Data Analytics and Decision Support

Introduction

Modern mechatronic systems generate data continuously. Sensors measure temperature, vibration, position, pressure, current, speed, and force; controllers record set points and responses; manufacturing systems log cycle times, defects, alarms, and maintenance events. Data alone, however, does not improve an engineering system. Improvement occurs when data are transformed into trustworthy information and then used to make defensible decisions.

Data analytics is the systematic process of inspecting, cleaning, transforming, summarizing, visualizing, and interpreting data. Decision support applies the resulting evidence to choices such as whether a machine requires maintenance, whether a process is stable, whether a controller needs adjustment, or whether a production line should be redesigned. This chapter emphasizes descriptive analytics and practical decision support because these are foundational skills for engineers working with spreadsheets, databases, Python, dashboards, and industrial information systems.

The central principle is that an analytical conclusion is only as reliable as the entire workflow that produced it. A visually attractive dashboard can still mislead if its data are incomplete, its measures are poorly defined, or its uncertainty is hidden.

Learning Objectives

After studying this chapter, you should be able to:

1. Distinguish data, information, evidence, and decisions.
2. Define an analytical question and identify suitable data sources and variables.
3. Detect and address missing values, duplicates, inconsistent units, outliers, and invalid records.
4. Calculate and interpret descriptive statistics for engineering data.
5. Select suitable charts and design informative dashboards.
6. Construct a reproducible analytical workflow.
7. Recognize uncertainty, sampling limitations, measurement error, and bias.
8. Interpret analytical results in relation to engineering constraints and risk.
9. Use a structured framework to support evidence-based decisions.

1. From Engineering Data to Decisions

A useful analytics chain is:

measurement → storage → cleaning → summarization → visualization → interpretation → decision → action → feedback

A measurement is an observed value produced by an instrument, sensor, test, or human observation. Data are organized measurements or records. Information is data placed in a meaningful context. Evidence is information that is sufficiently relevant and credible to support a claim. A decision is a choice among alternatives, usually involving objectives, constraints, risks, and consequences.

For example, a vibration sensor may record 4,000 acceleration samples per second. These samples are data. The calculated root-mean-square vibration over a ten-minute operating period is information. Comparing that value with a validated alarm limit and historical baseline produces evidence. Scheduling an inspection is the decision.

Analytics is often classified into four questions:

Descriptive analytics is the basis for the other categories. If the data definition, quality, or measurement process is defective, a sophisticated predictive model does not solve the underlying problem.

2. Framing an Analytical Problem

An analytical project should begin with a decision, not with a chart. A vague request such as “analyze the motor data” should be converted into a precise question:

• What decision is required?
• Who will make it?
• What is the response variable?
• Which explanatory variables may be relevant?
• What time period and operating conditions are included?
• What action follows each possible result?
• What accuracy, cost, latency, and safety requirements apply?

A variable is a measurable characteristic. A quantitative variable has numerical values, such as speed in revolutions per minute. A categorical variable identifies groups, such as product type or alarm class. A discrete variable is countable, such as the number of faults; a continuous variable can take values within an interval, such as temperature.

A data dictionary should be created before analysis. It records variable names, meanings, units, valid ranges, data types, collection frequency, source, and permitted missing-value codes. For example, “temperature” is not a complete definition unless the analyst knows whether it means bearing temperature, ambient temperature, or controller temperature and whether the unit is °C or °F.

Measurement scales are also important. Nominal categories have no order, such as sensor location. Ordinal categories have an order but unequal or unknown spacing, such as low, medium, and high severity. Interval measurements have meaningful differences but no meaningful zero, such as temperature in °C. Ratio measurements have a meaningful zero, such as mass, time, or absolute pressure. The scale affects which summaries and calculations are valid.

3. Data Acquisition and Quality

Engineering data may come from programmable logic controllers, supervisory control and data acquisition systems, laboratory instruments, maintenance-management systems, spreadsheets, databases, and application programming interfaces. Each source has possible failure modes.

