Access Official Technical PDF

Share Summary:LinkedInEmail

Fog computing for manufacturing offers an architectural approach to a familiar industrial challenge: collecting machine data at scale and turning it into useful information for equipment monitoring and prognosis. By placing computing and storage resources closer to factory-floor data sources, manufacturers can combine local data handling with cloud-based resources for computationally intensive analytics.

This approach is relevant to both legacy equipment, which may lack built-in sensing, and modern machines that produce data through controllers and connected sensors. A framework described in a 2017 manufacturing systems study brings together wireless sensor networks, gateway devices, private edge resources, public cloud computing, and machine learning. Its proof-of-concept demonstrations covered vibration and energy monitoring, as well as tool-wear prediction.

Why manufacturing monitoring needs a distributed architecture

Machine and process monitoring depends on access to timely, useful data. Yet collecting data across different equipment can be difficult. Legacy machines may have no sensors or networked controllers, while newer systems can use different communication protocols. That variation can complicate data exchange and integration.

Analysis also creates infrastructure demands. High-volume, high-speed data streams can require substantial storage, bandwidth, and computing capacity, particularly when machine-learning models must be trained on large datasets. Building and maintaining all of that capacity on site may be costly, while sending every raw data point to a centralized cloud can create network demands and delay.

Fog computing addresses this distribution challenge by placing computing, storage, and networking resources between factory-floor devices and centralized cloud data centers. The goal is not to replace the cloud, but to coordinate resources at different locations according to the work they need to perform.

How fog computing for manufacturing works

A fog-based monitoring system can be understood as a sequence of data collection, model development, and analysis. In the framework, the process has four main steps:

  1. Collect machine data. Sensors and communication adapters capture real-time data, with gateway devices acting as an interface between equipment and network infrastructure.
  2. Route data to suitable resources. Raw streams can be stored in a local private edge cloud, while selected sample or training datasets are sent to a public cloud for model development.
  3. Build diagnostic and prognostic models. Machine-learning algorithms use training data to develop models for tasks such as identifying conditions or predicting outcomes. Parallel computing can support larger workloads.
  4. Apply models to operating data. Once developed, models can be used to analyze live or stored machine data at the edge, supporting online monitoring and prognosis.

This division can reduce the need to move every raw dataset between locations. It can also allow manufacturers to keep sensitive data in a local environment while using public cloud resources for selected, compute-intensive tasks. The specific balance depends on a manufacturer’s infrastructure and data requirements.

Sensors, gateways, and protocols on the factory floor

The sensing layer can include devices that measure vibration, current, temperature, force, rotational speed, torque, or acoustic emission. Wireless sensor nodes can be useful for adding measurement capability to equipment that does not have suitable built-in sensors. A gateway collects data from those nodes and connects them to local or cloud-based systems.

In the proof-of-concept system, accelerometers measured vibration, while current transducers measured electrical current. The resulting measurements were used to monitor operating conditions and energy consumption. Sensor nodes transmitted data wirelessly to a gateway, which then sent it to a private cloud environment. The work demonstrated monitoring of pumps and CNC machines, including legacy equipment.

Communication standards and protocols also matter. Interoperability can be improved by using established interfaces, such as MTConnect, where supported. However, data available through a machine controller may not cover every monitoring need. Additional sensors may be necessary to measure conditions such as spindle vibration.

Using machine learning for equipment prognosis

Monitoring describes equipment or process conditions; diagnosis focuses on detecting and identifying faults; prognosis aims to estimate how a fault may progress or when a component may no longer perform its intended function. Remaining useful life is one term used for that predicted period. Data-driven methods can support these tasks by learning from measured data rather than relying solely on a complete physical model of the equipment.

Machine learning requires suitable training data and a way to evaluate model performance. In supervised learning, examples with known outcomes are used to train a model, and separate test data can be used for evaluation. Once assessed, a model can be applied to new measurements. Large training datasets can demand significant memory and processing capacity, which is one reason cloud and parallel computing are considered for manufacturing analytics.

The study demonstrated a random-forest model for predicting tool wear in milling operations using recorded force, vibration, and acoustic-emission signals. This illustrates how sensor data can feed a predictive model; it does not mean that the same model or input signals will suit every machine, process, or maintenance decision.

Practical considerations for implementation

For engineering and operations teams evaluating this architecture, several design questions deserve attention:

  • What data is needed? Select measurements based on the machine condition or process outcome to be monitored, and determine whether existing controller data is sufficient.
  • Where should data be processed? Consider which workloads belong on a local edge system and which can be performed using public cloud resources.
  • How will systems interoperate? Assess equipment interfaces, communication protocols, sensor adapters, and gateway requirements before scaling deployment.
  • How will models be validated? Define training and test datasets and evaluate whether a model is appropriate for its intended diagnostic or prognostic task.

A hybrid architecture can offer a way to combine local handling of machine data with scalable computing, but it still depends on sound sensing, reliable data integration, and appropriate model development. The framework is an architectural foundation, not a substitute for validating analytics against the equipment and operating conditions they are intended to support.

Conclusion

Fog computing for manufacturing connects machine-level data collection with edge and cloud resources for monitoring and predictive analytics. Wireless sensors and gateways can extend visibility to equipment, while a hybrid computing model can keep selected data local and provide access to additional processing capacity. For manufacturers, the practical implication is a path toward more distributed machine health monitoring and prognosis—provided that sensing, interoperability, data routing, and model validation are planned together.

Access Official Technical PDF

About The Author