Glossary¶
This glossary serves to define some of the technical terms used within Falkonry Patterns.
Entity¶
An Entity is any single thing from which time series data is continually collected from one or more sensors. Examples: transformer, pump, motor, welding machine, robot, etc. An entity need not strictly correspond to a single physical component or system. For example: A "pump" could be defined as the pressures and currents coming from the pump as well as the output temperature and flow rate of an adjacent coolant tank.
When Falkonry is used to train a model using data from multiple entities, all of the entities must be of the same type with similar behavior. They must have the same signals (names) and the same number of signals. For example: a set of transformers, a group of pumps, a series of motors, a collection of welding machines, etc.
Conditions¶
Conditions are what we are trying to extract from patterns found in time-series data. This is typically a measure of the state of some Entity. Conditions could be used to represent health, operating mode, risk or threat level, quality level, or almost any form of condition assessment.
Falkonry can also recognize two other conditions - Unknown and Gap:
- Unknown While a pipeline is monitoring the inflow or merely running a test revision, any new patterns that arise which were not seen when the pipeline was learning will be reported as the unknown condition. An unknown condition is displayed on the timeline alongside other condition labels. If a fact is added about that pattern, then Falkonry learns from it and replaces the unknown condition with the provided fact.
- Signal-Gap Ordinarily unevenly sampled data is automatically handled in Falkonry. However, if there is a long period of time during which any of the input signals have no data samples, then the assessment for such a period is reported as a gap condition and it is displayed on the UI with other assessment condition labels after learning.
Condition Model¶
A mathematical construct that converts signal data into assessment output. For example: a model converts a set of temperature and pressure values into a set of condition values (normal, fault, etc), explanation scores and confidence values. Falkonry condition models typically take data from a time window leading up to a point in time (assessment point) where it produces assessment output.
Sliding Window Model (S-Window Model)¶
A model in which assessments are separated by time rather than by event or signal. For example: a model which performs an assessment every 5 minutes would be a sliding window model.
Typically, if an asset is continuously operated & monitored and the condition could change continuously, it is generally suitable for the Sliding Window model approach.
Explicit Window Model (X-Window Model)¶
A model in which assessments are separated by event or signal value rather than by time. For example: a model which performs an assessment after each weld is completed would be an explicit window model.
Examples include: chemical batch processing, welding, cutting of metal billets/ingots, fault traces of electrical equipment, wire bending, a single cycle of oil pump, semiconductor wafer processes, etc.
Activities that are discrete in nature and have well defined start / end markers along with fixed and repetitive windows / periods of operation are best characterized by Explicit windows.
X-Window models require an explicit window signal (not part of the model) that defines boundaries of the windows. In the explicit window signal a value change indicates a new explicit window. This could be a categorical signal or a numerical signal.
Unsupervised model¶
A model which produces assessments that express similarity between training data and target data but which does not use human classification of training data to perform this grouping. That is, the "shape" of the training data alone defines the groups into which target data will be classified.
Semi-supervised model¶
A model which produces assessments that express similarity between training data and target data in which some human (or external system) classification of training data is used to define groups. That is, a limited number of classified examples are used to define groups into which target data will be classified. If target data does not fit into a defined group, target data may still be clustered into unsupervised groups.
Supervised model¶
A model which produces assessments that express similarity between training data and target data in which all training data is classified by a human (or other system) in order to define groups. That is, all training examples are classified and are used to define groups into which target data will be classified. If target data does not fit into a defined group, target data may still be clustered into unsupervised groups.
Datastream¶
Datastream is a data organizing construct within Falkonry. It consists of the loaded data (sources), signals, signal groups, entities, assessments, models, events, event groups and model output.
Assessment¶
A subset of a datastream which is typically used to segregate events and models in a way that allows the user to address a particular question using that data in a way that does not interfere with other questions the user wants to ask of the data.
An assessment shares signal data with all other assessments in a datastream. An assessment may have a different explicit window definition from other assessments in the same datastream. An assessment has unique events, event groups and models from all other assessments.
