that do not conform to the expected density distribution.
4 Autoencoders: Neural network models trained to reconstruct their input data. By learning the patterns of normal sensor readings, an autoencoder can reconstruct the data accurately. When presented with an anomalous data point, the reconstruction error is typically higher, indicating a potential anomaly.