LSTMDetector<T>
Models & Types · AiDotNet.AnomalyDetection.TimeSeries
Implements LSTM-based anomaly detection using prediction error.
Inheritance: ModelBase<T, Matrix<T>, Vector<T>> → AnomalyDetectorBase<T> → LSTMDetector<T>
Implements: IGeneratedParameterRegistrar<T>
For Beginners
LSTM (Long Short-Term Memory) learns patterns in sequential data and predicts the next value. Anomalies are detected when the actual value differs significantly from the predicted value.
How It Works
The algorithm follows the prediction-based formulation of Malhotra et al. (2015) and Hundman et al. (2018, NASA telemanom):
- Train an LSTM to FORECAST the next value of the series.
- For each point, compute the forecast error.
- High forecast errors indicate anomalies.
Because the core of the detector IS a forecasting LSTM, the model also exposes a normal
forecasting surface: Train learns a target series and
Predict returns one-step-ahead forecasts. The recurrent core runs on the
shared tensor Engine (tape-based BPTT) rather than a hand-rolled scalar loop.
Reference: Malhotra et al. (2015), “Long Short Term Memory Networks for Anomaly Detection in Time Series”; Hundman et al. (2018), “Detecting Spacecraft Anomalies Using LSTMs and Nonparametric Dynamic Thresholding”; Hochreiter & Schmidhuber (1997), “Long Short-Term Memory”.
Constructors
| Constructor | Summary |
|---|---|
LSTMDetector(int hiddenDim, int seqLength, int epochs, double learningRate, double contamination, int randomSeed) |
Creates a new LSTM anomaly detector. |
Properties
| Property | Summary |
|---|---|
int HiddenDim { get; } |
Gets the hidden dimension of the LSTM. |
int SeqLength { get; } |
Gets the sequence/window length. |
Methods
| Method | Summary |
|---|---|
Vector<T> Forecast(int steps) |
Returns one-step-ahead forecasts of the learned series in the ORIGINAL (denormalized) scale — the model’s underlying prediction surface (Malhotra et al. |
Vector<T> ScoreAnomalies(Matrix<T> X) |
|
void Fit(Matrix<T> X) |