What is Deep Learning?
Deep Learning is a form of Machine Learning in which artificial neural networks with many successive layers learn by themselves which features in raw data, such as images, audio, text or sensor signals, matter for making a prediction or decision. “Deep” refers to the depth of the network: the number of hidden layers between input and output. Deep Learning is particularly strong at recognising complex patterns in unstructured data and is the core of many modern AI applications, from visual inspection to language models. In industry it supports applications in both IT and OT.
🧠 How does Deep Learning work?
A neural network consists of layers of simple computing units (neurons). Each neuron multiplies its inputs by weights, adds a bias and passes the result through a non-linear activation function. The structure is always the same:
- Input layer - receives the raw data, for example the pixels of a camera image or a window of vibration measurements
- Hidden layers - each layer learns more abstract representations: in an image first edges, then shapes, then complete objects or defects
- Output layer - produces the result, such as a class (“good”/“scratch”), a probability or a predicted measurement
During training, the network compares its output with the correct answer. The backpropagation algorithm, popularised in 1986 by Rumelhart, Hinton and Williams, calculates how much each weight contributed to the error. An optimisation method such as gradient descent then adjusts the weights step by step. This repeats over many thousands of examples and requires a lot of compute, which is why training takes place on GPUs in a data centre or the cloud.
Training is followed by inference: the frozen model applies what it has learned to new data. Inference is much lighter and in the factory often runs on an edge device or industrial PC close to the machine.
🔄 What is the difference between Machine Learning and Deep Learning?
Deep Learning is a subfield of Machine Learning. The biggest difference lies in feature engineering: with classical algorithms an engineer decides which features the model receives, whereas with Deep Learning the network learns those features from the raw data itself.
| Aspect | Classical Machine Learning | Deep Learning |
|---|---|---|
| Features | Selected manually by an expert | Learned automatically from raw data |
| Data required | Often hundreds to thousands of records | Thousands to millions of examples, fewer with transfer learning |
| Compute | A standard CPU is sufficient | GPU or AI accelerator for training |
| Strong at | Tabular data, process data, time series with clear features | Images, audio, text, vibration spectra |
| Explainability | Relatively good (decision trees, regression) | Limited, a “black box” |
| Examples | Random forest, gradient boosting, regression | CNN, LSTM, transformer, autoencoder |
For tabular and process data from sensors, classical algorithms such as gradient boosting often perform as well as a neural network, with less data and better explainability. So do not choose Deep Learning because it is modern, but because the data is unstructured.
📜 How did Deep Learning emerge?
Neural networks have existed since the 1950s, but only around 2012 did they become practical, thanks to three factors: large labelled datasets, affordable GPUs and better training methods.
| Year | Milestone | Significance |
|---|---|---|
| 1982 | Hopfield network | Associative memory that stores and reconstructs patterns |
| 1986 | Backpropagation (Rumelhart, Hinton, Williams) | Practical way to train networks with hidden layers |
| 1997 | LSTM (Hochreiter and Schmidhuber) | Recurrent network that remembers long-range dependencies in sequences |
| 2012 | AlexNet wins ImageNet | Top-5 error of 15.3%, more than 10 percentage points better than the runner-up, trained on two GPUs |
| 2017 | Transformer (“Attention Is All You Need”, Google) | Foundation of modern language models such as GPT |
| 2018 | Turing Award for Bengio, Hinton and LeCun | Recognition of deep neural networks as a critical component of computing |
| 2024 | Nobel Prize in Physics for Hopfield and Hinton | For foundational discoveries that enable machine learning with artificial neural networks |
🧱 Which network types matter most?
| Model type | How it works | Industrial use |
|---|---|---|
| Convolutional Neural Network (CNN) | Filters slide across an image and recognise local patterns | Image and video analysis, Vision inspection, OCR of lot codes |
| Recurrent Neural Network (RNN) / LSTM | Remembers earlier steps in a sequence | Time series, speech, forecasting process variables |
| Transformer | Attention mechanism processes a whole sequence at once | Text generation, translation, code generation, vision transformers |
| Autoencoder | Compresses and reconstructs normal data | Dimensionality reduction, anomaly detection |
| Generative Adversarial Network (GAN) | Two networks compete: generator against discriminator | Image generation, synthetic training data, simulations |
The autoencoder is particularly useful in industry: it trains only on good products or normal process behaviour. A high reconstruction error then indicates a deviation, even if that defect type has never occurred before.
🏭 Where is Deep Learning used in industry?
| Application | Data | Example |
|---|---|---|
| Computer vision | Camera images | Detecting scratches, cracks or missing parts with Vision AI |
| Predictive Maintenance | Vibration, temperature, current | Recognising bearing damage in vibration spectra through condition monitoring |
| Anomaly detection | Process data, network traffic | Spotting deviations in behaviour, processes or cyber activity |
| Natural Language Processing | Failure reports, manuals | Chatbots and document analysis for maintenance teams |
| Audio recognition | Microphones near machines | Speech recognition, sound classification of pumps and compressors |
| Robotics and self-driving vehicles | Camera, lidar | Recognising objects and motion for robots and AGVs |
| Soft sensors | Historical process data | Estimating a hard-to-measure quality value from other measurements |
Training data often comes from a historian or time series database. That makes Deep Learning a building block of Industrial AI, digital twins and Industry 4.0.
🔧 How do you deploy Deep Learning in an OT environment?
