What is Fault Detection?

Some of the more common industries using fault detection include:

• Aerospace
• Automotive
• Electronics
• Industrial equipment
• Semiconductor
• Industrial manufacturing
• Materials processing
• Medical device
• Power generation and distribution
• Software

In many of these industries, especially aerospace, medical, and automotive, fault detection during design and operation is required because failures can be catastrophic and life-threatening. Other industries, such as semiconductor IC design and manufacturing, use fault detection as an integral part of the design process because making changes once manufacturing has started is prohibitive. And for all industries, the process is a cost-effective way to reduce redesign, rework, and repair costs.

Fault detection is the first step in identifying, understanding, and resolving faults in systems with techniques such as fault detection and diagnosis (FDD), fault detection and isolation (FDI), or fault detection, isolation, and recovery (FDIR). That is why fault detection is much more than looking for malfunctions. As part of the comprehensive FDIR process, it helps companies by detecting faults before they cause significant damage or degrade system performance. It provides design engineers, facilities managers, and maintenance teams with the information they need to make effective decisions about preventive maintenance and other corrective actions before abnormalities drive up maintenance costs, cause breakdowns, or result in injuries. 

Engineers who carry out fault detection categorize the multitude of available methods into a set of standard approaches based on where data comes from and how it is interpreted. Each approach has advantages and disadvantages, and engineers try to select the right one for the situation under evaluation.

Below are the most common approaches to fault detection.

Threshold-Based

This approach looks at system data and compares each parameter’s real-time value with predetermined limits. An alarm triggers when the limits are exceeded. A simple example is a temperature sensor on a car engine that turns on a warning light when engine coolant temperature exceeds 110 degrees Celsius and an error message when it exceeds 120 C. This methodology is simple, but it only detects faults that the system is designed to look for.

Model-Based

Engineers use mathematical models of the system to predict outputs from its real-world dataset in model-based fault detection. The monitoring tools calculate the difference between the predicted and actual values, called the residual. The system registers a fault when the residual exceeds the specified threshold. Some advantages of the model-based approach include the ability to detect unexpected faults and the use of historical data as the model. The downside is that creating the model can be difficult for complex systems.

Signal Processing

Instead of looking at the values in a data stream, this method looks at the frequency content of a signal and how it varies over time. The process uses a fast Fourier transform, statistical process control, or wavelet analysis to detect changes to the normal frequency pattern produced by each signal. It is very useful for mechanical fault identification, especially for rotating machinery, where vibration or acoustic changes can predict an equipment failure before it happens. The downsides are the complexity of signal processing and sensitivity to signal noise.

Data-Driven or Machine Learning Detection

This fault detection method uses historical data or data generated from simulation or a digital twin to create a representation of the system's behavior. The approach flags any data that doesn’t match the “normal” data as a deviation. This is not only a powerful failure detection method — it can also be used in predictive maintenance or as a diagnostic tool before a fault occurs. The advantages of this approach include handling highly complex systems well and identifying unexpected faults. The downside is the large amount of data needed and the black-box nature of the strategy.

Knowledge-Based Detection

A slightly different concept is to use logic or decision trees to evaluate a combination of data or signals rather than a single value or signal. This approach can use rule-based systems, fuzzy logic, fault trees, and expert systems. It works well with simpler, well-defined systems where data is scarce when a cascade of events leads to a fault. The downside is the amount of time needed to create the knowledge base and that it is limited to known fault modes.

Fault detection is the first step in a full fault management process. Once an issue is identified, the responding team should follow a standard process to control, understand, and resolve it. Every industry, and often every company in an industry, will have a different fault management process that works from identification to resolution and feeds back what was learned into the organization.

Although terms will vary from application to application, most steps in dealing with an identified fault fit into one of the following categories.

Fault Detection

There are multiple ways of implementing fault detection, and an important part in the full fault management process is ensuring that all detected potential faults are captured and passed along for further investigation and actions.

Fault Isolation

The investigating team needs to determine where the fault occurred in the system and identify the specific assemblies, subsystems, components, or routines affected by it. This often includes looking at sensor measurement data from other parts of the system to confirm an actual fault or eliminate a false detection.

Fault Identification and Diagnoses

In this step, the team determines what the fault was (identification) and why it happened (diagnosis). An important aspect of this step is also classifying the fault type — whether it was due to the sensors, the software, the machinery, a manufacturing defect, or another aspect of the system. An output of this step is also the determination of the cause and the assignment of a severity (such as minor, degrading, critical, safety) to the fault.

Fault Notification and Reporting

Once the above information is gathered, the team needs to create the appropriate documentation and notify the relevant parties of the fault, where it occurred, why it occurred, its severity, and who is impacted.

Fault Response

The fault response step, often executed in parallel with other steps, or right after detection, involves the team taking immediate action to mitigate the fault’s impact. More sophisticated control systems may have automatic response algorithms built into their software. It is also common to have standard operating procedures for operators to intervene when a fault triggers an alarm. The response may include shutting down all or part of the system, using redundant capabilities, or downgrading operations to keep the system running but at a diminished capacity. The goal of the fault response is to contain the fault while keeping as much functionality as possible.

Fault Recovery

The fault is fixed in the fault recovery step. This may involve a work order to repair or replace components, restarting software or hardware, adjusting operating parameters, or making changes to the hardware design or software code (bug fixes).

