One accurately predicted failure of a large asset is worth more than $100,000 in many industries. The latest research from global market intelligence provider IoT Analytics highlights, among many other things, that the median unplanned downtime cost across 11 industries is approximately $125,000 per hour. With critical unplanned outages in facilities in industries such as oil and gas, chemicals, or metals occurring several times a year, an investment into predictive maintenance can amortize with the first correct prediction.
Unfortunately, there is a flip side: the accuracy of many predictive maintenance solutions is lower than 50%. This creates headaches for maintenance organizations that often run to an asset to find it operating normally, eroding trust in the entire solution.
That said, vendors have made strides to increase prediction accuracy with more data sources and better analysis methods becoming available, including AI-driven analysis. There are positive signs that this determination for better prediction accuracy is helping end users: our research indicates that 95% of predictive maintenance adopters reported a positive ROI, with 27% of these reporting amortization in less than a year.
Knud Lasse Lueth, CEO at IoT Analytics, commented:
“Predictive maintenance continues to be one of the leading use cases for Industry 4.0 and digital transformation, especially in process industries where asset failures can quickly go into the hundreds of thousands of dollars. It is great to see that the market is moving ahead with AI integration into existing APM and CMMS systems and that prediction accuracies are improving. Nonetheless, we still have a long way to go as false alerts remain commonplace.”
General search interest in predictive maintenance and related concepts has been on the rise for the last 12 years. Online searches for the term have grown nearly threefold since we initiated coverage on the topic in 2017 and have outgrown condition-based maintenance and asset performance management (APM) related searches.
Indeed, predictive maintenance appears to be well on track to be the must-have killer application we made it out to be in 2021.
Fernando Brügge, senior analyst at IoT Analytics, added:
“Predictive maintenance is reaching new heights of maturity and sophistication thanks to the rapid advancements in artificial intelligence, hardware, and data engineering. We are at the point where these technologies enable us to collect, process, and analyze massive amounts of data from multiple sources, and use them to build more accurate and reliable models of machine health and behavior, as well as to determine potential courses of action to fix machine issues. In this way, predictive maintenance is not only a smart way to optimize equipment performance and lifecycle, but also a strategic way to enhance operational efficiency and competitiveness in a rapidly evolving industrial space.”
In this fourth installment of our predictive maintenance market coverage, we look at five important highlights to note about the market going into 2024:
Highlight #1: The predictive maintenance market is valued at $5.5 billion
The predictive maintenance market reached $5.5 billion in 2022. Uncertain economic conditions and other manufacturing priorities in the last two years resulted in 11% market growth between 2021 and 2022. With companies reinvesting in efficiency, safety, and operational performance, we expect the market for predictive maintenance to grow to 17% per year until 2028.
Our research indicates that industries with heavy assets and high downtime costs are driving the adoption of predictive maintenance solutions (e.g., oil & gas, chemicals, mining & metals).
Highlight #2: There are 3 different types of predictive maintenance, with anomaly detection on the rise
As the market has evolved, three noticeable predictive maintenance types have developed: indirect failure prediction; anomaly detection; and remaining useful life (RUL). The difference between these largely comes down to the objectives, methods of data analysis, and type of output/information they provide. RUL is the hardest to achieve due to resource demands and environmental factors that make it difficult to scale. Indirect failure prediction has been the most used approach, but our research indicates that anomaly detection is on the rise.
1. Indirect failure prediction
The indirect failure prediction approach generally takes a machine health score approach based on a function of maintenance requirements, operating conditions, and running history. This approach often relies on general analysis to yield this score, though supervised machine learning methods can be used if a significant amount of data is available.
- Scalability. Indirect failure prediction can be more easily scaled since they rely on equipment manufacturers’ specifications that are more or less the same across machines of the same type.
- Cost effective. Indirect failure prediction can use existing sensors and data, reducing the need for additional instrumentation.
- Failure time-window accuracy. Indirect failure prediction does not give a timeline of when machines will fail. This can be a problem for organizations with very costly downtimes (e.g., heavy equipment industries).
- Dependent on historical data. Indirect failure prediction’s effectiveness relies on the availability of extensive historical data for accurate modeling.
2. Anomaly detection
Anomaly detection is the process of finding and identifying irregularities in the data (i.e., data points that deviate from the usual patterns or trends). While the indirect failure prediction and RUL approaches use failure data to predict future failures, anomaly detection uses the “normal” asset profile to detect deviations from the norm. These deviations can indicate potential problems, such as faults, errors, defects, or malfunctions, that need to be detected and addressed before they cause serious damage or downtime. This approach makes it easier when there is not a good repository of failure data, and it often relies on unsupervised machine learning.
- Low data and hardware requirements. Anomaly detection models can identify issues without being trained on failure data. Further, since these models need less data, they do not demand high computing power.
- High scalability and model transferability. Anomaly detection models are trained on normal operation data, so they can easily be applied to different machines without retraining or adaptation.