A successful project starts small and stays close to the process question. A practical approach:
- Define the question - for example “what percentage of scratch rejects do I want to find automatically?” and decide how you measure success (detection rate, false rejects).
- Collect representative data - all product variants, lighting conditions, seasons and fault states. A model only learns what it has seen.
- Label carefully - let operators or quality staff do the labelling. Errors in labels become errors in the model.
- Use transfer learning - start from a pre-trained network and retrain only the last layers. For visual inspection this often means hundreds rather than millions of images.
- Validate on separate test data - data the model has never seen, including edge cases.
- Optimise for the edge - export to the open ONNX format (launched in 2017 by Facebook and Microsoft) and quantise to INT8. This makes the model roughly four times smaller than in FP32.
- Manage the model in production - monitor accuracy, retrain after product or process changes and record every model version through change management.
Typical values in practice:
| Parameter | Typical value |
|---|---|
| Inference time per image on an edge GPU | A few to a few tens of milliseconds |
| Compute of an edge module (e.g. NVIDIA Jetson Orin) | 34 to 275 TOPS at 7 to 60 W |
| Labelled images per defect class with transfer learning | Hundreds to a few thousand |
| Model size after INT8 quantisation | About a quarter of the FP32 model |
| Public anomaly detection benchmark (MVTec AD) | More than 5,000 images in 15 categories |
A deep learning model is not strictly deterministic in its execution time. Hard real-time actions, such as an ejector at a fixed position, therefore stay with the PLC, which uses the model output only as an input. Inference through edge computing keeps latency low and keeps the line running if the connection to the data centre drops.
🔐 What risks and rules apply to Deep Learning in OT?
Technical considerations:
- Data and compute - training requires large volumes of labelled data and GPU capacity
- Black box - a neural network is less explainable than classical algorithms; operators must be able to understand and override a recommendation
- Overfitting - an overly complex model memorises the training data and fails in new situations
- Model drift - wear, new raw materials or different lighting gradually make a model inaccurate; monitoring and retraining are required
Security: a vision PC or edge device with GPU drivers and AI libraries enlarges the attack surface of the OT network. In NIST AI 100-2 E2025, NIST describes attacks such as evasion (manipulating input) and poisoning (contaminating training data). Protect data pipelines and model files like any other OT system, safeguard data integrity and watch for supply chain risk when using pre-trained models. The international guidance on AI in OT of December 2025, co-authored by the Dutch NCSC, urges great caution with AI in safety-critical decisions; according to the guidance, language models should almost certainly not make safety decisions. The SIS remains deterministic.
Regulation:
- AI Act - since 2 February 2025, organisations must take measures to ensure the AI literacy of their staff. Deep Learning as a safety component counts as high risk. After the digital omnibus, those requirements apply from 2 December 2027 for critical infrastructure and from 2 August 2028 for machinery.
- Machinery Regulation - from 20 January 2027, safety components with self-learning behaviour require assessment by a notified body.
- Cyber Resilience Act - products with digital elements, such as smart cameras and AI software, must meet cybersecurity requirements from 11 December 2027.
- Standards - ISO/IEC 22989:2022 defines AI concepts and terminology, and ISO/IEC 23053:2022 describes a framework for AI systems using machine learning.
❓ Frequently asked questions
What is the difference between AI, Machine Learning and Deep Learning?
AI is the umbrella term for systems that perform tasks that normally require human intelligence. Machine Learning is the subfield in which models learn from example data. Deep Learning is a form of Machine Learning that uses multi-layer neural networks which learn features from raw data by themselves.
How much data does Deep Learning need?
A Deep Learning model trained from scratch often needs tens of thousands to millions of examples. With transfer learning, where you retrain a pre-trained network, industrial image inspection often works with hundreds to a few thousand labelled images per defect class. More important than the volume is that the data covers all variants and fault states.
Does Deep Learning always need a GPU?
For training, a GPU or AI accelerator is essential in practice, because the network repeatedly adjusts millions of weights. For inference, a compact edge module or an industrial PC with a small GPU or NPU is often sufficient. A quantised model can sometimes even run on a standard CPU.
Can Deep Learning replace a PLC or safety system?
No, Deep Learning does not replace a PLC or safety system. A deep learning model is probabilistic and not deterministic in its execution time, so the PLC performs the control and uses the model output only as an input. Safety functions stay in a certified SIS or safety relay.
When should you choose classical Machine Learning instead of Deep Learning?
Classical Machine Learning is usually the better choice for tabular and process data with clear features, such as temperatures, pressures and flow rates. Algorithms such as gradient boosting need less data and are easier to explain. Deep Learning pays off mainly for images, audio, text and vibration spectra.
Is Deep Learning covered by the AI Act?
Yes, every AI system based on Deep Learning falls under the AI Act, but most industrial applications such as quality inspection and predictive maintenance fall into a low risk class. If Deep Learning becomes a safety component of a machine or of critical infrastructure, high-risk requirements apply from 2 December 2027 or 2 August 2028. In addition, every organisation that deploys AI has had to take measures for the AI literacy of its staff since 2 February 2025.
📌 In summary
Deep Learning is a powerful AI technology that uses multi-layer neural networks to learn patterns and decisions from complex, unstructured data.
It is the engine behind many applications in modern industry, such as computer vision, predictive maintenance and anomaly detection. Success in OT requires representative data, inference at the edge, a PLC that stays in control, attention to security and compliance with the AI Act and the Machinery Regulation.