Apply¶
The process of using a model on a designated set of target signal data in order to produce a set of assessment output for that data. This is the inference part of machine learning.
Explanation Scores¶
Float valued numbers between -1.0 and +1.0 which indicate the relative contribution of each signal to a particular assessment output.
A higher score (closer to +1.0) suggests a higher correlation of the particular signal with the assessment outcome associated with the assessment point currently selected. In other words, the signal had a higher impact on determining the classification of the episode at the current assessment point.
A lower score (closer to 0.0) suggests a lower correlation of the particular signal with the assessment outcome associated with the assessment point currently selected.
A high negative score (closer to -1.0) suggests a high correlation of the particular signal with a classification other than the classification associated with the assessment point currently selected.
Live Monitoring¶
Once the model training is complete, deploying the model to accept real time data as input and to produce assessment output for that data is Live Monitoring.
Event¶
An interval of time with a defined start and end, for which a condition value is known or asserted. When used in a model as input, it is known as a "fact." Events are used to confirm assessment performance via evaluation. They also provide labels for supervised and semi-supervised learning.
Event group¶
A label which relates multiple events to each other. For example: A "pump failures" event group label could be applied to all the specific event instances where the pump failed. A single event may belong to multiple event groups.
Signal Group¶
A label which relates multiple signals to each other. For example: A "pump 1 mechanical" signal group label could be applied to all of the signals which describe the mechanical operation of pump 1 (e.g. RPM, inlet pressure, outlet pressure, etc).
Evaluation¶
The process of applying a model to a set of target data and then calculating a number of properties of the assessment output for that same target data. The properties include:
- Condition distribution
- Recall
- Precision
- F-Score
Running an evaluation is required to apply a model to a new set of target data. Creating an episode will also cause an evaluation to occur on the time period that the episode spans. However, performing an evaluation will not automatically create an episode.
Generalization Factor¶
A float value between 0.0 and 1.0 which is used to set a threshold that the model's confidence value needs to exceed in order to assign a condition to that assessment. This occurs in two stages for Falkonry models. The first stage applies to supervised/semi-supervised model output. If the supervised assessment's confidence value exceeds (1-GF) then the supervised model's condition assessment (i.e. a classification) will be applied. If not, then the second stage is checked.
The second stage applies to the unsupervised model output. If the unsupervised assessment's confidence value exceeds (1-GF) then the unsupervised model's condition assessment (i.e. an unsupervised group label) will be applied. If not, then the assessment will be labeled as "unknown." With Generalization value \< 0.5, Falkonry matches the supplied fact patterns increasingly restrictively. With values closer to 0.0 the patterns that are slightly different from learned patterns are classified as "unknown"
With Generalization value >= 0.5, Falkonry matches the fact patterns increasingly approximately (generally). With value of 1.0 every pattern is classified into one of the supplied fact classifications Unknown
When Falkonry discovers a pattern in the target data being assessed that is not sufficiently similar to examples in the trained data set, an \'unknown\' label is assigned. See "generalization factor" for more details of how "sufficiently similar" is defined.
\'unknown\' patterns could be an anomaly or a new condition. For Falkonry to recognize and label unknowns properly in the future, the data associated with an \'unknown\' pattern should be included in the training set of the next iteration of the model.
'U1' through 'U10'¶
Falkonry machine generated assessment output condition labels (unlabeled1,...,unlabeled10) for patterns discovered during unsupervised learning.
SVD¶
Signal Value Distribution - is a histogram of the signal values within an evaluation time period for a chosen condition for a chosen model. For example: distribution of "pressure" values for M8's "cavitation" condition. SVDs allow comparison of the overall behavior of a signal for a condition within a time window with other conditions that existed in the training set for that model.
SFD¶
Signal Frequency Distribution - also known as Power Spectral Density. Signal Frequency Distribution is the measure of signal\'s power content versus frequency
Episode¶
A time window used by Falkonry Patterns to denote a period during which some behavior of interest is likely to have occurred. There are many kinds of episodes (See other X episode definitions). The data within an episode will have an evaluation performed on it as part of the episode creation process.
