PdM Tech Deep Dive

Predictive Maintenance strategies are underpinned by the application of cutting edge technology that is used to derive insight from data.In this blog post you will discover the broad range of technologies available to predictive maintenance to form an effective strategy as well as future technological trends.

A predictive maintenance technology is one which is used to generate condition indicators from data which can then form part, or all, of a system health forecasting model.Many of the same technologies which are used to form condition indicators can be used to aggregate these indicators into a remaining useful life (RUL) model.

What Technologies are Used in Predictive Maintenance ?

Technologies from a broad range of mathematical domains are used in constructing predictive maintenance solutions.The core essence of PdM is the application of analytical techniques to abstract away information from the raw data that can be aggregated in some manner over time to create a predictive model.In this way, PdM can be considered a subset of the Predictive Analytics branch of statistics.

Below, we outline a set of technology groups and their utility in forming the crucial parts of a predictive maintenance solution.

Data Driven Methods for Predictive Maintenance

Data driven methods comprise all methods where conclusions are drawn from the data alone, through the testing of hypotheses. Through this verification process, models can be built that either classify, group or regress onto an estimation space, which can then be aggregated into a lifetime model.

Here are a range of data driven techniques commonly used within predictive maintenance:

• Machine Learning: An umbrella term encompassing methods that leverage the computational power of machines to ‘learn’ an inference target from the data in a semi-instructed fashion.Popular methods include; Random Forests, XGBoost Models, Ridge and Lasso Models, Bayesian Models, Support Vector Machines and K - Nearest Neighbours Clustering
• Deep Learning: An extension onto Machine Learning where neural networks are used to gain progressively high level features from the data through many sequential layers of abstraction.It is a technique that is common in the feature engineering stage of Predictive Maintenance.Some Deep Learning technologies commonly used: Convolutional Neural Networks, Standard Neural Networks, Recurrent Neural Networks, and Auto - Encoder architectures
• Time Series Models: In addition to Machine Learning methods there are a suite of time series methods that are similar in their design to a lot of machine learning methods.These are often very algorithmic techniques where a specific cost is minimised.Commonly, the objective with time series methods is to reconstruct a time series given past data and / or discover patterns in the data.Similarity methods are also included in time series methods, such as Dynamic Time Warping.Common techniques for time series reconstruction / prediction are Auto - Regression, ARMA and ARIMA models
• Anomaly Detection Models: Anomaly detection is a key method when detecting Black Swan Events in condition indicator data. Such events often precede the beginning of a major degradation and so can be used to form part of a predictive maintenance strategy. There are a class of models that focus on detecting these events given a suite of historical data that does not necessarily need to be classified as anomalous or not, but rather the anomaly detection approach can isolate points that are statistically likely to be anomalous. Some of these methods include Isolation Forests, the Local Outlier Factors algorithm, and Bayesian Models
• Clustering Models: The formation of clustering models from data can form a key part in the construction of condition indicator design when building up a predictive maintenance strategy.For example, the frequency of occurrence of a condition indicator into a specific cluster region could increase as failure starts to occur.To determine clusters from data, both supervised and unsupervised methods can be used.Two of the most common methods for clustering are K - Means and the DBSCAN algorithms

State Estimation for Predictive Maintenance

State estimation forms a core part of traditional control theory, where a state of a system is to be estimated without being, necessarily, directly measured.Such methods are also applicable in the domain of PdM if the system state to be estimated is indicative of system health, or, the state to be estimated is simply the Remaining Useful Life.

