Blog | xLM

#016: Can your PM do this? - Part II

Written by Nagesh Nama | 06.19.24

#016: Can your PM do this? - Part II

Is your PM predictably intelligent? Introducing ContinuousPdM!

Note: This is Part 2 of the ContinuousPdM series. Part 1 of this series can be found here.

There are several steps in developing a Predictive Maintenance model. These steps are discussed in detail below.

Data Gathering

Leveraging data from diverse sources enhances the quality of predictive maintenance immensely. Predictive maintenance offers a significant advantage over preventive maintenance by minimizing unplanned downtime through the anticipation of failures and the prevention of both over-maintenance and under-maintenance. The effectiveness of this approach heavily relies on the quality of data harnessed by the machine learning-based artificial intelligence tool.

Sensors

For predicting failures in the future, predictive maintenance utilizes historical data from sensors which measure values for a multitude of parameters such as:

Vibration

One of the most widely used sensor across the PdM industry especially for rotating machines is the Piezo accelerometer or the MEMS accelerometer, both of which are used for vibration measurements. The data from these sensors can be used to find target faults in bearing condition, gear meshing, misalignment, load condition, etc.

Sound

Ultrasonic microphones are used to record sounds at high frequencies from machinery, and can be used to detect pressure leaks, imbalance etc.

Temperature

Sensors such as RTDs, and thermo-couplers are used to measure process temperatures, for example. Other parameters can include bearing temperature, lubrication temperature, cooling water temperature, etc..

These were a few of the sensors from which data is usually gathered when we consider the context of predictive maintenance. Data from various other sensors such as oil quality, magnetic field, operational field, proximity, electrical and many more can be gathered to predict faults in machinery.

Maintenance (CMMS) Data

A consolidated archive of historical data consisting of maintenance tasks performed on an asset over time is crucial for modeling. This can comprise of past records component replacements, duration of downtime due to component failures, can be used for failure pattern recognition, Remaining Useful Life (RUL) estimation, and maintenance optimization.

Data Pre-processing

Data from a multitude of sources is used to prepare a machine learning model which is capable of performing predictive maintenance, and hence it is imperative to make sure that the data is well-aligned and organized to the chosen model. Various pre-processing techniques need to be applied in order to clean the data, ensure its consistency and is properly formatted.

Handling Missing Data

Deletion Techniques

This technique removes entire rows or records that contain any missing values. It is simple but can lead to loss of a significant amount of data and reduced statistical power.

Imputation Techniques

Missing values can also be replaced using mean/median imputation where the missing data is replaced with either mean, median or mode of the available data. Regression models can be used to perform Regression Imputation to handle missing values or Multiple Imputation techniques such as Multivariate Normal Distribution, Chained Equations (MICE) or algorithms such as Random Forests, k-nearest neighbors, or neural networks can be used to impute missing values.

Model-based Techniques

Maximum Likelihood Estimation models estimate missing parameters by estimating the likelihood function of the observed data without imputing the missing data. Expectation-Maximization (EM) Algorithm is an iterative method that finds the maximum likelihood estimates of parameters in probabilistic models with missing data.

Handling Outliers

Statistical methods, like calculating z-scores, standard deviation, median absolute deviation, or determining the inter-quartile range, are valuable tools for detecting potential outliers that fall beyond the normal range of a particular parameter.

Machine learning algorithms such as Local Outlier Factor (LOF) which computes the local density deviation of a data point with respect to its neighbors to identify outliers, along with Isolation Forests and one-class SVMs can be used. Ensemble methods which combine multiple such algorithms can also be used in tandem to improve accurate outlier identification.

While outliers can notably decrease model accuracy across various datasets, anomalies in parameter values are essential for predictive maintenance tasks. Therefore, it is imperative to approach statistical techniques with caution.

Data Normalization

Some of the commonly used normalization techniques in predictive maintenance include z-score normalization which scales features to have a mean of 0 and a standard deviation of 1. Other techniques used for normalization are log transformation, decimal scaling, and min-max scaling.

Feature Engineering

Time based Feature Engineering

Lag Analysis

This technique involves creating new features by shifting the time series data by a certain number of time steps (lags). This captures the temporal dependencies and autocorrelation in the data. It also involves calculating statistical measures (mean, max, min, etc.) over a rolling window of time steps. This can capture short-term patterns and trends.

Decomposition

Time series decomposition is a technique that breaks down a time series into its constituent components, typically trend, seasonality, and residual components. This allows for better understanding and modeling of the underlying patterns and behaviors present in the time series data. The decomposed components can be used as separate features or combined in various ways to improve the performance of predictive models.

Stationarity Tests

Stationarity tests are statistical tests used to determine if a time series data is stationary or non-stationary. They are crucial before applying time series forecasting models, as most models assume stationarity. Common stationarity tests include the Augmented Dickey-Fuller (ADF) test and the Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test, which test for the presence of unit roots and stationarity around a deterministic trend, respectively.

