PluginMLIntegration

The PluginMLIntegration provides a standardized interface for connecting machine learning models (e.g., Tensorflow.js, ONNX) to the Velo Plot. It handles data extraction, asynchronous inference, and high-performance visualization of predictions and confidence intervals.

Core API

registerModel(model)

Register a custom model implementation that satisfies the MLModelAPI interface.

chart.ml.registerModel({
  id: 'my-nn-forecaster',
  name: 'Forecasting Model',
  type: 'forecasting',
  async predict(data) {
    // data.x and data.y are plain arrays extracted from series
    const prediction = await myLoadedModel.predict(tf.tensor(data.y));
    return {
      x: futureXArray,
      y: predictionArray,
      confidence: confidenceIntervalArray
    };
  }
});

runInference(modelId, seriesId)

Runs analysis on a specific data series. It returns the PredictionResult directly.

visualizeResults(result, config)

Renders the result on the chart overlay. This is extremely efficient as it avoids creating new heavy-weight series for transient predictions.

• showConfidenceInterval: Renders a translucent band around the prediction.
• intervalOpacity: Control the transparency of the confidence band.
• lineStyle: Customize the appearance of the prediction curve.

visualizePredictions(result, config)

Intent-revealing alias for visualizeResults, for the prediction-overlay use case.

trainModel(modelId, { x, y })

Trains a small native regression model on the fly. Creates a native linear-regression model if modelId does not exist yet. Returns fit diagnostics for a residual plot:

const fit = chart.ml.trainModel('trend', {
  x: [[0], [1], [2], [3], [4]],   // feature rows
  y: [2.1, 3.9, 6.2, 7.8, 10.1],  // targets
})
// fit -> { coefficients, intercept, fitted, residuals, r2, rmse }
chart.addSeries({ id: 'residuals', type: 'scatter', data: { x: fit.fitted, y: fit.residuals } })

listModels()

Returns the descriptors of all registered models.

Model audit (supported native models)

velo-plot ships native, dependency-free models. External frameworks (TensorFlow.js, ONNX) can be bridged via registerModel.

Native model type	Capability	Limits
linear-regression	OLS fit via normal equations (general N×N inverse)	Linear relationships only; no regularisation
neural-network	Feed-forward inference (relu/sigmoid/tanh)	Inference only — no native backprop training
signal-processor	Low/high/band-pass filtering	First-order (EMA-based) filters

Statistics helpers (chart.ml.stats): fft (naive O(n²) DFT — fine for small windows), mean, standardDeviation, correlation.

Honest scope note: the native NN performs inference only. For training deep models, load weights from an external framework and bridge via registerModel. Native trainModel covers linear/multivariate regression.

Scientific Application

Specifically designed for:

• Real-time Signal Denoising: Using autoencoders to predict clean signals.
• Anomaly Detection: Visualizing probability scores across a time series.
• Electrochemical Forecasting: Predicting peak positions in future CV cycles.
• Trend Extrapolation: Using LSTMs to forecast multi-variable trends.

Known limitations

• Native simple NN/regression only — not a full ML framework integration (accepted v3 gap).