Data fusion can improve reliability by allowing one source of information to validate another

Harshal Pore, Managing Director, Basler India.

How are advances in industrial sensors and machine vision changing the way smart factories capture, interpret and use real-time information from the physical production environment?

Industrial sensors and machine vision systems are becoming more capable, faster and more connected. Cameras can capture higher-resolution images at higher speeds, while modern interfaces and processing platforms allow image data to be analysed much closer to the machine.

The important shift, however, is not simply that factories are collecting more data. It is that different sources of data can increasingly be interpreted together.

Consider a production line where a vision system begins detecting a higher number of surface defects. On its own, the camera can identify the problem, but not necessarily its cause. If that visual information is correlated with machine sensor data – temperature, vibration, cycle time or tooling history – a pattern may emerge. The manufacturer can then begin distinguishing an isolated quality issue from a developing process problem.

This is where data fusion becomes valuable. The camera provides visual context; conventional sensors provide physical measurements; production systems provide historical and operational context. Together, they provide a more complete picture of the production environment.

What is the significance of combining data from sensors, machine vision, PLCs, robots and other industrial systems, and how does data fusion improve the accuracy and reliability of manufacturing decisions?

The real value of industrial data appears when individual signals are no longer treated in isolation. A sensor may indicate that a condition has changed, a camera may show what changed physically, while a PLC or robot controller can provide the sequence and operating state in which the event occurred.

Combining these sources gives manufacturers something that a single data stream cannot provide: context.

For example, suppose a vision system detects that a component is slightly out of position. The image alone may indicate a deviation, but PLC status can reveal whether the machine was in a normal production cycle, a robot controller can show the actual movement and timing, and sensor data can indicate whether any process was outside its normal range. Taken together, these signals can help distinguish a genuine process anomaly from an expected variation or an inspection error.

Data fusion can also improve reliability by allowing one source of information to validate another. In a well-designed system, a vision result does not necessarily have to be accepted or rejected in isolation. It can be checked against machine state, sensor readings or production history.

This is particularly important as manufacturers adopt AI and machine learning. Models become more useful when they have access to meaningful contextual variables rather than images or sensor readings alone. The objective is not to create one enormous pool of data, but to connect the right data points at the right moment.

In practical terms, this creates a common operational picture of the production process. Decisions about quality, machine intervention, process adjustment and traceability can then be based on multiple independent signals rather than a single measurement.

The architecture is also becoming more open. Industrial cameras, controllers, robots and software increasingly need to exchange information across vendor boundaries. Standards and interfaces such as GenICam, GigE Vision, OPC UA and industrial Ethernet are therefore important building blocks for a data-fused factory.

How is machine vision evolving from a conventional quality-inspection tool into a broader source of process intelligence – for example, in assembly verification, process monitoring, traceability and predictive quality?

Machine vision has traditionally answered a relatively straightforward question: is the product acceptable or defective? That remains one of its most important applications, but the role of vision is expanding.

Modern vision systems can verify whether the correct component has been assembled, check orientation and position, confirm the presence of fasteners, read codes and characters, track products through production and identify process deviations. AI-based vision solutions can also help recognize patterns and anomalies that are difficult to define using conventional rule-based inspection.

This changes the question being asked of the vision system. Instead of only asking, “Is this product defective?”, manufacturers can ask, “What is happening in this process, and what can the visual information tell us about it?”

When these records are analysed over time, manufacturers can identify recurring patterns and process drift. This creates the basis for predictive quality: intervening before a deviation affects a larger number of products.

This broader role is reflected in the evolution of the vision ecosystem itself. Basler, for example, has expanded beyond image acquisition into software, AI-based image processing and embedded vision, supporting applications where image data needs to become part of a wider automation and analytics workflow.

How can the combination of sensor data, visual information and historical production data help manufacturers detect anomalies early and move from reactive maintenance towards predictive and even prescriptive maintenance?