Frequency based Feature Engineering

Autocorrelation Analysis

It measures the correlation between a time series and its own lagged values, revealing repeating patterns and seasonality. It helps identify the optimal lag order for modeling and forecasting time series data.

Seasonal Subplots

This technique involves creating subplots of the time series data grouped by a seasonal component (e.g., months, days of the week) to visually inspect and identify seasonal patterns and trends

Spectral Analysis

It involves transforming the time series data from the time domain to the frequency domain using techniques like Fast Fourier Transform (FFT). This allows identifying dominant frequencies, periodicities, and cyclical patterns in the data.

Feature Engineering for Classification

Numerical Transformations

These techniques transform numerical features to a different scale or distribution, such as log transformation, square root transformation, or Box-Cox transformation. They are used to make the data more suitable for certain algorithms or to reduce the influence of outliers.

Encoding Categorical Data

This involves converting categorical variables into a numerical representation that can be used by machine learning algorithms. Common techniques include one-hot encoding, label encoding, target encoding, and others, each with its own advantages and trade-offs.

Dimensionality Reduction

These methods aim to reduce the number of features in a dataset while retaining most of the relevant information. Popular techniques include Principal Component Analysis (PCA), t-SNE, Factor Analysis of Mixed Data (FAMD), and others. They can help with visualization, reducing noise, and improving model performance.


 

Machine Learning Modeling

In the realm of predictive maintenance, various machine learning models are commonly employed, such as decision trees, random forests, support vector machines, and neural networks like LSTMs. The selection of a model hinges on factors like data intricacy, type of problem, interpretability requirements, and trade-offs between performance and resources. Furthermore, the amalgamation of multiple models through ensemble methods has the potential to enhance accuracy significantly.

We adapt a 2-step approach to predict maintenance window in the near-future. The first step includes the prediction of parameter values in future, and the second step includes the prediction of the type of failure using the values predicted using time-series analysis.

STEP 1: Time Series Forecasting

Autoregressive Integrated Moving Average (ARIMA) models play a significant role in time series forecasting within the realm of predictive maintenance. These models excel at capturing trends, seasonality, and autocorrelation present in the data, making them a valuable tool for predicting future sensor values. On the other hand, Recurrent Neural Networks (RNNs), specifically Long Short-Term Memory (LSTM) networks, are adept at handling sequential data such as time series. Their strength lies in capturing long-term dependencies, rendering them highly effective in forecasting future sensor values.

STEP 2: Error Classification Modeling

Support Vector Machines (SVMs) stand out as robust supervised learning algorithms proficient in executing both binary and multi-class classification tasks. Within the realm of predictive maintenance, SVMs prove invaluable for categorizing equipment condition as either "healthy" or "faulty," as well as for discerning various failure types.

Moreover, decision trees and random forests play a pivotal role in predictive maintenance by categorizing equipment condition, pinpointing failure modes, and shedding light on the key features that contribute to failures.

Complementary algorithms such as Isolation Forest, One-Class SVM, and Local Outlier Factor find utility in anomaly detection, a type of unsupervised classification.

In the context of predictive maintenance, anomaly detection serves to flag abnormal equipment behavior or sensor readings, which could serve as early indicators of potential failures or deterioration.

 

Dashboarding and Reporting in Power BI

In our journey to revolutionize preventive maintenance, we’re developing powerful ways to visualize data using Power BI. These visualization techniques are part of our current updates and showcase potential insights and actionable data for various teams .

Here’s a closer look at how these visualizations can benefit engineers, operations, and facilities teams once they are fully developed.

Engineering Predictive Maintenance Dashboard

 

Our Engineering Predictive Maintenance Dashboard provides a glimpse into how engineers can filter and analyze data by date, machine category, and critical parameters such as Tool Wear, Torque, RPM, and Process Temperature. This visualization will be essential for:

Identifying Failure Modes

Engineers will be able to pinpoint potential failure modes in any component of the manufacturing line before conducting any testing or process changes.

Cost and Downtime Analysis

The dashboard features graphs displaying the total cost and downtime associated with selected parameters. By selecting specific data points, engineers can visualize the financial impact and operational downtime, helping them to prioritize maintenance activities and allocate resources more effectively.

Parameter Gauging

Gauge charts provide real-time insights into the selected product ID values against the average values for all products in the category. This comparative analysis helps in identifying anomalies and making informed decisions on maintenance needs.

Operations Predictive Dashboard

 

The Operations Predictive Dashboard is tailored for line technicians, focusing on minimizing downtime and ensuring smooth production flows.

Predictive Failure Analysis

The line chart showcases future trends of predicted failures, enabling technicians to anticipate and address issues before they escalate. This proactive maintenance approach reduces unplanned downtime and ensures continuous production.

