ThisAI Glossary is DDS's practical resource for understanding the key terms ofArtificial Intelligence applied to industry and energy: Deep Learning, PINNs, XAI, LLM, RAG, Digital Twin, Edge AI or Predictive Maintenance. Each entry in this AI glossary includes a concise definition and a real plant example. If you want to go beyond this AI glossary, you can also check theofficial definition of Artificial Intelligence on Wikipedia or theIEEE standards portal.
A finite, ordered set of instructions that a system follows to solve a problem or produce a result from input data.
A type of machine learning in which the model is trained with labelled examples, i.e. known input-output pairs.
A type of machine learning in which the model discovers patterns in the data without pre-existing labels. Useful for grouping, segmenting or detecting anomalies.
An approach in which an agent learns to make decisions through trial and error: it receives rewards for good actions and penalties for bad ones.
The convergence of industrial IoT and artificial intelligence: sensors, PLCs and gateways run AI models to make real-time decisions without always relying on the cloud.
A neural network that learns to compress and reconstruct its own input data. It is mainly used to detect anomalies: if the reconstruction departs from the original, something is off in the process.
A set of AI techniques that identify data, events or behaviours that deviate significantly from a process's normal pattern. It underpins asset monitoring and industrial cybersecurity.
A system that uses AI agents to carry out tasks autonomously on the user's behalf, capable of deciding, planning and using tools to achieve complex goals.
A machine learning technique in which the model itself selects the most informative data for a human to label, reducing annotation cost and speeding up training.
A technique that lets a neural network dynamically weigh which parts of the input are most relevant to each prediction, learning to "pay attention" to the key elements of a sequence. It is the core mechanism of Transformers.
A set of policies, controls and responsibilities that ensure AI systems are safe, ethical, traceable and compliant with regulations throughout their entire lifecycle.
A model that predicts the next value in a sequence from its previous values. In time series it takes the AR form, and in generative AI it describes how a language model produces each token conditioned on the ones it has already written.
A function each neuron applies to the weighted sum of its inputs to produce its output, introducing non-linearity into the network. Without it, a neural network could only represent linear transformations; ReLU, sigmoid and GELU are the most commonly used.
An algorithm that lets a neural network adjust its internal parameters by propagating the prediction error backwards.
A set of technologies and techniques for processing data volumes too large, fast or varied for traditional systems.
A probabilistic model that represents variables and their causal dependencies as a graph. It allows the probability of a future event to be calculated from what has already been observed.
A systematic error in an AI model that produces unfair or skewed results, usually inherited from unrepresentative training data. Detecting and correcting it is key to reliable decision-making.
A technique that normalises each layer's activations in a neural network using the mean and variance of the training batch, stabilising and speeding up learning.
A Transformer-based language model, introduced by Google in 2018, that interprets each word by considering both its preceding and following context. It is used as a foundation for classifying, searching or extracting information from text.
A decoding algorithm that, at each step of sequence generation, keeps the k most probable hypotheses (the so-called beam width) instead of keeping only the best one, improving the result's quality compared with greedy search.
A method for optimising expensive-to-evaluate functions that builds a probabilistic surrogate model of the objective and uses an acquisition function to choose the next point to try, balancing exploration and exploitation.
The discipline that enables machines to interpret images and video: detecting objects, reading text, identifying defects or measuring distances.
A type of neural network specialised in processing images. It uses convolutional layers that automatically detect visual patterns (edges, textures, shapes).
The branch of AI that looks for cause-and-effect relationships, not just correlations. It allows questions like "what would happen if…?" to be answered and interventions with real impact on the process to be designed.
An unsupervised learning technique that groups data into sets (clusters) based on similarity, without prior labels. It is used to segment, explore data and discover hidden profiles.
A table summarising a classifier's performance by crossing actual classes with predicted ones: true and false positives and negatives. It underpins metrics such as precision, recall or F1.
A change over time in the relationship between the input variables and what the model predicts, so that the learned logic is no longer valid. Unlike data drift, here the concept itself changes, not just the data distribution.
A measure of similarity between two vectors based on the cosine of the angle they form, ranging from -1 to 1, independent of their magnitude. It underpins comparing embeddings by meaning.
A prompting technique that asks a language model to reason step by step, showing the intermediate steps before giving the final answer. It improves performance on multi-step tasks and makes its reasoning more transparent.
A self-supervised learning method that learns representations by pulling similar examples (positive pairs) closer together in vector space and pushing dissimilar ones (negative pairs) apart, without needing labels.
A statistical framework that converts any model's point predictions into sets or intervals with guaranteed coverage, assuming only that the data are exchangeable.