A sensor may drift, saturate, be incorrectly calibrated, or lose communication. A database may contain duplicate records after a system restart. A manually entered maintenance record may use inconsistent spelling. A timestamp may be recorded in local time in one system and Coordinated Universal Time in another. Sampling intervals may vary, making direct comparisons invalid.

A practical quality assessment considers the following dimensions:

Quality rules should be explicit. For example, a record may be rejected if a rotational speed is negative when reverse rotation is impossible, if a timestamp is absent, or if a temperature lies outside the sensor’s specified range. A value outside an expected operating range is not automatically an error: it may represent a genuine fault. It should therefore be flagged for investigation rather than silently deleted.

Cleaning Operations

A defensible cleaning process generally includes:

1. Preserve the original raw data as read-only evidence.
2. Record the source, extraction time, and software version.
3. Standardize variable names, units, dates, and categorical labels.
4. Identify duplicate records and determine whether they are genuine repeats.
5. Quantify missing values by variable, time period, and operating condition.
6. Apply documented validity rules.
7. Investigate outliers using engineering knowledge and source records.
8. Save a cleaned dataset and a change log.

Missing values require judgment. If a sensor temporarily fails, interpolation may be inappropriate, especially during a rapid transient. For a slowly changing process, interpolation may be acceptable for visualization but not for safety-critical calculations. Common strategies include deletion, imputation with a statistical value, interpolation, model-based estimation, or explicit treatment as a separate state. Every strategy adds assumptions and should be documented.

An outlier is an observation that differs markedly from the rest of a dataset. It may be a data-entry error, a sensor fault, a rare but valid event, or evidence of a process change. An outlier should not be removed merely because it makes a result inconvenient.

4. Descriptive Statistics

Let a dataset contain n observations x₁, x₂, …, xₙ.

Measures of Location

The arithmetic mean is:

Here, x̄ is the sample mean, xᵢ is observation i, and n is the number of observations. The mean uses every value and is useful for approximately symmetric data without extreme outliers. It has the same units as x.

The median is the middle ordered observation. For an odd number of observations it is the central value; for an even number it is commonly the average of the two central values. The median is less sensitive to extreme observations than the mean.

The mode is the most frequently occurring value or category. It is particularly useful for categorical data.

The weighted mean is:

where wᵢ is the weight assigned to observation i. Weights must be non-negative and should represent a justified quantity such as exposure time or production quantity. A weighted mean is not automatically more accurate than an ordinary mean.

Measures of Spread

The range is maximum minus minimum. It is easy to calculate but depends strongly on extreme values.

The sample variance is:

The sample standard deviation is:

The denominator n − 1 is used when the sample standard deviation estimates the variability of a larger population. If the complete finite population is being described rather than estimated, the population variance uses n in the denominator. Standard deviation has the same units as the measured variable, whereas variance has squared units.

The interquartile range is:

where Q₁ is the first quartile and Q₃ is the third quartile. The IQR describes the middle 50% of observations and is relatively robust to extreme values.

The coefficient of variation is:

It is dimensionless and is often reported as a percentage, CV × 100%. It is meaningful mainly when the mean is positive and sufficiently far from zero. It is useful for comparing relative variability among measurements with different scales, but not when the variable can be negative or the mean is near zero.

Worked Example: Cycle Time

Suppose six cycle times, in seconds, are 10, 11, 10, 12, 9, and 14. The mean is:

The ordered data are 9, 10, 10, 11, 12, 14, so the median is (10 + 11)/2 = 10.5 s. The range is 14 − 9 = 5 s. The deviations from the mean are −1, 0, −1, 1, −2, and 3 s; their squared sum is 16 s². Therefore:

and:

The mean exceeds the median because the 14-second observation pulls the mean upward. An engineer should investigate whether that observation represents a fault, a deliberate interruption, or normal variability.

Percentiles and Rates

The pth percentile is a value below which approximately p percent of observations fall. Percentile calculations can differ slightly among software packages because interpolation conventions vary; the method should be recorded when results are used formally.

Rates must include a denominator. “Five failures” has little meaning without the number of operating hours, units produced, or opportunities. A failure rate may be expressed as failures per 1,000 operating hours or defects per million opportunities. Comparisons are valid only when the denominators and definitions are comparable.

