An AI Approach to Managing Motor–Driven Pump Reliability

Overview

Asset-health data from connected equipment is valuable—if you know how to use it. If you don’t, it’s an expensive collection of data sets. For example, let’s say you’ve collected substantial Internet of Things (IoT) measurement data, but you can’t fully extract value from it without subject matter experts (SMEs), which are disappearing from the reliability industry. The solution? Automate the process of converting data to insight. In this paper, we’ll discuss the InsightCMTM Analytics Toolkit, a fresh approach to condition monitoring and predictive-maintenance analytics.

Traditional Analytics Challenges

Predictive maintenance (PdM) pilot projects often generate difficult questions:

What data do we need?

Pure data-science analytics are beneficial because they let you “work with the data you have.” But aside from convenience, does the data available provide the insight the analytics need to predict failure? Convenient data often comes from a process control or asset management system. And while process variables can provide asset health insight, those systems really were designed to operate a production line. For example, motor vibration and current signatures are common reliability measurements, but typically aren’t valued in a production control system. Data coverage gaps lead to failure-mode blind spots that are difficult to find in machine-learning environments.

What model-training expertise do we need?

Machine-learning models look for patterns and correlation amongst a set of unlabeled, unitless data. This translates to “working with the data you have,” but also, it means that the model needs to experience different states to detect patterns. A SME helps train the model, effectively telling it, “This is normal operation for these inputs.” Think of everything a machine goes through that can cause the model to formulate a new state: Environmental changes, operating speeds, materials handled—all normal conditions need evaluation before any measurement can be identified as anomalous. This requires SME time and knowledge.

Do we need failure data?

If you want the model to know that the machine is in a failure state, the model needs to learn what failure looks like.

According to Chris Coleman and Ed Deuel, Specialist Leaders at Deloitte Consulting, “The reason you generally can’t go straight to predictive maintenance is that it takes time to set up data collection processes, connect sensors to the machines, and most importantly, the asset has to fail at least a few times in order to baseline the predictive algorithms. The more a machine fails over time, the better the predictions should logically become.”1

Asset-Specific Analytics Can Help

Asset-specific analytics combine engineering-based asset knowledge (digital twin) with machine learning to diagnose problems, predict failure, and, ultimately, assess operation risk.

The InsightCM Analytics Toolkit

The InsightCM Analytics Toolkit integrates asset-specific analytics from The DEI Group with sensor data from InsightCM installations. Toolkit components, available from NI, include all InsightCM and The DEI Group software elements to go from data to dashboard.

InsightCM Analytics Toolkit Features PreMA for Pumps – Base PreMA for Pumps – Full

PreMA Software

 

Health Dashboard

Anomaly Detection

Diagnostics

  • What is my overall asset health score?
  • Which asset should be closely watched for failure?

Prognostic – What is the remaining useful life (RUL) of each component?

Maintenance Optimizer – When should I service which components?

Reliability Dashboard

 

 

About The DEI Group

The DEI Group has helped customers such as the US Navy, Chevron Shipping, GE Power, Chevron Power, TVA, Ontario Power Generation, Duke Energy, and others deploy solutions that change the status quo of reliability programs for more than 37 years. Their development resources combine PhD-level academic knowledge and industry experience in fields including data science, plant engineering, equipment maintenance, and a variety of engineering disciplines.

Combining Engineering Knowledge with Machine Learning

Unlike a pure data-science approach, asset-specific analytics understand wearable components and dominant failure modes and account for installation, manufacturing, and operational variances. Engineering knowledge is unique to pump type: The model for a vertical, multistage, axially split pump is different than the model for a vertical, single-casing pump with discharge through an axial-flow column.

Automatically Diagnose Dominant Failure Modes

With voltage and current sensors, NI monitoring hardware, NI InsightCM, and associated toolkits, you can detect the following failure modes (as applicable) on a motor-driven pump:

Pump

  • Casing worn
  • Casing wear ring degraded
  • Flange gasket leaking
  • Impeller worn
  • Impeller key loose sheared
  • Impeller loose on shaft
  • Pump shaft bent
  • Pump shaft unbalanced
  • Thrust collar wear
  • Thrust collar key loose
  • Active thrust shoes worn
  • Inactive thrust shoes worn
  • Pump non-drive end (NDE) sleeve bearing worn
  • Pump NDE bearing labyrinth seal leaking
  • Pump drive-end (DE) sleeve bearing worn
  • Pump DE bearing labyrinth seal leaking
  • Inboard seal faces worn
  • Inboard single coil spring worn
  • Inboard seal O-rings worn
  • Outboard single coil spring worn
  • Outboard seal O-rings worn

 

Foundation

  • Pump motor misalignment
  • Soft foot (loose foundation bolts)
  • Flexible coupling sleeve wear
  • Coupling end rings leaking

Motor

  • Stator windings shorted
  • Stator winding insulation worn
  • Stator windings connections loose
  • Stator core lamination loose
  • Stator wedges loose connections
  • Rotor bars cracked degraded
  • Rotor core lamination damaged
  • Rotor core air gap static eccentricity
  • Rotor core air gap dynamic eccentricity
  • Motor shaft bent
  • Motor shaft flinger leaking
  • Motor DE bearing worn
  • Motor DE bearing oil ring leaking
  • Motor NDE bearing worn
  • Motor NDE bearing oil ring leaking
  • Gating element degraded
  • Cooling fan dirty

Match Sensor Measurements to Analytics

The InsightCM Analytics Toolkit takes the guesswork out of selecting sensors for PdM. For motor-driven pumps, the InsightCM Analytics Toolkit requires three-phase voltage and current measurements to capture the dominant electrical and physical failure modes.