NOVEL Episode¶
An episode which encompasses a time period which has been classified by a model as an "unknown" condition. Novel episodes typically indicate that something has occurred within the system that has not been seen before.
ALERT Episode¶
An episode which encompasses a time period which has been classified by a model as a condition which the user has previously specified as being "important" so that they want to be notified by Falkonry when the condition occurs again.
User Episode¶
An episode which encompasses a time period defined by a user. The specific conditions contained within the episode are irrelevant to the episode being called a "user episode."
Digital Twin¶
Any algorithmic, heuristic or mathematical model of a system which can use signal data from a real-world instance of the system in order to better understand the past, current or potential future behaviors of that system. A Falkonry digital twin is a condition model which describes or predicts the behavior of an entity.
Dashboard¶
A UI element in Falkonry AI Cloud which displays a variety of information including signal charts and Falkonry output that can be used for continuous monitoring purposes.
F-Measure¶
The harmonic mean of recall and precision. A higher F-measure indicates that both precision and recall were higher. In many applications it is necessary to trade off recall for precision (or vice versa). It is important to understand the application in order to decide whether precision, recall or a balance of both is required for the use case to be successfully resolved. That is, a higher F-score isn't always better.
Precision¶
Fraction of what was predicted that was classified correctly (false positives do count against precision). Higher precision → higher probability that when you say something is an X it really is an X. High precision is important in applications which cannot tolerate false positives.
Recall¶
Fraction of what is actually a certain class that you classified correctly (false positives don\'t count against recall).
Higher recall → higher probability that you classified as X all of the things that are actually X, even if many of the things you called X are incorrectly classified. High recall is important in applications which cannot tolerate false negatives.
Signal Gap¶
This is a Falkonry system generated classification when a signal included in a model does not have data available for longer than the signal gap threshold. Note: Time periods shorter than the gap threshold will have signal data values filled in by interpolation or held at the last known value for that signal by Falkonry.
Also a condition in the assessment output of a model which is applied when a signal gap (per the 1st definition) occurs.
Note
Signal gap threshold may be set by API or via Falkonry Patterns UI
EHE (Event Horizon Estimation)¶
An algorithmic approach which calculates the time remaining until some target condition occurs. For example: if the target condition is a pump seal failure, then the EHE returns a time from now until the pump seal failure occurs.
When applied to part or system failure, this is commonly referred to as TTF (Time to Failure) or RUL (Remaining Useful Life). Because the target condition can be any condition of interest, not just part failure, Falkonry refers to the capability as Event Horizon Estimation.
Connection¶
A user named set of processes which is used to load signal data into Sources. Files placed into the directory specified for each Connection, or data provided via MQTT connection will be loaded and tagged in a way which links that Source to the Connection used to load it via a connectionID.
Source / Connected Source¶
A Falkonry representation of data which can be used to create a signal. A source must be mapped in order to assign that source's data to a specific signal name belonging to a specific entity in a specific datastream.
Assessment output (conditions) are stored as sources so that they can be used as signals in another model. This is necessary to create a model of models.
Edge Analyzer¶
A Docker container instance of the services and configuration data (including a model) required to accept signal data and to apply a Falkonry model to that data. This is typically used for live monitoring in real time. Each instance is complete so that no connection to web services is required to use it. Edge analyzers are intended to be light-weight so that they can be easily deployed in low compute environments within the network close to the endpoint devices generating the signal data (aka the "edges" of the network).
Edge analyzers do not contain the ability to train a new model so their models are static unless updated by a Falkonry Patterns.
MQTT (Message Queue Telemetry Transport)¶
Publish-subscribe-based messaging protocol, standardized by the OASIS technical committee, which is used to communicate device data. It is often used when a very large network of small devices needs to be monitored or managed via the Internet e.g. parking sensors, underwater lines, energy grid, etc. The messages are encapsulated in several predefined control packets, which are designed to minimize the network footprint. Widely adopted standard. Good for server-to-server and server-to-client communication.
Model of Models¶
A model which takes as signal input the assessment output of another model. Also known as a "chained model."