A training strategy that presents examples to the model in order of increasing difficulty, from simple to complex, to speed up convergence and improve generalisation.
The amount of text, measured in tokens, that a language model can hold in mind at once. It determines how much information fits between the query and the response before the model starts to forget.
An explanation that indicates the minimal change in the input variables that would have been enough for the model to give a different output.
A statistical technique that identifies the moments when a time series' distribution changes (mean, variance or pattern), splitting it into segments with distinct properties. It is used to detect regime transitions that are not one-off faults but sustained changes.
Continuous tracking of a machine's physical parameters (vibration, temperature, current, oil) to detect changes that indicate a developing fault. It is the data foundation that predictive maintenance models work on.
A branch of machine learning based on neural networks with many layers. Capable of learning hierarchical representations of complex data without manual feature engineering.
A gradual change in the distribution of the data reaching a model in production compared with the data it was trained on. If not monitored, the model silently loses accuracy.
A model that splits data into branches according to rules like "if X exceeds a threshold, then…", forming an easy-to-interpret tree-like structure.
A virtual replica of a physical asset, process or plant, fed in real time with real-world data and enriched with AI models to simulate, predict and optimise.
A generative model that learns to create new data starting from random noise and reversing a diffusion process step by step. It underpins many of today's image generators.
A regularisation technique that randomly deactivates a percentage of neurons at each training step, forcing the network to learn redundant representations and reducing overfitting.
A data architecture that combines a data lake's flexible, low-cost storage with a data warehouse's reliability, governance and query performance, serving both BI and machine learning on a single platform.
A record of a piece of data's journey throughout its lifecycle, from its origin and transformations to its destination, enabling the information's reliability to be audited, traced and validated.
A structured, organised collection of data (whether tables, images, signals or text) that serves as the basis for training, validating and evaluating machine learning models. Its quality and size directly determine the model's performance.
A continuous flow of data connecting all of a product or asset's information throughout its lifecycle (design, manufacturing, operation and maintenance) enabling traceability and real-time decisions. It is broader than the digital twin, which replicates a specific asset.
A technique that artificially expands a training set by applying transformations to existing data (rotations, noise, crops, scaling), increasing its diversity to improve generalisation and reduce overfitting.
A digital representation of a physical asset with automatic data flow only in the physical→digital direction: it reflects the asset's real state but does not act on it, unlike the digital twin.
A set of techniques that adjust a model trained in a source domain so it works in a target domain with a different data distribution and little or no labelling.
A decentralised data architecture in which each business domain publishes and maintains its own data as a product, with federated governance and a common self-service platform.
An algorithm that measures the similarity between two time series that may run at different speeds, stretching or compressing the time axis to find the best alignment. It outperforms Euclidean distance when two curves have the same shape but are not synchronised.
A dense numerical representation of a piece of data (word, image, product, signal) in a vector space where similar elements end up close to one another.
A set of techniques that let an AI model explain its decisions in a way humans can understand. Essential in industrial and regulated environments.
Running artificial intelligence models directly on the device (PLC, industrial camera, gateway or machine) instead of sending data to the cloud. It reduces latency, network cost and dependence on connectivity.
A technique that combines several models to obtain a better, more reliable prediction than any of them individually, reducing bias and variance.
A regularisation technique that stops training when the error on the validation set stops improving, preventing overfitting.
One complete pass of the training algorithm through the entire dataset. If the dataset has 10,000 examples and the batch size is 100, one epoch equals 100 iterations; the number of epochs is a hyperparameter tuned with early stopping.
Adjusting a pre-trained model with data specific to a particular case. It allows a general model to be adapted to a specific domain without training it from scratch.
A distributed training technique in which several plants or devices train a shared model without sharing their raw data: only the model updates are exchanged.
A large-scale AI model pre-trained on massive amounts of general data (text, code, images) that is later adapted to specific tasks through fine-tuning or prompting.
A metric that summarises a classifier's performance as the harmonic mean of precision and recall. It is especially useful when classes are imbalanced.
A centralised repository that stores and manages the features used by ML models, ensuring consistency between training and inference and their reuse across projects.
The process of transforming raw data into variables (features) that better represent the problem for the model, using techniques such as normalisation, aggregation or creating new indicators. It often delivers more improvement than the algorithm itself.
An approach in which a model learns to solve a task from a very small number of labelled examples, useful when data are scarce or expensive to annotate.
A capability that lets a language model invoke external tools or functions: it detects when one is needed and generates a structured call (usually in JSON) whose result it incorporates into its reasoning. It is one of the foundations of agentic AI.
