oscura 0.5.1__py3-none-any.whl → 0.7.0__py3-none-any.whl

oscura/analyzers/classification.py

@@ -0,0 +1,659 @@
+"""Intelligent signal classification pipeline for automatic signal type identification.
+
+This module provides multi-method signal classification to automatically identify
+signal types from waveforms without manual configuration.
+
+Key capabilities:
+- Classify signals: digital, analog, PWM, UART, SPI, I2C, CAN
+- Multi-method classification: statistical, frequency domain, pattern recognition
+- Confidence scoring for each classification
+- Batch classification with parallel processing support
+- Extensible architecture for adding new classifiers
+
+Classification methods:
+- Statistical features: mean, variance, duty cycle, edge density
+- Frequency domain: FFT analysis, dominant frequencies, spectral characteristics
+- Time domain patterns: Protocol-specific signature detection
+- Rule-based classification: Feature threshold matching
+
+Typical workflow:
+1. Extract features from signal (statistical, frequency, pattern)
+2. Apply classification rules based on feature values
+3. Return best match with confidence score and alternatives
+
+Example:
+    >>> from oscura.analyzers.classification import SignalClassifier
+    >>> classifier = SignalClassifier()
+    >>> result = classifier.classify(signal, sample_rate=1e6)
+    >>> print(f"{result.signal_type}: {result.confidence:.2f}")
+    uart: 0.94
+    >>> print(f"Features: {result.features}")
+    Features: {'duty_cycle': 0.52, 'edge_density': 0.042, ...}
+
+References:
+    IEEE 181-2011: Transitional Waveform Definitions
+    DISC-001: Automatic Signal Characterization
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any, ClassVar, Literal
+
+import numpy as np
+
+if TYPE_CHECKING:
+    from numpy.typing import NDArray
+
+SignalType = Literal["digital", "analog", "pwm", "uart", "spi", "i2c", "can", "unknown"]
+
+
+@dataclass
+class ClassificationResult:
+    """Result of signal classification.
+
+    Contains detected signal type, confidence score, features used for classification,
+    and alternative matches for ambiguous cases.
+
+    Attributes:
+        signal_type: Detected signal type (digital, analog, uart, spi, etc.)
+        confidence: Confidence score (0.0-1.0), higher is more confident
+        features: Dictionary of extracted features used for classification
+        secondary_matches: Alternative classifications with confidence scores
+        reasoning: Human-readable explanation of classification decision
+
+    Example:
+        >>> result = classifier.classify(signal, sample_rate=1e6)
+        >>> if result.confidence >= 0.8:
+        ...     print(f"High confidence: {result.signal_type}")
+        >>> for alt_type, alt_conf in result.secondary_matches:
+        ...     print(f"Alternative: {alt_type} ({alt_conf:.2f})")
+    """
+
+    signal_type: SignalType
+    confidence: float
+    features: dict[str, float]
+    secondary_matches: list[tuple[SignalType, float]] = field(default_factory=list)
+    reasoning: str = ""
+
+
+@dataclass
+class ClassifierRule:
+    """Rule-based classification criteria.
+
+    Defines feature thresholds for identifying a specific signal type.
+
+    Attributes:
+        signal_type: Signal type this rule identifies
+        conditions: Dict mapping feature name to (min, max) thresholds
+        weight: Importance weight for this rule (default 1.0)
+        required_features: Features that must be present and within range
+
+    Example:
+        >>> rule = ClassifierRule(
+        ...     signal_type="digital",
+        ...     conditions={"variance": (0.2, 1.0), "edge_count": (100, float('inf'))},
+        ...     weight=1.0
+        ... )
+    """
+
+    signal_type: SignalType
+    conditions: dict[str, tuple[float, float]]
+    weight: float = 1.0
+
+
+class SignalClassifier:
+    """Multi-method signal classifier with extensible architecture.
+
+    Combines statistical analysis, frequency domain features, and pattern
+    recognition to automatically identify signal types from waveforms.
+
+    Classification Methods:
+        - statistical: Mean, variance, duty cycle, edge statistics
+        - frequency: FFT-based frequency analysis, spectral characteristics
+        - pattern: Protocol-specific pattern detection (UART, SPI, etc.)
+
+    Attributes:
+        methods: List of classification methods to use
+        threshold: Minimum confidence to report classification (default 0.5)
+
+    Example:
+        >>> # Default classifier uses all methods
+        >>> classifier = SignalClassifier()
+        >>> result = classifier.classify(signal, sample_rate=1e6)
+        >>>
+        >>> # Custom classifier with specific methods
+        >>> classifier = SignalClassifier(methods=["statistical", "frequency"])
+        >>> result = classifier.classify(signal, sample_rate=1e6, threshold=0.7)
+    """
+
+    # Classification rules for different signal types
+    RULES: ClassVar[list[ClassifierRule]] = [
+        # Digital signal: bimodal distribution, many edges
+        ClassifierRule(
+            "digital",
+            {"variance": (0.15, 1.0), "edge_density": (0.01, 1.0)},
+            weight=1.0,
+        ),
+        # Analog signal: continuous values, low edge count
+        ClassifierRule(
+            "analog",
+            {"variance": (0.01, 0.5), "edge_density": (0.0, 0.02)},
+            weight=1.0,
+        ),
+        # PWM signal: regular duty cycle, periodic
+        ClassifierRule(
+            "pwm",
+            {"duty_cycle": (0.1, 0.9), "periodicity": (0.6, 1.0), "edge_density": (0.01, 0.5)},
+            weight=1.2,
+        ),
+        # UART: specific bit timing patterns, moderate edge density
+        ClassifierRule(
+            "uart",
+            {"uart_score": (0.6, 1.0), "edge_density": (0.01, 0.2)},
+            weight=1.3,
+        ),
+        # SPI: clock + data patterns, high edge density
+        ClassifierRule(
+            "spi",
+            {"spi_score": (0.6, 1.0), "edge_density": (0.1, 1.0)},
+            weight=1.2,
+        ),
+        # I2C: ACK patterns, clock stretching
+        ClassifierRule(
+            "i2c",
+            {"i2c_score": (0.5, 1.0)},
+            weight=1.1,
+        ),
+        # CAN: specific encoding
+        ClassifierRule(
+            "can",
+            {"can_score": (0.6, 1.0)},
+            weight=1.1,
+        ),
+    ]
+
+    def __init__(self, methods: list[str] | None = None) -> None:
+        """Initialize signal classifier.