5. Relationships Between Variables

Covariance indicates whether two variables tend to vary together, but its units depend on both variables. The Pearson correlation coefficient is:

where x̄ and ȳ are sample means. The coefficient lies between −1 and +1 when both variables have nonzero variation. Values near +1 indicate a strong positive linear association, values near −1 a strong negative linear association, and values near 0 little linear association.

Correlation does not establish causation. A third variable may influence both measured variables, the relationship may be nonlinear, or the observations may be time-dependent. Correlation is also sensitive to outliers and restricted ranges. Scatter plots should accompany correlation values.

A simple linear model is:

where y is the response, x is a predictor, β₀ is the intercept, β₁ is the expected change in y for a one-unit increase in x, and ε represents unexplained variation. The model assumes, for basic interpretation, an approximately linear relationship and appropriately behaved errors. A fitted line can summarize association; it does not by itself prove a physical mechanism.

6. Visualization and Dashboards

A chart should answer a question. Line charts are suitable for trends over time, scatter plots for relationships, histograms for distributions, box plots for comparing distributions, bar charts for category comparisons, and control charts for monitoring process behavior over time. Pie charts are usually inferior when precise comparisons among several categories are required.

Good visualizations:

• label axes with units;
• state the time period and population represented;
• use an honest scale and avoid unnecessary three-dimensional effects;
• distinguish missing values from zero;
• identify thresholds and reference values;
• show sample size where relevant;
• use accessible colors and alternative visual cues;
• annotate unusual events such as a tool change or sensor replacement.

A dashboard is an interface for monitoring and decision support, not merely a collection of charts. Its design should follow the user’s decisions. A production supervisor may need current throughput, downtime, defect rate, and active alarms. A maintenance engineer may need vibration trends, operating hours, temperature, and inspection history. A senior manager may require aggregated availability, quality, cost, and risk indicators.

A useful dashboard separates:

1. Context: asset, time period, operating mode, and data freshness.
2. Key performance indicators: concise measures linked to objectives.
3. Diagnostics: trends, distributions, comparisons, and contributing factors.
4. Action information: thresholds, ownership, escalation rules, and next steps.

A KPI should have a precise definition, numerator, denominator, unit, target, update frequency, owner, and data source. For example, overall equipment effectiveness is commonly represented as:

Each component is expressed as a proportion between 0 and 1 under the chosen definitions. Availability may be operating time divided by planned production time; performance may compare actual output with ideal output during operating time; quality may be good units divided by total units. Because definitions vary across organizations, an OEE value is meaningful only when its components and exclusions are documented.

7. Reproducible Analytical Workflows

Reproducibility means that another competent person can obtain the same result from the same source data by following the recorded procedure. Reproducibility is not limited to programming; a carefully documented spreadsheet can be reproducible, while an undocumented script may not be.

A robust workflow commonly contains:

1. Raw layer: immutable source files or database extracts.
2. Preparation layer: scripts or documented transformations for cleaning and joining.
3. Analysis layer: calculations, statistical summaries, and tests.
4. Presentation layer: tables, charts, dashboard outputs, and conclusions.
5. Documentation layer: data dictionary, assumptions, validation checks, and change history.

Good practice includes version control, meaningful file names, parameterized date ranges, automated quality checks, environment or package records, and separation of data from code. Random samples or simulations should use a recorded random seed when exact reproduction is required.

A workflow should also be validated. Check totals against source-system totals, inspect a sample of transformed records, test boundary conditions, verify units, and confirm that filters do not unintentionally exclude relevant operating modes. A second person should review critical calculations, especially those affecting safety, compliance, or production release.

8. Uncertainty, Bias, and Responsible Interpretation

Uncertainty is not the same as error. Error is the difference between a measured value and the true value, while uncertainty describes limited knowledge about that value or conclusion. Sources include sensor resolution, calibration uncertainty, sampling variation, missing observations, model assumptions, and changing operating conditions.