A logical system that admits intermediate degrees of truth between 0 and 1 instead of strictly binary values. Fuzzy controllers apply "if-then" rules over linguistic variables and convert the result back into a numerical signal.
A control engineering discipline that monitors a system, decides whether a fault exists and locates its type and origin. It combines model-based methods (comparing readings with what a physical model predicts) with signal- or data-based methods.
An architecture in which two neural networks compete against each other: one generates synthetic data and the other tries to tell whether it is real or fake. The result is increasingly realistic samples.
An optimisation algorithm that adjusts a model's parameters by following the "descent" of the error. It is the mathematical engine behind training in most Machine Learning models.
A family of AI models capable of creating new content (text, code, images, CAD designs or process recipes) from natural-language instructions.
An ensemble learning technique that sequentially combines weak decision trees, where each new tree corrects the previous one's errors; XGBoost is its most widespread implementation.
A set of programmable rules and filters placed between the user and a language model to ensure safe, reliable and compliant responses, blocking improper inputs or outputs.
A simplified variant of a recurrent neural network that uses two gates (update and reset) to capture temporal dependencies with fewer parameters than an LSTM, training faster and with lower memory usage.
An optimisation method inspired by biological evolution: it evolves a population of candidate solutions through selection, crossover and mutation until converging on a sufficiently good solution.
A neural network that operates on data structured as graphs (nodes and connections) propagating information between neighbouring nodes through message passing, so that the learned representations capture the system's topology.
A hyperparameter tuning method that exhaustively evaluates all combinations of a predefined grid of values, usually with cross-validation.
A non-parametric method that treats an unknown function as a probability distribution over functions, so that each prediction comes with its uncertainty.
A model configuration parameter that is NOT learned during training but decided beforehand (learning rate, number of layers, batch size).
When an LLM answers with something that sounds coherent but is factually false or made up. It is one of the main risks of using generative AI in critical environments.
A probabilistic model of a system that transitions between states which are not directly observed and are inferred from a sequence of measurable observations.
The phase in which an already trained model is used to make predictions on new data. Unlike training, inference is fast and runs in production.
The discipline that develops systems capable of performing tasks that normally require human intelligence: perception, reasoning, language, decision-making.
An anomaly-detection algorithm that builds trees with random partitions: observations that end up isolated with few splits receive a high anomaly score. It does not require labelled data.
A data structure that represents entities (equipment, products, people, events) and the relationships between them as a graph, enabling complex reasoning about the domain.
A recursive algorithm that estimates a dynamic system's state from noisy, incomplete measurements, combining a prediction phase and a correction phase to achieve more accurate estimates.
A technique that transfers knowledge from a large, accurate model (the "teacher") to a smaller, faster one (the "student"), retaining most of the performance at a fraction of the computational cost.
An AI model trained on large amounts of text that can generate, summarise, translate or answer in natural language. GPT, Claude, Llama or Mistral are well-known examples.
A recurrent neural network variant with memory cells and gates (input, forget and output) that captures long-term dependencies and mitigates the vanishing gradient problem.
An efficient fine-tuning technique that freezes the original model's weights and trains only a few small added low-rank matrices, drastically reducing the memory and cost of adapting an LLM.
A mathematical optimisation method that seeks the best value of a linear objective function subject to linear constraints. It is a classic tool for optimally allocating limited resources.
A remote-sensing technology that measures distances using laser pulses and generates three-dimensional representations of the environment.
A hyperparameter that sets how much a model's weights change at each gradient descent step. If it is too low, training takes too long; if it is too high, the model oscillates and never converges.
A prediction of the amount of electricity (power in kW or energy in kWh) that will be needed at a given time and place, from hours to years ahead. Grid operators and plants use it to plan generation, purchases and start-ups.
A function that quantifies the difference between the model's predictions and the actual values; training consists of adjusting the parameters to minimise it. Mean squared error is common in regression and cross-entropy in classification.
The branch of AI in which machines learn patterns from data, without being explicitly programmed for each case.
A set of practices and tools for deploying, monitoring and maintaining Machine Learning models in production reliably and at scale. The equivalent of DevOps for AI.
AI models capable of processing and integrating several types of data at once (text, image, audio, video or signals) to achieve a more complete understanding than a single data type would give.
A standardised document accompanying an AI model that describes its purpose, training data, metrics, limitations and intended uses, facilitating transparency and auditing.
A neural network architecture that replaces dense layers with a set of specialised sub-networks ("experts") and a routing network that activates only the most suitable ones for each input, scaling model size at much lower computational cost.
A centralised repository that manages the lifecycle of machine learning models: it versions, tags, documents lineage and controls each model's promotion between development, testing and production.