There are a variety of methods available to estimate system state.Here are a few:

• Kalman filter: A classical state estimation method that fuses system measurement with a model prediction. The Kalman solution provides the optimal fusing of prediction and estimation when presented with Gaussian process and measurement noise
• Complementary filter: A filtering method for fusing the measurements from multiple sensors and / or prediction models.Usually designed in the frequency domain, transfer functions from input to estimation error are fused in an optimal way
• Bayesian Update Filter: A class of filters that recursively propagate a Bayesian solution to an estimation problem. Bayesian update filters commonly have a similar algorithmic structure to a Kalman filter but implement mathematical solutions to the Bayesian estimation problem at each iteration of the filter
• Particle Filter: A class of Bayesian update filters that utilise Monte - Carlo sampling methods for propagating forward a prediction model.Particle filters can be extremely powerful tools in multimodal problems where some knowledge of the estimation map is known
• SLAM: SimuLtaneous Acquisition and Mapping is an extension onto a particle filter, where the map is simultaneously built as it is estimated

Time and Frequency Domain Methods for Predictive Maintenance

There is an extremely large set of time domain techniques for the analysis of data in Predictive Maintenance problems; here are a few:

• Anomaly Detection Models: Anomaly detection models are a common time series method where the goal is to detect abnormal signals in real time.Often informed through domain knowledge, peak detection algorithms form an important subset of anomaly detection models.For example, if a specific failure is known to have a particular peak characteristic, then the peak detection algorithm can be designed robustly prior to observing any real - world data
• Windup Detection: Windup detection is an advanced method for the detection for specific types of control failure.The failure mode that is being detected is a breakdown in control action due to evolving plant dynamics.This can present as more energy being put into the system without the expected output.The result is a windup in system state often causing further degradation
• Similarity Detection: If failure modes follow a specific time series history, similarity detection can be used to identify whether the observed signal history is indicative of a failure.If it is, the system can be flagged before the failure propagates too far
• Discrete Filtering: In addition to the aforementioned estimation methods, signals can be passed through a discrete filter to isolate frequency bands.If failures are known to exist in certain frequency bands the magnitude of a filter response can be tracked to detect failures in this frequency region
• Heterodyne Locking: Similar to filtering, Heterodyne Locking can be used to capture frequency information in real time, by comparing the signal to a reference frequency.The reference frequency may be the frequency where a known failure mode exists and so it can be tracked to detect failure.Heterodyne Locking is particularly important when the phase characteristics(with respect to a reference signal) is important in detecting degradation

A few techniques of frequency domain analysis are listed below:

• Frequency Spectrum Estimation: The core of all frequency based techniques – the Fourier Transform. This method seeks to identify the frequency information present within a signal, often achieved through the application of the Fast Fourier Transform algorithm. _ Power Spectrum Estimation: Use this technique to understand the Power content of your signal over the domain of observable frequencies.Welches method is often used if the signal is known to have several core fundamental frequencyes present, and the signal is stationary _ Spectrogram Calculation: This method expands power spectrum estimation to apply over the time domain.This technique is useful if you want to observe the change in frequency component as a function of time, i.e. if a signal is not stationary

Systems Engineering and Risk Modelling for PdM

Systems engineering principles cover a broad range of development techniques – all focused around the proper delivery of a solution.As such, they have penetrated all areas of engineering and Predictive Maintenance is no exception.There are many methods comprising systems engineering techniques, but here are a few that are particularly useful when designing a predictive maintenance strategy:

• Requirements Capture: Requirements capture is the principle of identifying the core requirements to a solution for your problem.There are many tools available to aid in the capturing of requirements as well as different methodologies in requirements capture.For example, more traditional methods would summarise system requirements as a hierarchy of statements citing the required performance of the solution.Alternative more agile methods express requirements as user stories.Both methods are valid, and must be implemented correctly to realise value for the development of a predictive maintenance strategy

During the verification stage of any project several techniques are used to not only confirm the proper function of a solution, but also understand the behaviour of the solution under different scenarios.In essence this is where exposure to risk caused by the implementation of a solution is evaluated.There are many methods and tools available to perform this analysis, here are a few:

