TensorFlow Lite is Google's toolkit for running trained machine learning models on small devices, and it has since developed into LiteRT, which Google describes as a high-performance, cross-platform inference runtime for edge devices. In manufacturing, you train a model with TensorFlow or another framework, convert it to the compact .tflite format, and run it on an industrial PC, gateway or microcontroller next to the machine.
What is the workflow from data to a deployed factory model?
Deploying deep learning is mostly engineering around the model. A dependable pipeline has seven steps:
- Define the decision. "Flag a bearing likely to fail within a week" or "reject cracked castings." A vague goal gives a vague model.
- Collect and label data. Record sensor signals or images with the operating context and the engineers' judgement of what was good or bad.
- Split data honestly. Hold out whole days, shifts or machines for testing, not random rows, otherwise the model looks better than it is.
- Train and evaluate. Build a baseline first, such as a simple threshold or a gradient-boosted model, so you know whether deep learning adds anything.
- Convert and optimise. Export to .tflite and reduce size and latency with quantisation.
- Deploy and integrate. Run on the edge device, expose results over MQTT or OPC UA, and let the PLC or SCADA act on them.
- Monitor and retrain. Track input data, predictions and outcomes; retrain when the process or the camera changes.

What does quantisation do, and what are the risks?
Quantisation stores a model's numbers at lower precision so it is smaller and faster. Google's documentation describes three post-training options:
- Dynamic range quantisation: weights become 8-bit, about 4 times smaller with roughly 2 to 3 times faster CPU inference, and no calibration data is needed.
- Full integer quantisation: weights and activations become integers, about 4 times smaller with larger speed-ups and support for accelerators and microcontrollers. It needs a representative dataset of roughly 100 to 500 samples.
- Float16 quantisation: about half the size, mainly useful for GPU acceleration.
The same documentation warns that accuracy can fall, particularly for smaller networks, so you must test the quantised model against your acceptance limits, and use quantisation-aware training if the drop is too large. For an inspection model, re-check the escape rate and false reject rate after conversion, not just overall accuracy.
Should you use TensorFlow or PyTorch?
Either will do. TensorFlow with Keras has mature deployment tooling and a long history in edge deployment. PyTorch is widely used in research and many open-source vision models. LiteRT supports direct export and quantisation from PyTorch, TensorFlow and JAX into the .tflite format, so the framework you train in no longer locks you to one runtime. Pick the one your team and your data science tools already know, and focus effort on data and validation instead.
Also consider whether deep learning is needed at all. For many sensor problems, such as vibration features on a motor, a simpler model is easier to explain, runs on a tiny device and is quicker to validate. Our predictive maintenance guide shows how far simple methods go.
What manufacturing problems suit TensorFlow-style models?
- Image classification and defect detection for surface flaws, assembly checks and sorting.
- Time-series classification of vibration, current or acoustic data for condition monitoring.
- Anomaly detection using autoencoders trained on normal operation.
- Tiny models on microcontrollers that listen for a bearing noise or detect an unusual vibration signature without a network connection.
What should you check when you convert a model?
Conversion is where many projects lose time. Confirm that every layer in your model is supported by the runtime on your target device, since unusual custom operations may not convert. Compare outputs of the original and converted models on the same few hundred inputs and look at the largest differences, not just the average. Measure latency on the real device, not on your laptop, with the camera or sensor feed attached and other applications running. Check memory use over hours, not minutes, to catch leaks. If the target is a microcontroller, check that the model fits in flash and RAM with room to spare for the rest of the firmware.
Keep the preprocessing identical between training and deployment. A different image resize method, colour order or sensor scaling is a classic source of models that work in the notebook and fail on the line.
How do you keep a deployed model trustworthy?
Treat it as a controlled piece of equipment. Version every model and record the training data and settings that produced it. Test it on a fixed reference set before each release. Log inputs, outputs and the human decision that followed so you can measure real-world performance. Alert when the input data drifts from what the model saw in training, such as a new product, a replaced camera or a sensor recalibration. Define a fallback: if the model is unavailable, the process should continue safely on conventional control or manual inspection.
Connecting a model to control equipment should follow the site's management of change process. Our edge computing guide covers the hardware and security side.
When is a gateway enough, and when do you need a GPU?
A model that classifies a few vibration features once a second runs happily on a modest industrial PC or even a microcontroller. An image model analysing several frames a second at full resolution will usually want a GPU or an accelerator. Work out the numbers first: images per second, image size and acceptable latency, then test a candidate device with your actual model. Buying hardware before measuring is the most common and most expensive mistake in edge AI projects. NVIDIA's Jetson range, for example, spans entry-level modules to high-end ones, and its own figures are not comparable across generations, so benchmark with your workload.
What skills and jobs does this lead to?
ITJobsWatch reported a median UK salary of £75,000 for permanent machine learning roles across 2,526 advertisements in the six months to 23 October 2026. Most of those roles are in software and data teams, but industrial employers increasingly need engineers who can take a model from a notebook to a machine. The scarce skill is deployment and monitoring, not building yet another classifier.
How to learn TensorFlow and edge deployment for free
- Start with the free Deep Learning with TensorFlow course, then compare with Deep Learning with PyTorch.
- Build foundations in Machine Learning with Python and learn deployment in MLOps, Model Deployment and Monitoring.
- Browse all free AI courses.
- Practise in the Machine Learning Engineer job simulation and the Computer Vision Engineer job simulation. They are practice, not employment.
- Check the AI and machine learning jobs page for the skills employers list.
Start free
Begin with the free Deep Learning with TensorFlow course, try the first five free days of the Machine Learning Engineer job simulation, and look at the AI and machine learning jobs page. If you want a self-paced route with five graded industrial projects, see the EDWartens AI Module for its current price and syllabus.