A measurement should not be reported with unjustified precision. If a temperature sensor has resolution of 0.1 °C and substantial calibration uncertainty, reporting 72.38461 °C implies knowledge that the instrument does not possess.

Sampling bias occurs when the observed data do not represent the population relevant to the decision. Examples include collecting vibration data only during daytime operation, analyzing successful production lots while excluding rejected lots, or evaluating a machine only after maintenance. Survivorship bias occurs when failed or discontinued cases are absent. Selection bias, measurement bias, and confirmation bias can all distort conclusions.

When comparing groups, consider confounding variables. A machine may show longer cycle times on the night shift, but product mix, staffing, temperature, or scheduled cleaning may differ between shifts. Stratification, controlled experiments, careful observational design, and transparent sensitivity analysis can reduce misleading interpretations.

Ethical and governance responsibilities include:

• protecting personal and operationally sensitive data;
• limiting access according to legitimate roles;
• retaining records according to policy and regulation;
• documenting automated decisions and allowing review;
• avoiding discriminatory or unsafe use of proxy variables;
• preserving audit trails;
• communicating uncertainty rather than overstating confidence.

In safety-critical systems, an analytics result should support—not silently replace—engineering judgment, validated procedures, and required approvals. Artificial intelligence can detect patterns and generate summaries, but it may reproduce biased training data, hallucinate explanations, or fail under changed operating conditions. AI-generated code and conclusions require testing against known cases, source verification, and human accountability.

9. Decision-Support Framework

A practical decision-support process has six stages:

1. Define the decision and consequences. Identify alternatives, costs, risks, and constraints.
2. Specify evidence requirements. Determine the variables, accuracy, time window, and acceptable uncertainty.
3. Assess data quality. Record exclusions, missingness, measurement limitations, and possible bias.
4. Analyze and compare alternatives. Use descriptive statistics, visualizations, engineering models, and domain knowledge.
5. Apply thresholds and escalation rules. Distinguish warning conditions from action conditions.
6. Review outcomes. Compare the decision with what happened and update the process.

A threshold is not automatically a decision rule. If a vibration value exceeds a limit, the appropriate action may depend on load, speed, sensor validity, trend, redundancy, and safety procedures. Thresholds should therefore be linked to operating context and response plans.

Decision matrices can compare alternatives against weighted criteria. If alternative j has normalized score sᵢⱼ on criterion i and weight wᵢ, with weights summing to 1, a simple weighted score is:

This method assumes that criteria can be meaningfully scored on a common scale and that the weights represent decision-maker priorities. It is a transparent aid, not an objective substitute for judgment. Sensitivity analysis should test whether the preferred alternative changes when weights or scores vary.

10. Illustrative Case: Predictive Maintenance for a Conveyor Drive

Background

A packaging line uses a motor-driven conveyor. The maintenance team reports occasional belt stoppages and wants to determine whether vibration and motor temperature data can support earlier intervention. This is an illustrative case rather than a report of a specific company.

Problem

The team has three months of records containing timestamp, motor speed, load state, vibration RMS, motor temperature, alarm status, and maintenance events. The immediate decision is whether to schedule an inspection when vibration increases, while avoiding unnecessary shutdowns.

Analysis

The analyst first creates a data dictionary and confirms that vibration is in mm/s RMS, temperature is in °C, and timestamps are in UTC. Duplicate timestamps are found after two controller restarts. The raw records are retained, while exact duplicates are flagged and removed from the analysis table. Several temperature values are missing during communication interruptions. They are not interpolated for alarm decisions.

A line chart shows that vibration increases gradually over ten operating days, but one isolated spike occurs during a high-load test. The spike is checked against the maintenance log and confirmed as a genuine operating event. It is retained and annotated. A scatter plot suggests that vibration is higher at greater load, so comparisons are stratified by load state rather than treating all observations as equivalent.

The team summarizes daily median vibration, 95th percentile vibration, operating hours, and number of valid readings. The median reduces the influence of short transients, while the 95th percentile identifies worsening upper-end behavior. The dashboard displays data freshness, the percentage of missing readings, current vibration, recent trend, load condition, and maintenance history.