• FMEA: Fault Mode Event Analysis is the construction of the fault modes of a system and the subsequent events that may follow if any one of those faults occur.Often used as a process to discover fault patterns, it can be used to help focus efforts on building conditional indicators for the largest impact failure modes
• Fault Tree Analysis: Related to FMEA, FTA focuses on the construction of isolated fault events and how they relate to each other.Known(or proposed) failure rates of each event can be attached to each node in the fault tree along with the failure relationship between each node.Constructing such a picture of the system can help understand which components  can be classified as ‘high risk’ and therefore where a predictive maintenance strategy should be focused
• Risk Priority Numbering: This method is the application of a numerical value to the risk associated with a step in a process or a part within a system.Often a crucial step in FMEA, RPN can be used to better understand the maintenance liability of an existing process
• Uncertainty Quantification: A crucial step in communicating a Predictive Maintenance strategy is in quantifying the uncertainties associated with an estimate.There are many methods to quantify the different uncertainties that exist in the predictive maintenance pipeline, but some methods include: confidence interval calculation,
  sensitivity analysis and model averaging
• Risk Management Models: A risk management model is any process that attempts to identify and isolate the risk associated with doing or not doing.Predictive Maintenance can form a crucial element of mitigating risk associated with unforeseen service cost, but there are other techniques for containing risk pertaining to other parts of the pipeline.For example, Kelly - optimal resource allocation specifies the amount of available resource that should be allocated given the likelihood of an adverse event resulting from the allocation of the resource
• Monte Carlo Simulations: Monte Carlo simulations summarise tools to perform large scale simulations where input parameters to a specific model are varied given specified distributions.In the context of risk management, the model under test would be the predictive maintenance strategy embedded into a model of your distribution framework.Parameters to vary could be stock availability, and prediction efficacy; the result being a range of cost variation that you can factor into your forecasting tools

Future Trends in Predictive Maintenance Technologies

Many of the methods mentioned in this post are very old and are rooted in core areas of mathematics. Looking to the future, what are some of the trends in PdM technologies ?

In this blog post, we have taken a glance at time - honored methods that originate from fundamental domains of mathematics and engineering.As we shift our gaze towards the horizon, what emerging trends can we anticipate in the realm of PdM technologies ?

IoT Enablement

Industry is becoming awash with data.Many of our systems are connected to the cloud or on premise data stores, or we are actively looking how and what we need to connect to the cloud.Having such a large data capability across industry has accelerated interest in IoT enablement whilst singularly increasing the sheer amount of data that is being collected and stored.

This increase in data awareness is only going to advance the utilisation of data driven methodologies in industry, helping businesses realise further value data.In addition, the cross pollination of data from company to company will increase as organisations form strategic partnerships and attempt to expedite their value realisation goals.

Data Sparsity

With the increased data advocacy that we see in industry, manipulation and storage become real concerns as the cost for processing and storing data increases with scale.Methods for storing such as embedding the data into latent spaces will become more popular in addition to more classical applications such as compressed sensing.

However, although sparsity of data is something that may be advocated for in the increasingly connected industry, there are unique challenges that come with it. For example:

• Information loss
• Aliasing

We will see increased use of data driven methods to compensate for increased sparsity, as overall apetite for information will undoubtedly increase.

In conclusion, the exciting world of Predictive Maintenance(PdM) operates on the cutting edge of technology, extracting valuable insights from data.We’ve embarked on a journey exploring the array of technologies that fuel predictive maintenance strategies, understanding their role in establishing effective system health forecasting models.

The Evolving Domain of PdM

Over the course of this blog post, the critical technologies underpinning Predictive Maintenance(PdM) across various industries have been illuminated.The exploration has been broad, spanning from the tools to generate condition indicators to systems of validation and verification.

Yet, it’s to be noted that this is but a snapshot of the present.The field of predictive maintenance is an ever - evolving landscape, consistently marked by the emergence of pioneering technologies poised to disrupt established norms.As ongoing transformations are observed, it’s essential to stay informed about the latest advancements.The commitment at Predixus is to ensure that we serve as a reliable source for insight.