An advanced control technique that, at each instant, uses a dynamic model of the process to predict its future behaviour and solve a constrained optimisation over a moving horizon, applying only the first computed action and repeating the calculation at the next step.
A numerical method that estimates results and probabilities by running thousands of simulations with input variables randomly sampled according to their distribution.
An approach in which a single model learns several related tasks at once, sharing internal representations that act as regularisation and improve each task's performance.
A computational model inspired by how the brain works, made up of layers of artificial neurons that transform inputs into outputs.
The branch of AI concerned with processing and understanding human language: comprehension, generation, translation, text classification.
A natural language processing task that locates entities such as people, organisations, places, dates or quantities in a text and classifies them by category.
A phenomenon in which a model memorises the training data instead of learning generalisable patterns. It works very well on the known and very poorly on the new.
A technology that converts text contained in scanned images or PDFs into editable, software-processable digital text, the basis of industrial document automation.
A computer vision task that locates the objects present in an image and classifies each one, drawing a box around its position. It combines localisation and classification in a single pass.
A technique that converts a categorical variable into a numerical vector of N elements, where only the position corresponding to the category is 1 and the rest are 0, allowing models to process non-numerical data.
The apparent motion pattern of pixels between consecutive video frames, represented as a field of displacement vectors.
A property of a system indicating whether its internal state can be deduced from the measurements available over a finite time interval. If a system is not observable, no estimator will be able to reconstruct those variables.
A vendor-independent industrial communication standard (IEC 62541) that carries plant data along with its semantic description and built-in security.
Neural networks that incorporate known physical equations of the process during training. They combine real data with physical laws for more reliable, defensible predictions.
An instruction or question given to a generative AI model (typically an LLM) to obtain a specific response. The prompt's quality directly affects the response's quality.
A maintenance strategy based on AI models that analyse sensor data (vibration, temperature, consumption) to anticipate faults before they occur and plan the intervention.
Two complementary classification metrics: precision measures what proportion of the model's alerts are correct, and recall measures what proportion of actual cases it manages to detect.
A metric that evaluates a language model by measuring how surprised it is when predicting a text; the lower the perplexity, the better the model anticipates the next word.
A statistical dimensionality-reduction technique that transforms correlated variables into a smaller set of principal components, ordered by the variance they explain, retaining as much information as possible.
A population-based optimisation algorithm in which a set of particles moves through the search space, adjusting its trajectory based on its own best position and the swarm's best position.
A discipline that integrates condition monitoring, fault diagnosis and remaining-life prognosis to decide when to intervene on an asset.
A set of points with 3D coordinates, often with colour or intensity, representing the surface of objects or environments captured with LiDAR, photogrammetry or 3D scanners.
Determining an object's position and orientation, or a person's joint positions, from camera images.
A curve showing a variable's average effect on a model's prediction, averaging over the rest of the variables. It is used to read an opaque model without opening it.
Analytics that, besides forecasting what will happen, propose the action to take by combining predictive models with optimisation and the process's real constraints.
A set of techniques that reconstruct, verify and improve a process's real flow from event logs (case identifier, activity and timestamp) extracted from systems such as ERP or MES. It shows how the process really happens, not how it is documented.
An optimisation technique that reduces a model's numerical precision (for example from 32 to 8 bits) so it takes up less memory and runs inference faster, with minimal loss of accuracy.
A technique that combines an LLM with your own knowledge base. Before answering, the model retrieves relevant information from your documents and uses it to generate the response.
Short form of Neural Network. A computational model made up of layers of interconnected processing units that learn patterns from data.
A machine learning algorithm that combines many decision trees trained on random subsets of the data. It tends to give reliable predictions that resist overfitting.
A deep neural network designed for sequential data or time series that retains a "memory" of previous inputs to influence the current output.
A simplified mathematical representation of a complex physical system that reduces computational cost by orders of magnitude while preserving its dominant behaviour.
A training technique in which human evaluators' preferences are used as a reward signal to align a language model's responses with what is considered helpful and safe.
A metric that summarises in a single number a classifier's ability to distinguish between classes across all thresholds. It ranges from 0 to 1 and remains reliable even with imbalanced classes.
A set of techniques, such as L1, L2, dropout or early stopping, that penalise a model's complexity during training to prevent overfitting and improve its generalisation to new data.
A second retrieval stage in which a more accurate model reorders the initially retrieved documents by relevance, reducing noise before passing them to an LLM. It is a key piece of RAG systems.
An estimate of the time (in operating hours, cycles or equivalent units) that an asset can keep running within specification before reaching a failure threshold. It is the central output of asset prognostics models.