Taking Three-Phase Voltage and Current Measurements

High-speed voltage and current measurements include valuable data sets such as in-rush current, unbalance, and sideband measurements that are used to detect failure modes such as broken rotor bars and worn insulation.

Motor-current signature analysis (MCSA) installations use sensors that attenuate high voltage/current into levels suitable for direct instrumentation input:

  • Current sensors loop around electrical connections to the motor (or the protective relay secondary inputs)
  • Voltage measurements are made directly to NI instrumentation or via measurement potential transformers that connect to the same input modules as low-voltage current transformers

 

 

Figure 1. Split-core current sensors typically install in the motor control cabinet. Safety first: Always de-energize the asset/cabinet for install.

The motor-control center (MCC) makes for an easier installation point because you can monitor many motors from a single location—unlike vibration, which requires installation at each physical asset location.

SAFETY NOTE: Always de-energize equipment before working in the motor control cabinet, even when you don’t need to cut circuits for installation.

Figure 2. A single NI Continuous Monitoring System can monitor multiple assets because of the shared voltage bus and available expansion slots for current sensor input modules.

Note: The MCSA Toolkit connects InsightCM to NI Continuous Monitoring Systems for I/V measurements and is required for the InsightCM Analytics Toolkit.

Detecting Mechanical Problems with Electrical Signals

Traditionally, SMEs use voltage/current measurements to detect electrical failure modes (broken rotor bars or worn insulation) and vibration signatures to detect physical failure modes (roller-element bearing failures or shaft misalignment). But with the InsightCM Analytics Toolkit, you can extract features for both electrical and mechanical motor-driven pump components from high-speed, time-based voltage/current waveform data. Physical defects impact electrical signals through the air gap eccentricity between the rotor and stator. And we can see additional disturbances, such as indications of a worn impeller, through torque and power-spectra ripples. Configuring vibration and process data into the model provides additional failure evidence to the diagnostic engine, reinforces mechanical degradation findings, and helps you diagnose with more context.

Configure Pump Analytics Models

Asset-specific analytic models don’t need as much training, but they do need to understand the monitored asset’s physical properties. The asset-specific component is a “built-in SME” that supervises the training and matches sensor groupings to correlating failure modes.

Figure 3. The InsightCM Analytics Toolkit uses pump-specific analytics. You can configure the type of pump in InsightCM to inform the model. (Image courtesy of The DEI Group)

You can configure individual assets using the following information:

  • Pump and motor survey from nameplate
  • Sensor data
  • Maintenance history

 

After configuration, the model enters learning mode, using time-based waveform data to establish the parametric relationships that represent failure modes of interest. The learning process observes the equipment through its expected operating range to build a high-confidence baseline model. Learning time duration depends on type of equipment and operating range to be observed (constant load/variable load).

Once learning mode is complete, the Health & Reliability Dashboard displays failure mode information, including:

  • Health, criticality, and reliability ratings by failure mode
  • RUL
  • A reliability outlook for the next 30/60/90/120 days for the whole asset train and each component (motor/pump)

Interact with Your Data

Different stakeholders need varying levels of data and information to do their jobs. With the InsightCM Analytics Toolkit, you get three levels of information to address different workflows.

High-Level Dashboards and Maintenance Schedule

Use the InsightCM Reliability Dashboard to interactively explore current and future operations risks. You also can assess individual component (impellers or bearings) risk at this level. The Schedule Optimizer combines asset-health data with a user-input availability schedule (production schedule and planned outages) and cost/risk constraints to generate a by-asset component repair schedule. Use the schedule to repair assets only when needed (lowering maintenance costs) and maximize time by replacing other worn components while the asset is offline (boosting your return on investment).

Figure 4. View overall asset health scores in real time and 30/60/90/120-day risk assessments on the Reliability Dashboard included with the InsightCM Analytics Toolkit.

Failure Mode Information and Model Inputs

Knowing which pump is adding risk to your operations is helpful. Knowing the specific part that is going to fail, and when, drives key performance indicators in the right direction. You can use the PreMA software package (included with the InsightCM Analytics Toolkit) to better understand the measurements driving the results, or to zoom in on a failure mode.

Figure 5. Launch The DEI Group PreMA application software (included with the InsightCM Analytics Toolkit) directly from InsightCM to inspect specific failure modes, associated measurement data, and suggested maintenance actions.

Raw Sensor Data

Use InsightCM to access raw data with industry-standard analysis tools to spot-check components, troubleshoot sensor issues, and confirm a diagnosis from the analytics engine.

Figure 6. InsightCM includes industry-standard calculations, viewers, and analysis tools to diagnose equipment fleets from plants around the world.

Deploy as Part of Your Digital Transformation

Analytics are a complex yet powerful technology for improving plant reliability. The DEI Group’s analytics and NI InsightCM are adjacent technologies in the industrial IoT software stack and combine to help companies modernize asset maintenance with a complete solution–from sensor data to informational dashboards.

Figure 7. InsightCM has the hardware, sensor, software, and analytics support you need to connect a variety of assets to your IoT network as part of a digital transformation program.