Data Science

Predictive Maintenance Services and Solutions - Overview

Navdeep Singh Gill | 17 December 2018


XenonStack White Arrow

Thanks for submitting the form.

Overview of Predictive Maintenance

Predictive maintenance is critical in diverse application areas, such as manufacturing industry, information, and technology, aerospace, heavy-machinery industry, etc. to estimate the future performance of a subsystem or a component to make RUL(Remaining Useful Life) estimation. If we can accurately predict when an asset (which may be any machine or a server or a turbine) will fail, then we will be able to make planned maintenance decision in advance to avoid sudden failures, reduce the maintenance cost, as well as streamline operational activities. This may seem a bit theoretical, but this is feasible using modern-day data technologies available. We can monitor our assets(can be any device or machine, which may lead to breakdown) in real-time via sensors and based on the sensor data patterns we can predict when we are going to have a break-down of assets. Preventive Maintenance is the maintenance that is performed on a regular basis on a piece of equipment to reduce the possibility of breaking down or increasing the lifetime of the equipment. Such maintenance is scheduled based on the usage or on time lapsed from last maintenance. A typical example of time-based preventive maintenance is an air conditioner which is serviced every year before the summer starts. Preventive maintenance is a primitive method of maintenance that is costly because we cannot predict the remaining useful life of an asset or equipment. We do maintenance checkups regularly based on usage or time lapsed from the last checkup, even when the maintenance is not at all needed. This is where the concept of predictive analytics comes in. Using IoT and machine learning, we can predict when an asset is going to break down, and hence we can perform a maintenance check-up on it when and in which part it is needed. Merits of Predictive Maintenance over Preventive Maintenance practices are:
  • Predictive maintenance helps in Identifying key parameters and determining the likelihood of breakdown outcomes.
  • Predictive maintenance helps in Optimizing business decision-making systematically using real-time and old relevant data.
  • Helps in better planning of spares inventory, so that the inventory or maintenance team doesn't face phases of pressure.

Data Analytics For Predictive Maintenance

Predictive maintenance depends upon the monitoring process of the equipment by performing actions on the device to achieve required objective. Data analytics for Predictive Maintenance is categorized into three:
  • Predictive
  • Descriptive
  • Prescriptive

Descriptive Analytics:

This particular analytics is performed to discover the actionable details of specific equipment using historical data of that particular equipment. It is used to improve the maintenance process and reduce the inefficiencies regarding the equipment operation. Examples are: Performance Degradation Detection: in this case, the performance of the equipment is monitored so that early warning of failure could be detected and further the equipment can be restored to a practical and cost-effective state. Maintenance effectiveness estimation: it is one of the statistical analysis of the equipment performed before and after the maintenance actions. This is useful for the customer to determine whether the maintenance process applied by him is effective or there is a need to change them.

Cases of Predictive Analytics:

It covers the prediction of future events like failures by learning from historical data. Techniques include Event-based Failure Prediction, Sensor-Based Failure Prediction, and Model-based Failure Prediction. Event-based Failure Prediction: Event data of equipment is economical, so it is readily available for performing an analysis. The data is collected and transmitted to the operational database. After that, data mining is performed to find the significant occurrences of events within the specified time windows. Once it is done, temporal rules are defined to predict future failures. Therefore, the defined rules are applied to the incoming events to predict the potential occurrences of failures within a time window. Model-based Failure Prediction: this technique is used to predict the failures that do not frequently occur in the field. Therefore, physical models are developed to simulate those failures. The physical models are developed using mathematical equations representing the electrical and mechanical details of the equipment. These models include failure modes and degree of failures. Therefore, these models can easily be simulated with the data by respective specified failure nodes and degree of failures. Sensor-based Failure Prediction: sensor data is beneficial to obtain information about the pre-failure conditions. To implement it model is developed by training it using historical sensor data. Once done, the model is applied to real-time data to predict the failures.