- Failure time-window accuracy. As with indirect failure prediction, anomaly detection models do not give a timeline of when machines will fail, which can be a problem for organizations with very costly downtimes.
- Presence of false positives. While most solutions in the market can distinguish between critical and noncritical anomalies, the choice of unsupervised machine learning models is still important as it can affect how well this distinction can be made (e.g., autoencoders and generative adversarial networks do not capture the complexity of normal operations).
3. Remaining useful life (RUL)
RUL is the expected machine life or usage time remaining before the machine requires repair or replacement. Life or usage time is defined in terms of whatever quantity is used to measure system life (e.g., distance traveled, repetition cycles performed, or the time since the start of operation).
This approach relies on condition indicators extracted from sensor data—that is, as a system degrades in a predictable way, data from the sensors match the expected degradation values. A condition indicator can be any factor useful for distinguishing normal operations from faulty ones. These indicators are extracted from system data taken under known conditions to train a model that can diagnose or predict the condition of a system based on new data taken under unknown conditions.
Predictions from these RUL models are statistical estimates with associated uncertainty, resulting in a probability distribution.
- Failure prediction time-window. RUL is especially useful for industries where maintenance is very costly and needs advanced planning, such as heavy-equipment industries.
- Output robustness. Since RUL estimates rely on high-quality and detailed data, they tend to be more robust and reliable.
- Resource demand. Training large models requires powerful computing hardware, especially if done on-premises.
- Model transferability and scalability. Different environments and usage patterns can cause different failure modes for the same type of equipment. This means the model needs to be retrained for each specific case, reducing its scalability and generalizability.
Highlight #3: Predictive maintenance software tools have 6 common features
Software is the largest segment of the predictive maintenance tech stack, making up 44% of the predictive maintenance market in 2022. Our report shows that even though most successful predictive maintenance software vendors specialize in industries or assets, there are six common features between their various solution software suites:
Feature 1: Data collection
Data collection tools within predictive maintenance software collect, normalize, and store data on asset health/condition parameters. They also collect other data types needed to identify and predict upcoming issues, such as business and process data.
Feature 2: Analytics and model development
Analytics and model development tools within predictive maintenance software analyze, interpret, and communicate data patterns, including analytics discovery (e.g., RCA, AD modules) and modeling (e.g., feature engineering and model selection and testing).
Feature 3: Pre-trained models
Pre-trained models are just that: ready-to-use models typically designed for specific assets in specific industries. These models include capabilities and references for specific assets or failure modes (e.g., fouling for heat exchangers, wear and corrosion for fans, or valve leakage for compressors). These are meant to help end users see examples of models so they can build on them or develop custom predictive maintenance algorithms.
Feature 4: Status visualization, alerting, and user feedback
Status visualization, alerting, and user feedback tools within predictive maintenance software automatically communicate asset-related data/insights for various personas. These insights often include status dashboards and automatic alerts that trigger work orders or corrective actions, maintenance planning, and optimization. These tools also enable users to provide feedback concerning the accuracy of alerts.
Feature 5: Third-party integration
Third-party integration enables users to connect their predictive maintenance software to other software systems and workflow management tools, such as ERP, MES, CMMS, APM, and Field Service.
Feature 6: Prescriptive actions
Prescriptive action features typically suggest the optimal actions to take in case of an (upcoming) failure. These actions are typically prioritized by criteria that are set when the algorithm is designed. The actions that are prescribed by the software vary depending on the nature and urgency of the issue. They may require multiple steps or interventions. For instance, some actions may involve automatically adjusting the equipment parameters or informing the maintenance and operation teams about the necessary procedures to ensure equipment efficiency.
Highlight #4: Integration into the maintenance workflow is becoming important
In its early days, predictive maintenance was mostly a standalone solution developed by startups to address specific customer needs. However, our report highlights a notable trend of sophisticated predictive maintenance solutions integrating into larger APM and computerized maintenance (CMMS) solutions.
APM is a strategic equipment management approach designed to help optimize the performance and maintenance efficiency of individual assets and entire plants or fleets. APM aims to improve asset efficiency, availability, reliability, maintainability, and overall life cycle value.
Various APM vendors are introducing predictive maintenance software tools within their APM offerings. The solutions aim to tie the different capabilities into one thread:
- Knowing when a machine will fail and mapping how failures could affect production or output
- Estimating how much fixing or preventing an issue will cost
- Making recommendations on whether it is worth fixing or preventing a problem
By including a sophisticated predictive maintenance solution in an end-to-end asset flow, APM players are trying to become the main partners for their customers’ digitalization journeys.
Highlight 5: Successful standalone solutions vendors specialize in an industry or asset
Our research found that 30% of predictive maintenance vendors offer standalone, industry- or asset-specific solutions. By tailoring their efforts to specific niches in which they have acquired domain knowledge, they can discern the types of equipment and industries in which their solutions offer the most end-user benefits.
Download a copy of the full report from IoT Analytics to learn more.