Approach and Solution

The team establishes a two-stage response. A warning requires verification of sensor health and review of the trend. A confirmed high reading under comparable load requires an inspection during the next planned stop. An emergency response is reserved for a validated value above the safety limit or for combined evidence such as abnormal noise, temperature, and vibration.

The decision rule is documented with units, operating conditions, responsible personnel, and escalation time. A weekly review compares alerts with inspection findings. This prevents the dashboard from becoming an isolated reporting tool.

Outcome

The analysis does not claim that vibration alone predicts failure. Instead, it provides earlier visibility and reduces ambiguity about when to inspect. The first review identifies that one apparent trend was partly caused by a change in load mix. After stratification, the remaining increase is smaller but still operationally relevant.

Learning Points

1. A trend must be interpreted in operating context.
2. Missing data and sensor condition belong on the dashboard.
3. Genuine extreme observations should not be deleted automatically.
4. An alert must specify an action and an owner.
5. Analytics supports maintenance judgment; it does not eliminate engineering responsibility.

11. Contemporary Practice and Employability

Engineers increasingly work with time-series databases, edge computing, cloud platforms, digital twins, industrial Internet of Things devices, and machine-learning services. These technologies can reduce latency and support fleet-wide analysis, but they also increase requirements for cybersecurity, data governance, interoperability, and model monitoring.

A mechatronics graduate should be able to inspect a CSV file, query structured data, calculate statistics, build a clear chart, explain a dashboard metric, and challenge an apparently convincing conclusion. Employers also value communication: an engineer must translate technical evidence into a decision that operators, managers, and safety personnel can understand.

The most transferable skills are precise problem formulation, unit awareness, data-quality judgment, statistical reasoning, reproducibility, visualization, documentation, and ethical accountability.

Best-Practice Checklist

Before releasing an analysis, ask:

• Is the decision clearly stated?
• Are variables, units, time zones, and denominators defined?
• Has the raw data been preserved?
• Are missing values, duplicates, and outliers documented?
• Are summaries appropriate for the measurement scale and distribution?
• Do charts show context, units, sample size, and relevant limits?
• Can another person reproduce the result?
• Have uncertainty, confounding, and bias been considered?
• Does each alert have an owner and a response?
• Are privacy, security, safety, and regulatory requirements satisfied?

Chapter Summary

Data analytics converts measurements into evidence for engineering decisions. A reliable process begins by defining the decision and the variables, then acquiring and assessing data systematically. Cleaning must address completeness, validity, consistency, accuracy, uniqueness, and timeliness without deleting genuine fault information.

Descriptive statistics summarize location, spread, distribution, and relationships. The mean, median, standard deviation, interquartile range, percentiles, rates, and correlation each have specific assumptions and interpretations. Visualization communicates trends, comparisons, distributions, and relationships, while dashboards connect indicators to actions.

Reproducible workflows preserve raw data, document transformations, validate results, and enable independent review. Uncertainty, sampling limitations, measurement error, confounding, and bias must be communicated honestly. Decision support combines analytical evidence with engineering knowledge, constraints, risk, and responsibility. AI and advanced analytics can extend capability, but they do not remove the need for data governance, validation, human oversight, or ethical judgment.

Key Terms

Data analytics: Systematic examination and transformation of data to produce useful information.

Decision support: Evidence and tools that help people compare alternatives and select actions.

Data dictionary: Documentation defining variables, units, types, sources, and valid values.

Descriptive analytics: Analysis concerned with what happened in observed data.

Diagnostic analytics: Analysis investigating why an outcome may have occurred.

Predictive analytics: Analysis estimating likely future outcomes.

Prescriptive analytics: Analysis recommending possible actions.

Mean: Arithmetic average of numerical observations.

Median: Middle value of ordered observations.

Standard deviation: Measure of typical dispersion around the mean.

Interquartile range: Difference between the third and first quartiles.

Outlier: Observation unusually distant from the main pattern.

Correlation: Standardized measure of linear association between two variables.

Dashboard: Visual interface presenting indicators and information for monitoring or decisions.

Reproducibility: Ability to obtain the same result using the same data and documented procedure.