The strongest maintenance insights often emerge when different types of evidence are considered together.

Imagine a machine where vibration levels have gradually increased. At the same time, a vision system begins detecting small changes in component positioning. Neither signal alone may be sufficient to justify stopping & correcting the machine. When combined with historical production data, however, the relationship may become significant.

If similar combinations of signals have previously preceded a bearing failure, alignment problem or quality deviation, the system can flag the condition earlier than a conventional reactive maintenance approach.

This supports a progression from reactive maintenance to predictive maintenance. With sufficiently mature data and models, manufacturers can move towards prescriptive maintenance as well – not simply identifying that a problem is developing but helping determine what intervention should be considered and when.

The objective should not be to remove engineering judgement from the process. AI and analytics are most useful when they give maintenance and production teams better evidence on which to base decisions.

What technologies and architectures – such as edge computing, AI/ML, industrial IoT and real-time analytics – are needed to convert the enormous volume of sensor and vision data into actionable insights at the shop-floor level?

One of the practical challenges is the sheer volume of data generated by modern cameras and sensors. Continuously transmitting every image to a central server or cloud environment can create problems around bandwidth, latency, cost and cybersecurity.

This makes edge computing increasingly important. Processing can take place close to the machine, where the system can analyse images, run AI models and combine results with sensor information. Only the relevant result, metadata or selected images then need to be passed to higher-level systems.

Interoperability is equally important. Industrial environments typically contain equipment from multiple vendors, so technologies and standards such as GenICam, GigE Vision, OPC UA, MQTT and industrial Ethernet can help connect different layers without creating isolated data islands.

Basler's pylon Software Suite and its work in embedded vision are examples of this broader direction. The emphasis is moving from simply acquiring an image to providing the software, processing and integration capabilities required to turn image data into useful information within an industrial environment.

The key is to process data at the right level. Time-critical decisions should generally happen at the edge, while aggregated information and historical data can be passed to higher-level analytics systems.

Beyond improving quality and reducing downtime, what measurable benefits can manufacturers expect from integrating sensing, machine vision and data fusion, and what are the key challenges they must overcome to achieve meaningful RoI?

The business case ultimately has to go beyond the number of cameras installed or the volume of data collected.

Depending on the application, integrated sensing and vision can contribute to higher first-pass yield, lower scrap and rework, improved OEE, faster inspection, better traceability and reduced unplanned downtime. It can also reduce dependence on repetitive manual inspection and provide more consistent process information across shifts and production locations.

However, achieving these benefits requires more than deploying technology.

Poor lighting can undermine an otherwise capable vision system. Inconsistent sensor data can make analytics unreliable. Legacy equipment may not expose the required information. AI models require representative training data and ongoing validation. Connecting OT systems to IT infrastructure introduces cybersecurity considerations, and organisations need people who understand both automation and data technologies.

There is also a more fundamental question manufacturers should ask before starting a data project: what decision are we trying to improve?

If that question cannot be answered, collecting another million images or adding another group of sensors is unlikely to create meaningful RoI. The strongest implementations usually begin with a clearly defined production or business problem and then determine what combination of sensing, vision, analytics and automation is actually required.

Ultimately, the future of the smart factory is not about seeing more or measuring more for its own sake. It is about making better use of the information already being generated. When visual information, sensor data and production history are brought together, the factory can move from simply detecting events to understanding patterns, anticipating problems and responding earlier.

That is where machine vision and industrial sensing are heading: from isolated inspection and measurement technologies towards an integrated layer of process intelligence.

(The views expressed in interviews are personal, not necessarily of the organisations represented.)

Mr Harshal Pore is an avid entrepreneur and business leader with over 18 years of experience in Industrial Automation and Computer Vision technologies. He leads Basler India (formerly Alphatechsys Automation), helping customers address complex machine vision challenges and leverage vision technologies to improve production efficiency, quality, and automation.