Two approaches are available:

Failure Prediction using anomaly detection: the deviation from the standard behavior of the equipment is determined by applying the trained model. The model is trained using the historical data where a large number of measurements are related to the normal operation, and few of them represents the incidents of failures. At the time of detection, an anomaly score for each new measurement is calculated on the basis on the distance of measurements from the previous learning clusters. If any unusual score is obtained, it is concluded as a failure. Failure Prediction using classification: this particular method is used to predict the failure of complex data. In this case, a model is trained using historical data which consist of sufficient samples of historical incidents for each type of failure. When the obtained model is applied, the partitions the data into normal and pre-failure data and further classifies them binomial into various classes of failures.

Prescriptive Analytics

It includes the generation of recommendations by results obtained so that failure rates could be reduced while meeting the operational objectives. While performing the maintenance of equipment, it becomes difficult to determine and understand the global impact of applied technique. Therefore, an overall optimization framework is developed that considers the outputs of the predictive maintenance algorithms along with different cost estimates and operation constraints and recommends a complete maintenance plan that results in maximum operation efficiency and minimum maintenance cost.

Hybrid approach for Predictive Maintenance

It is the combination of the physical model and statistical analysis part of events and operations performed by equipment. A physical model can simulate normal and faulty behavior. Once a statistical model is learned from simulated and field data, it can predict the severity of fault mode over real-time data. When the failure is predicted by assets, the reactive maintenance is performed by producing an alert that motor must be repaired or replaced based on the type of failure.

Challenges Summary

Predictive analysis poses various challenges such as :

Predictive Maintenance Solution Stack

We have sensors attached with assets to monitor critical parameters in real-time such as temp, voltage, CPU usage, etc. Process the data from sensors and analyze them in real-time to identify and report critical events without wasting any time (e.g., sudden increase in temperature, irregular voltage change) and take preventive measures accordingly. After processing the data, store the data in the data warehouse so that analysis can be on that data for predictive analysis to compute RUL(Remaining Useful Life) estimation of assets. Building a machine learning algorithm to predict the breakdown condition in advance from the collected data. Automate all the above steps in a continuous pipeline, without any need of human help. Building a data visualization tool to visualize all the collected data in a human-readable format.

Critical parameters and features of the real-time predictive system :

For the above implementation, our focus should be on identifying unique problems that affect operational and production impacts while managing risk. It’s important to understand what metrics the organization is focusing on. Answering these questions to determine the critical goals of a PdM project and ensure completion: The predictive maintenance solution provides event-based prediction by learning from historical data as well as learning from the new real-time data. Prediction based on the data directly from sensors and learning the normal behavior of assets and identifying any abnormal behavior. Prediction of fail events using classification models for different failure events.
  • Collection of a massive amount of real-time data from multiple assets at different geographical locations in different data forms.
  • Storing huge amount of data in a structured manner so that it can be used for analysis via machine learning algorithms.
  • Selection and training of an efficient machine learning model (for various assets), over the different data source.
  • Building a scalable data pipeline, that is strong enough to accept huge amounts of data and process it in a structured way.
  • What significant assets are likely to fail? When and why do we accept they will fail
  • How would the assets’ crash impact organization, operations or product costs? What does downtime cost?
  • How can data-driven decisions be integrated within the constraints of your existing maintenance practices?

Predictive Analysis Architecture

For Iot based predictive analysis, we have sensors to monitor data at different locations, all the assets namely A1, A2, A3, A4 are connected to a common network via the CoT(cloud of things). The cloud of things transfers the real-time data to nifi which is data ingestion software which via Kafka creates a data stream that can be consumed directly. Now we had two options available to process the stream of data:
  • Batch processing
  • Stream processing
In Batch processing, we store the data in a data warehouse and apply machine learning algorithms and another analysis tool later on to predict the breakdown of our assets. In-Stream processing, the data stream is consumed by spark streaming services and it analysis the real data time and based on the real-time data it does predictions and identifies any breakdown events and reports them. The data is then stored in the data warehouse. At last, we have a dashboard that can be used to visualize all our data and gather business intelligence.

Data Collection and Data Processing Solution:

Data Collection and Data Processing Solution

Predictive Maintenance Analytics Pipeline

Identifying and prioritizing available data sources

The increase in asset connectivity and use of smart devices may have generated large volumes of available data. It is not required or advised to address this entire universe of possible data. Instead, begin to predict failures on a single asset by focusing on the usable, existing data sources related specifically to it.

