SignalSharp is a .NET library focused on time series analysis, change point detection, and signal smoothing/filtering tasks. It provides implementations of common algorithms optimized for clarity and efficiency.
Algorithms to identify points in time where the statistical properties of a signal change.
-
PELT (Pruned Exact Linear Time):
- Detects multiple change points efficiently.
- Supports various cost functions to target specific types of changes, including mean, variance, event rate, and distribution shape.
- Includes a
PELTPenaltySelectorfor automatic penalty selection using BIC, AIC, or AICc for supported cost functions.
-
CUSUM (Cumulative Sum):
- Detects shifts in the mean of a signal.
- Works by accumulating deviations from an expected level and triggering when a threshold is exceeded.
Methods to reduce noise from measurements.
- Savitzky-Golay Filter: Smooths data by fitting successive sub-sets of adjacent data points with a low-degree polynomial. This helps preserve signal features better than a simple moving average.
- Moving Average: A basic smoothing technique that calculates the average of data points within a sliding window.
Methods for forecasting future values based on historical data.
- Simple Exponential Smoothing (SES): A forecasting method for univariate time series without a clear trend or seasonal pattern. It computes future values based on a weighted average of past observations, with recent observations given more weight. The forecast is a constant value, which is the last smoothed level calculated from the data.
- Holt's Linear Trend Method: Also known as Double Exponential Smoothing, this method is for data with a trend but no seasonality. It supports both additive and multiplicative trends and can include a "damped" trend to improve long-term forecast accuracy. The implementation can also automatically find optimal smoothing parameters via a built-in grid search.
- Linear Extrapolation: Forecasts future values by fitting a simple linear trend to a recent window of data points and extending that line into the future. The size of the fitting window is configurable.
Hyperparameter tuning methods to find the best-performing parameters for a model.
- Grid Search: An exhaustive search method that systematically evaluates a grid of parameter combinations to find the one that yields the best performance. It supports parallel processing and an "adaptive refinement" feature that performs a second, more focused search around the best initial result.
- Nelder-Mead: A direct search method that "crawls" toward the minimum of a function in a multi-dimensional space without requiring derivative information. This makes it well-suited for objective functions that are noisy or non-differentiable. It can be configured to perform multiple restarts from different starting points to increase the likelihood of finding a global minimum.
- Signal Padding: Methods for common padding modes (
Constant,Mirror,Nearest,Periodic). - Statistical Functions: Basic statistics (
Mean,Variance,StdDev,Median,Normalization, etc.).
Install the package via the dotnet CLI:
dotnet add package SignalSharpOr use the NuGet Package Manager in your IDE.
For detailed examples and API documentation, please refer to the official documentation.
Here's a quick example of how to use the PELT algorithm for change point detection:
using SignalSharp.Detection.PELT;
using SignalSharp.CostFunctions.Cost;
// Create a sample signal
double[] signal = { 1, 1, 1, 5, 5, 5, 1, 1, 1 };
// Initialize PELT algorithm (using default L2 cost, MinSize=1, Jump=1)
var options = new PELTOptions();
var algo = new PELTAlgorithm(options);
// Detect change points with a manually chosen penalty
int[] breakpoints = algo.FitAndDetect(signal, penalty: 2.0);
Console.WriteLine($"Manual Penalty Change Points: {string.Join(", ", breakpoints)}"); // Expected: [3, 6]
// --- Or use Automatic Penalty Selection (requires a likelihood cost function) ---
// Configure PELT with Gaussian Likelihood (supports BIC/AIC)
var likelihoodOptions = new PELTOptions { CostFunction = new GaussianLikelihoodCostFunction(), MinSize = 2 };
var likelihoodAlgo = new PELTAlgorithm(likelihoodOptions);
// Create the selector
var selector = new PELTPenaltySelector(likelihoodAlgo);
// Choose selection method (e.g., BIC)
var selectionOptions = new PELTPenaltySelectionOptions(PELTPenaltySelectionMethod.BIC);
// Fit data and select penalty
var result = selector.FitAndSelect(signal, selectionOptions);
Console.WriteLine($"Auto Penalty (BIC): {result.SelectedPenalty:F4}");
Console.WriteLine($"Auto Penalty Change Points: {string.Join(", ", result.OptimalBreakpoints)}");
// Expected: likely [3, 6], penalty will varyContributions are welcome! If you have ideas, suggestions, or bug reports, feel free to open an issue or submit a pull request.