Predictive Repair Intelligence

Manufacturers worldwide traditionally rely on a concept called scheduled preventative maintenance to keep their products in good operating condition, whether they are vehicles on the road, commercial airliners in the sky, or wind mills in the remote outposts. A typical characteristic of such a mode of maintenance is that many perfectly good parts get tossed as a price for optimizing the uptime, since in all cases, an unexpected breakdown of the machinery is by far the most expensive mishap any company could face.

PRI EMPHASIS

A far more ideal method for fleet management is what’s known as predictive maintenance, whereby an intelligent algorithm takes in numerous measurements and predicts the pending component failure with a high degree of accuracy. From hereon, we refer to this algorithm as Predictive Repair Intelligence or PRI. In general, predictive models of this sort work either by classification for a breakdown within a prediction horizon, or by means of a Remaining Useful Life (RUL) assessment. While the application of supervised learning to such a problem can seem mundane, dealing with a real-life PRI case proves to be anything but straightforward. Here, in the rubber meets the road scenario, academic clarity counts much less than hard knuckled experience with real data, domain folklores, and the wild west of IoT sensors.

In a typical Data Science engagement, our team of experts walks into a hardware centric environment, where the client staff are of a deterministic logic mindset, with little prior exposure to statistical ways of problem solving. Design experts debug component failures by diving deeply into individual cases, juggling up to 4 or 5 failure instances maximum in his head. While information exchange is both interesting and informative, domain guidance alone rarely directly yields a workable set of relevant input signals and reliable labeled samples.

VIRTUOUS CYCLE

The key challenge starts with dimension reduction from over 200 signals down to a manageable set of 50 or fewer. Furthermore, the accuracy of failure labels is often untrustworthy, as labor cost for detailed diagnosis is far greater than making sacrificial part replacements. “Throwing parts at the problem” is considered the most practical way to eventually drive out the true cause of a warning light. Given these adverse initial conditions, BigRio has developed a general framework for attacking this class of problems, which we call the Virtuous cycle, whereby we iteratively apply visualization to gain intuition, and test out hypotheses using numerical / statistical techniques.

Visualization techniques are especially useful when dealing with a noisy feature set compounded with high data volume. Our study into battery aging, for example, leveraged an advanced clustering technique which illustrates how battery charging data can be used to distinguish between the fastest and slowest aging units.

PREDICTIVE MAINTENANCE

Predictive maintenance, and IoT data science in general, deals with high dimensional noisy data, where any one sensor measurement means little in driving an outcome, but many of them collectively can produce accurate classifications, given the proper application of data preprocessing and algorithmic fine tuning. In our parlance, we use the machine learning model to serve as a Virtual Sensor, in cases where the ideal hardware, dedicated sensor does not exist. A real-life example is that whereas a humidity sensor can often pinpoint system leakage, no such sensor can be manufactured economically to withstand the rough and tumble conditions of a driving car. A software algorithm which incorporates various sensor signals that are sensitive to humidity, on the other hand, can serve as a virtual substitute in gauging humidity effects. We believe many example applications of this powerful concept exist in the manufacturing world, and application of advanced Data Science in complex mechatronics hold the key to uncover these opportunities for improving the quality of fleet health.

High equipment reliability is critical to the productivity and profitability of manufacturing companies. Traditional run-to-failure maintenance management is known to be vastly suboptimal because unscheduled downtime translates into high labor and maintenance costs and opportunity cost due to the idle time.

Four modes of maintenance:

Corrective – don’t fix until it breaks.
Preventive – schedule maintenance based on age only.
Predictive – schedule maintenance based on age and condition.
Cost matrix – optimize overall maintenance cost.

While preventative maintenance circumvents disruptions to operations, it incurs unnecessary cost by discarding many parts still in good working condition. An ideal maintenance policy takes into account the current condition of the mechanical components and schedules repair or replacement based on forecasted REMAINING USEFUL LIFE (RUL). This is predictive maintenance.

REMAINING USEFUL LIFE “RUL” PREDICTION

It is possible to get a read on the remaining useful life of a machine by monitoring the right predictor signal. For example, excessive vibration is a good indication that bearings are wearing out and component failure is an eventuality. While vibration signals are noisy in nature and defy simple cutoff rules, it is possible to characterize the gradual degradation process by means of artificial neural network training, as shown below.

In this illustration, the simple feed-forward neural network is trained using a sequence of vibration signals and its time indices and is able to reliably calculate the remaining useful life for test cases with widely different phase II durations.

Application of NEURAL NETWORKS to RUL prediction

In real life, IoT data can be much more complex and multifaceted than the vibration signals in the last example. The data can arrive with or without schema in the textual, audio, video, imagery and binary form, sometimes multi-lingual and encrypted. They are often streamed in real time and may require a substantial infrastructure upgrade in terms of storage, integration, and analytic requirements. Once captured, relating the relevant predictor signals to the failure event is generally a challenging process befuddled by high dimensionality, obscure causality, and complex distributions.

Neural Networks have properties that make them particularly useful for solving problems with this level of difficulty. Deep Neural Networks are known for their strengths in the following areas:

Implementing regression or classification by applying nonlinear transformation to linear combinations of raw input features.
Hidden layers and high neuron counts increase the expressiveness of the Neural Network.
Self-learning through back propagation mechanism.
Auto detects composite events that serve as precursors to the phenomenon of interest.
Excellent in multi-modal learning, matching the multi-modality of IoT.
Well-suited for temporal pattern recognition

CEP AND STREAM PROCESSING

A major part of the predictive maintenance challenge is in sorting out the relevant predictors in a sea of noisy and interdependent sensor signals. Before getting into the training phase, a complex event processing (CEP) effort must first be conducted to both sort out the candidate signals and identify the applicable regime of signal lifecycle. Often, subject matter experts provide important clues as a starting point. In such cases, a direct investigation of the event sequence, time window, and predicate relationships can be derived by an iterative computation.

In other cases, the complexity justifies the employment of various neural network techniques. Recurrent Neural Networks (RNN), for example, are an excellent tool for the study of time-series events that occur in sensor data. Automatic mapping of raw signals into a high-dimensional space facilitates the classification of all the sensor inputs into clusters of similar behaviors. Advanced visualization allows investigators to efficiently identify useful patterns that can drive out the likely suspect signals.

Once signals are identified and deep learning models are trained, the entire predictive maintenance module can be deployed for real-time monitoring of warning signs and give operational staff clear metrics for scheduling optimal maintenance procedures. Typically, the runtime environment consists of a stream processing front end capable of picking out the early warning signs and the neural network executable, which makes prediction on these signals in their operational regime.