Fault Review

Once operations are restored, the team will conduct a post-fault review or a full root cause analysis or use the information for functional safety assessments. The goal of this step is to minimize the likelihood of the fault recurring, recommend corrective actions to recover and reduce its impact when it does, and provide information to decision-making stakeholders for related faults.

By its nature, fault detection involves gathering data about a system and monitoring it to detect faults. In many applications, fault detection systems fall into the big data category because of the volume and complexity of the available information. Where traditional data analytics struggles with complex data, nonlinear system behavior, unknown faults, and gradual degradation, artificial intelligence (AI) techniques are ideal for these situations. AI methods can learn about faults, adapt their analytics, detect faults that humans or traditional data analytics would miss, and provide early detection of faults. The amount and complexity of data created by Internet of Things (IoT) implementations make AI solutions essential for accurate fault detection.

Most current implementations of AI in fault detection fall under the deterministic subset of AI — machine learning. In machine learning, historical data or synthetic data from simulation is used to train systems to recognize faults. Engineers are also increasingly using large language models (LLMs) to support deep learning, data summarization, log analysis, and root cause analysis.

Some good examples of AI techniques for fault detection are:

• Supervised learning: Learns from labeled data using well-established algorithms.
• Unsupervised learning: Learns by building a normal model, then looks for deviation.
• Semisupervised learning: Learns through both supervised and unsupervised learning.
• Deep learning: Learns by training neural networks to automatically recognize patterns.
• Reinforcement learning: Learns by trial and error in an automated fault detection and recovery system.
• Physics-informed neural networks (PINNs): Learns by trial and error using physical laws.
• Digital twins: Learns by comparing the actual data with simulated data. Digital twins also support predictive maintenance.

Engineers have a variety of approaches for implementing fault detection in their applications. Although details vary across industries, the following suggestions for effective implementation of fault detection solutions apply.

1. Identify scope: Define the systems covered, what is considered a fault, the consequences of missing a fault or of false positives in the system, and the severity of each type of fault.
2. Understand the data: Identify all sensors, missing sensors, what data is produced, how to access it, data accuracy, and where faults can be found in the data.
3. Choose simulation solutions: Determine the role simulation will play in fault detection, including its use of digital twins, physics models, and model-based fault detection. Then identify the simulation tools the team will need.
4. Define the fault detection approach: Decide which fault detection approach or collection of approaches best fits the system.
5. Build a baseline solution: Establish an easy-to-deploy baseline solution, usually with one of the simple approaches, and measure the performance.
6. Create the full fault detection solution: Construct a robust, comprehensive fault detection solution that leverages information from the baseline solution and incorporates more sophisticated AI tools.
7. Design to support a fault management process: Include the needs of the full fault management process for the system in the design of the baseline and full solutions. This includes defining the approach to root cause analysis, what role functional safety will play, and how and when key stakeholders are informed of faults.
8. Focus on actionable information: Establish a clear definition of the information that automatic algorithms and human decision-makers need and ensure that all outputs from the solution produce usable information.
9. Validate and iterate: Use an interactive, continuous improvement approach to monitor and refine the solution.

Supporting Implementation With Simulation and Functional Safety Tools

A key part of most modern fault detection and diagnosis implementations is the use of both simulation and functional safety tools as part of a comprehensive solution. Teams use tools like the Ansys Twin Builder simulation-based digital twin platform for mechanical systems or Synopsys Electronics Digital Twin (eDT), Synopsys Virtualizer, and Synopsys Platform Architect for integrated circuits to create the virtual representations they need for model training or as baseline models for comparison with real-time data. In addition, multiphysics tools like Ansys Mechanical structural finite element analysis software, Ansys Fluent fluid simulation software, and Ansys HFSS high-frequency electromagnetic simulation software are used to provide physics-based behavior to the diagnostic tools.

Functional safety tools like Ansys medini analyze system-oriented safety analysis software for mechanical systems and Synopsys VC tools, including VC Functional Safety Manager, VC Z01X, VC Formal FuSa, and VCS Functional Verification Solution, are used to meet various international standards for functional safety in electrical and integrated circuits.

Combining Fault Detection With Other Processes

Many applications, especially industrial and building systems, benefit from using the same data to optimize energy consumption, improve performance, and enable predictive maintenance. For example, building automation systems and HVAC systems implementations are leading the way in deploying FDD software to provide fault detection alongside energy management. Manufacturing applications, especially in highly automated systems and robotics, combine their fault detection with performance optimization.

Fault Detection in IC Design and Manufacturing

Fault detection in integrated circuits is unique because much of it occurs at the design stage, before manufacturing. The goal of this approach is to identify faults caused by logic errors, signal integrity issues, or voltage drop. The cost of changing a chip’s design is prohibitive, so virtual representations are used with test patterns to find and remove potential faults before the design is validated and put into production. Engineers deploy a mix of tools, such as the Synopsys VC Z01X Fault Simulator, Synopsys ZeBu Server, and Synopsys TestMAX, to exercise the design and detect faults using the same real-world detection methods.

The key to effective fault detection in IC design is to integrate these tools into the design flow, including generating the test patterns for virtual testing. Rigorous testing is also conducted after manufacturing and before shipping to identify any faults caused by manufacturing issues. Once detected, the team uses fault management to isolate and address the fault, including scrapping the chip, downgrading the clock speed, or turning off faulty parts of the chip.