Uncertainty: Limited knowledge about a measurement, estimate, or conclusion.

Bias: Systematic distortion causing results to differ from the relevant truth or population.

Confounding: Distortion caused when another variable influences an apparent relationship.

KPI: Key performance indicator linked to an objective or operational outcome.

Data governance: Policies and controls for data quality, access, security, use, and retention.

Review Questions

Short Answer Questions

1. Distinguish data, information, evidence, and a decision using a mechatronics example.
2. Why should an outlier be investigated rather than automatically deleted?
3. State one appropriate use for a line chart and one for a scatter plot.
4. What is the purpose of a data dictionary?
5. Why does correlation not establish causation?

Long Answer Questions

1. Explain a complete workflow for cleaning sensor data, including missing values, duplicates, units, and validity checks.
2. Compare the mean, median, standard deviation, and interquartile range, including situations in which each is useful.
3. Describe the design of an effective engineering dashboard for a production supervisor.
4. Explain reproducibility and identify practices that make an analytical project reproducible.
5. Discuss how uncertainty, sampling bias, and confounding can affect an engineering decision.

Analytical / Application Questions

1. A machine produces cycle times of 8, 9, 9, 10, and 19 seconds. Calculate the mean, median, range, and sample standard deviation, and explain which summary best represents a typical cycle if the 19-second event was a verified interruption.
2. A vibration dashboard shows a rising trend, but the machine operated under higher loads during the same period. Design an analysis to determine whether the trend is attributable to degradation, load, or both.
3. A team must choose between two sensors using accuracy, cost, installation time, and maintainability. Propose a weighted decision matrix, state its assumptions, and explain how sensitivity analysis could test the recommendation.

Multiple Choice Questions with Answers

Question 1

Which type of analytics asks what happened?

A. Descriptive analytics

B. Predictive analytics

C. Prescriptive analytics

D. Diagnostic analytics

Correct Answer: A. Descriptive analytics

Question 2

For the observations 2, 4, 6, and 8, what is the arithmetic mean?

A. 4

B. 5

C. 6

D. 20

Correct Answer: B. 5

Question 3

Which measure is generally least affected by a single extreme observation?

A. Mean

B. Range

C. Median

D. Variance

Correct Answer: C. Median

Question 4

What is the main purpose of a data dictionary?

A. To increase sensor sampling frequency

B. To define variables, units, and valid values

C. To replace all missing values automatically

D. To generate dashboard colors

Correct Answer: B. To define variables, units, and valid values

Question 5

A scatter plot is most appropriate for examining which situation?

A. The relationship between motor load and temperature

B. The names of maintenance technicians

C. A single machine serial number

D. The order of categorical alarm labels

Correct Answer: A. The relationship between motor load and temperature

Question 6

What does a correlation near zero indicate?

A. Certain absence of any relationship

B. A strong causal relationship

C. Little linear association, although nonlinear association may remain

D. That both variables have zero units

Correct Answer: C. Little linear association, although nonlinear association may remain

Question 7

Which practice most directly supports reproducibility?

A. Deleting the raw dataset after analysis

B. Changing calculations manually without recording them

C. Documenting transformations and preserving source data

D. Using a three-dimensional chart

Correct Answer: C. Documenting transformations and preserving source data

Question 8

Why should missing values not automatically be treated as zero?

A. A missing value may represent sensor failure rather than a measured zero

B. Zero is never a valid engineering value

C. All missing values are always outliers

D. Zeros cannot be stored in databases

Correct Answer: A. A missing value may represent sensor failure rather than a measured zero

Question 9

Which statement about correlation is correct?

A. It proves that one variable causes another

B. It measures linear association under its stated conditions

C. It can only have positive values

D. It is unaffected by outliers

Correct Answer: B. It measures linear association under its stated conditions

Question 10

In a decision-support dashboard, an alert should ideally include which additional element?

A. An unexplained decorative animation

B. A defined response, owner, and escalation rule

C. Only the rawest available data

D. A hidden calculation with no units

Correct Answer: B. A defined response, owner, and escalation rule