Detailed Failure Insights

A table chart lists predicted failures with details like downtime, risk categories, and failure types. This detailed view allows operations teams to understand the root causes and take corrective actions promptly.

Parameter Condition Cards

Conditional formatting on parameter cards provides a quick visual indication of whether values are within acceptable ranges. Green indicates within 10% of the average, yellow for 10%-20%, and red for beyond 20%, enabling technicians to focus on critical areas needing attention.


Interactive Graphs and Reports in Power BI

Enhanced Data Analysis

Power BI’s interactive graphs help users get answers quicker and more accurately. Various charts allow for detailed analysis and visualization of data, making it easier to identify trends, anomalies, and patterns. By using slicers, users can filter data by future dates, risk categories, machine types, or parameter values to predict potential failures and their impacts.

Interactive Reports for Operations Dashboard

Users can interact with the reports by changing parameters to predict which component might fail due to specific risk categories and the predicted risk percentage. For example, by adjusting slicers, they can see how selecting different risk categories impacts the predicted failures and associated risks, providing a before-and-after snapshot of the potential outcomes.

Interactive Reports for Engineering Dashboard

In the Engineering Dashboard, users can see how changing parameters for a given date affects predicted failures, downtime, and costs. By using slicers, engineers can input future dates and parameter values to find and predict the failures, allowing them to make informed decisions on maintenance activities.

 

Facilities Management Benefits

 Facilities management teams play a crucial role in maintaining the readiness and availability of spare parts and managing the overall maintenance strategy. The ContinuousPdM visualizations will offer significant benefits:

Spare Parts Management

By predicting equipment failures and maintenance needs, facilities teams can optimize spare parts inventory, reducing excess stock and associated costs. This ensures that the right parts are available when needed, without over-investing in inventory.

Maintenance History and Aging Analysis

The dashboards integrate data on equipment age and maintenance history, providing a comprehensive view of each component's lifecycle. This information helps in planning and scheduling maintenance activities more effectively, ensuring that only necessary maintenance is performed, thus avoiding unnecessary downtime.

Investigation and Process Understanding

When there are changes in production quality, facilities teams can use the dashboards to quickly identify potential causes and investigate the process efficiently. This reduces downtime associated with lengthy investigations and helps in maintaining consistent production quality.

 

Key Performance Indicators (KPIs)

To effectively plan and manage preventive maintenance, our ContinuousPdM solution will incorporate essential KPIs and metrics, including:

  • Mean Time Between Failure (MTBF): The average time between failures for a particular asset, helping to predict when maintenance should be scheduled.

  • Mean Time to Repair (MTTR): The average time required to repair a failed asset, providing insights into maintenance efficiency.

  • Expected Next Failure Date: Predicts the next likely failure date based on historical data and current operating conditions.

  • Failure Rate: The frequency of failures over a specified period, allowing for trend analysis and proactive maintenance planning.

  • Maintenance Costs: Tracking the costs associated with both preventive and corrective maintenance activities.

  • Downtime: The total time that production is halted due to maintenance activities, crucial for evaluating the impact on productivity.

  • Spare Parts Inventory Levels: Monitoring the availability and usage of spare parts to optimize inventory management.

  • Compliance with Maintenance Schedules: Ensuring that maintenance activities are performed as per the predefined schedules to avoid unexpected failures.

 

ContinuousPdM - Delivered as a Managed Service

In every service we offer, the software app is continuously qualified. Also the customer's instance is continuously validated. In each run, 100% regression is performed.  In the case of ContinuousPdM, test data will be introduced continuously to validate the model’s output.

 

Conclusion

The ContinuousPdM powered by Power BI are more than just tools for predictive maintenance; they represent a significant step towards a smarter, data-driven approach to managing manufacturing operations. Once fully developed, these dashboards will offer tailored insights for engineers, operations, and facilities teams, enabling proactive decision-making, optimizing maintenance schedules, and enhancing overall operational efficiency.

This integrated architecture is crucial for manufacturing operations. It not only streamlines operations but also enhances predictive maintenance, reduces downtime, and improves overall efficiency. By harnessing the power of predictive analytics, manufacturers can optimize their production processes, ensure product quality, and maintain a competitive edge in the market.

As we continue to refine our predictive maintenance models, the integration of diverse data sources and advanced analytics will only improve, providing even greater accuracy and value. Embrace the future of maintenance with ContinuousPdM and transform your preventive maintenance program into an intelligent, data-driven strategy that maximizes productivity and minimizes downtime.

ContinuousPdM is your answer to go from Preventive to Predictive/Prescriptive/Cognitive.

 

xLM in the News

 
 

Latest AI News


What questions do you have about artificial intelligence in Life sciences? No question is too big or too small.