Collecting targeted data

The targeted data may reside in remote locations and get it to the analysis pipeline includes sensors, meters, supervisory control, etc. An efficient solution has to be flexible enough to collect data from all of these remote data sources to learn and continually make better, more informed business decisions.

Determining analytics pipeline

Establishing an advanced analytics pipeline based on the specific operation. For example, Cloud analytics should be balanced enough to reduce the burden of streaming perishable PdM data on our cloud deployment. A distributed approach is followed to detect and respond to local events at spark streaming step as they happen, taking action immediately on streaming data, while simultaneously integrating additional data sources in the cloud. Spark Streaming is used to analyze streaming data in memory for real-time response and filter out unnecessary data rather than relaying it to the cloud. Early detection of performance degradation by monitoring the changes in key performance indicators over time compared to the ideal performance. Assessment of the effectiveness of maintenance activities and actions by comparing before and after the performance for better maintenance planning.

Predictive Maintenance

Reactive: Alerting on the basis of asset repaired or replaced. Prescriptive Analytics: Operating recommendation and maintenance optimization. Classification Approach: this approach is used to predict the possibility of failure in next specified steps. Regression Approach: this approach is used for the prediction of RUL (Remaining Useful life), i.e., prediction of time before the next failure. The Deep Learning and Machine learning models for predictive maintenance:
  • Deep Belief Network
  • Autoencoders
  • Recurrent Neural Network (LSTM)
  • Convolutional Neural Network
  • Restricted Boltzmann Machine
  • Consensus Self-Organizing Models (COSMO)

Combining and analyzing data to gain precise insights

We start by analyzing available data to establish the parameters of normal operation for a machine. This enables the formulation of rules through condition monitoring for analyzing the real-time data coming directly from device sensors on the spark streaming via the data collecting pipeline. After analyzing the real-time data, we added historical and third-party data such as reliability models and logs to uncover meaningful correlations, patterns, and trends with the anomalies generated by the real-time data rules, to signal potential failures. These patterns can be used to further refine our rules and offer actionable insights in real time.

Operationalizing and preventive actions

Turning insights provided by data analysis into deliverable action by integrating a risk assessment for all assets into our operation through a single dashboard. For example, in case a potential problem is uncovered, the analytics dashboard triggers an event which allows us to send out automated alerts to concerned parties, such as location, estimated replacement components and suggested corrective action to avoid a catastrophic event. Then, by capturing wear characteristics data from the replaced parts, we can continuously refine our machine learning models and learn from the new performance insights.

Concluding Predictive Maintenance

  • We can Reduce unplanned downtime of assets to 3.5%
  • Improve the quality of equipment effectiveness
  • Reduce maintenance cost of assets
  • Increasing ROI on assets, hence increasing profits

How Can XenonStack Help You?

XenonStack follows the Test Driven Development Approach in the development of Enterprise-level Applications following Agile Scrum Methodology.
  • Text, Predictive and Statistical Modeling & Analytics: Our Data Science Services provides text analytics solutions to extract data from text sources for sentiment analysis, text clustering and categorization using Natural Language Processing algorithms. Predict the Future, anticipate the change, nurture agility that boosts your bottom line. Build Predictive and Statistical Models for Data Products.
  • WorkBench For Deep Learning Services & Machine Learning Services: At XenonStack we offer Data Science Solutions for Data scientists and researchers to Build, Deploy Machine learning and Deep Learning algorithms at scale with automated On-premises and Hybrid Cloud Solutions.
  • Decision Science Solutions: XenonStack Data Visualization & Artificial Intelligence Services enables you to develop Interactive Dashboard, Data Product in form of Chat Bots, Digital Agents and Decision Science framework, recommendation systems with progressive Web Applications.

Transform your
Enterprise With XS

  • Adapt to new evolving tech stack solutions to ensure informed business decisions.

  • Achieve Unified Customer Experience with efficient and intelligent insight-driven solutions.

  • Leverage the True potential of AI-driven implementation to streamline the development of applications.