A information-filtering system that ranks items according to what it predicts will interest a user, based on their previous behaviour or that of similar users.
An explainable AI (XAI) technique that assigns each variable a value indicating how much it contributed to a specific model prediction, making its decisions auditable.
Artificial data generated by simulation or AI that mimic real data. They allow models to be trained when real data are scarce, costly or sensitive.
A machine learning model that approximates a computationally expensive physical simulation, making it possible to evaluate the system's behaviour at new design points much faster.
A computer vision task that classifies each pixel in an image into a predefined category, generating a mask that delimits each class's regions.
A technique in which the model generates its own labels from unannotated data, learning to predict one part of the input from another. It drastically reduces the need for manual labelling.
A search technique that interprets the query's intent and meaning, not just the exact words, to return relevant results even if they are phrased differently.
A supervised learning algorithm that classifies data by finding the optimal hyperplane that maximises the margin between classes; with kernel functions it also solves non-linear problems.
A software model that estimates in real time a variable that is difficult or costly to measure, based on other easy-to-measure process signals. It is also called an inferential sensor.
A technique that combines data from several sensors to obtain a more accurate estimate with less uncertainty than any of them would offer individually.
A process that infers a system's unmeasured internal variables from noisy or incomplete measurements and a model of the process.
An approach that trains models, usually for control or reinforcement learning, in a simulator and transfers them to physical equipment, closing the gap with techniques such as domain randomisation.
A technique that balances imbalanced datasets by generating synthetic examples of the minority class through interpolation between nearby neighbours, instead of duplicating existing ones.
A family of statistical methods that model the time until an event, such as a failure, handling censored data: assets that have not yet failed by the study's cut-off date.
A technology that converts speech into written text. It is also called ASR or speech-to-text.
A supervisory control and data acquisition system that centralises field signals, displays the process's state and allows action to be taken on it. It is the usual source of the historical data used to train industrial models.
A representation of how a signal's energy is distributed across frequencies over time. It converts a time signal into an image that vision networks can classify.
An unsupervised neural network, proposed by Kohonen, that projects data with many variables onto a two-dimensional grid while preserving neighbourhood relationships: similar process states end up close together.
A technique that reuses a model trained in one domain (with lots of data) to solve a similar problem in another domain (with little data). It saves months of training.
A neural network architecture that appeared in 2017 and transformed AI, especially in language. It is the foundation of modern LLMs: GPT, Claude, BERT, Llama, Mistral.
A set of statistical and deep learning techniques (ARIMA, Prophet, LSTM, Temporal Fusion Transformer) used to anticipate future values of variables that evolve over time.
A preliminary natural language processing step that splits a text into minimal units (tokens), words or fragments, so a model can process it.
A non-linear dimensionality-reduction technique that projects high-dimensional data into two or three dimensions while preserving local neighbourhood relationships, useful for visualising clusters.
A parameter that adjusts how sharply or flatly a generative model's probability distribution is shaped when choosing the next token. Low values give more deterministic outputs and high values, more varied ones.
A family of unsupervised methods that discover the latent topics of a document collection by grouping words that tend to appear together. LDA is the most widespread algorithm.
The opposite of overfitting: the model is too simple and fails to capture the data's real patterns. It predicts poorly both in training and in production.
A set of techniques that estimate the confidence level of a model's predictions, delivering a distribution of results instead of a single value. It is key to reliable risk management in critical environments.
A dimensionality-reduction technique that projects data with many variables into 2 or 3 dimensions while preserving local structure and much of the global structure, faster than t-SNE.
A technique for assessing a model's stability by training it several times with different data partitions, ensuring its performance does not depend on the luck of the split.
A database designed to store and index embeddings (numerical vectors) and retrieve them by semantic similarity, not just exact matching. It is the key piece of RAG systems.
An architecture that applies the Transformer mechanism to images: it divides the image into patches, converts them into a sequence and uses self-attention to capture global relationships. It rivals CNNs in image classification.
A generative autoencoder that learns a probabilistic latent distribution, which makes it possible to generate new data and measure how unusual a sample is.
Numerical values that connect a neural network's neurons and are adjusted during training. They are the parameters the model "learns".
A signal decomposition into components localised in both time and frequency, useful when brief events that the Fourier transform would dilute are of interest.
Acronym for Explainable AI. A set of techniques that make a model justify its decisions in an understandable way.
See the full definition underE · Explainable AI.
A family of real-time computer vision models capable of detecting and classifying multiple objects in an image in a single pass, optimised for very low latency.
A model's ability to solve tasks or recognise classes it never saw during training, relying on general knowledge or auxiliary descriptions.
No results
No terms match your search. Try a different keyword.