+
+        Args:
+            methods: Classification methods to use. Default uses all methods.
+                Available: "statistical", "frequency", "pattern"
+
+        Raises:
+            ValueError: If unknown method is specified
+        """
+        available_methods = {"statistical", "frequency", "pattern"}
+        if methods is None:
+            self.methods = ["statistical", "frequency", "pattern"]
+        else:
+            # Validate methods
+            unknown = set(methods) - available_methods
+            if unknown:
+                raise ValueError(f"Unknown methods: {unknown}. Available: {available_methods}")
+            self.methods = methods
+
+    def classify(
+        self,
+        signal: NDArray[np.floating[Any]],
+        sample_rate: float,
+        threshold: float = 0.5,
+    ) -> ClassificationResult:
+        """Classify single signal.
+
+        Extracts features using configured methods and applies classification
+        rules to determine signal type.
+
+        Args:
+            signal: Signal data array (voltage samples)
+            sample_rate: Sample rate in Hz
+            threshold: Minimum confidence for primary classification (0.0-1.0)
+
+        Returns:
+            ClassificationResult with signal type and confidence
+
+        Raises:
+            ValueError: If signal is empty or sample_rate is invalid
+
+        Example:
+            >>> result = classifier.classify(signal, sample_rate=1e6)
+            >>> print(f"Type: {result.signal_type}, Confidence: {result.confidence:.2f}")
+            Type: uart, Confidence: 0.94
+        """
+        if len(signal) == 0:
+            raise ValueError("Cannot classify empty signal")
+        if sample_rate <= 0:
+            raise ValueError(f"sample_rate must be positive, got {sample_rate}")
+        if not 0.0 <= threshold <= 1.0:
+            raise ValueError(f"threshold must be in [0, 1], got {threshold}")
+
+        # Extract features using configured methods
+        features: dict[str, float] = {}
+
+        if "statistical" in self.methods:
+            features.update(self._extract_statistical_features(signal))
+
+        if "frequency" in self.methods:
+            features.update(self._extract_frequency_features(signal, sample_rate))
+
+        if "pattern" in self.methods:
+            features.update(self._detect_digital_patterns(signal, sample_rate))
+
+        # Classify from features
+        signal_type, confidence, alternatives = self._classify_from_features(features, threshold)
+
+        # Generate reasoning
+        reasoning = self._generate_reasoning(signal_type, features)
+
+        return ClassificationResult(
+            signal_type=signal_type,
+            confidence=confidence,
+            features=features,
+            secondary_matches=alternatives,
+            reasoning=reasoning,
+        )
+
+    def classify_batch(
+        self,
+        signals: list[NDArray[np.floating[Any]]],
+        sample_rate: float,
+        threshold: float = 0.5,
+    ) -> list[ClassificationResult]:
+        """Classify multiple signals.
+
+        Classifies signals sequentially. For large batches, consider using
+        multiprocessing for parallel processing.
+
+        Args:
+            signals: List of signal arrays to classify
+            sample_rate: Sample rate in Hz (same for all signals)
+            threshold: Minimum confidence for primary classification
+
+        Returns:
+            List of ClassificationResult objects
+
+        Raises:
+            ValueError: If signals list is empty
+
+        Example:
+            >>> results = classifier.classify_batch(signals, sample_rate=1e6)
+            >>> for i, result in enumerate(results):
+            ...     print(f"Signal {i}: {result.signal_type}")
+        """
+        if not signals:
+            raise ValueError("Cannot classify empty signal list")
+
+        return [self.classify(signal, sample_rate, threshold) for signal in signals]
+
+    def _extract_statistical_features(
+        self,
+        signal: NDArray[np.floating[Any]],
+    ) -> dict[str, float]:
+        """Extract statistical features from signal.
+
+        Features:
+            - mean: Average voltage level
+            - variance: Signal variance (normalized)
+            - min/max: Voltage range
+            - duty_cycle: Fraction of time signal is high (for digital)
+            - edge_count: Number of transitions
+            - edge_density: Edges per sample
+
+        Args:
+            signal: Signal data array
+
+        Returns:
+            Dictionary of statistical features
+
+        Example:
+            >>> features = classifier._extract_statistical_features(signal)
+            >>> print(features['duty_cycle'])
+            0.52
+        """
+        mean = float(np.mean(signal))
+        variance = float(np.var(signal))
+        min_val = float(np.min(signal))
+        max_val = float(np.max(signal))
+        voltage_swing = max_val - min_val
+
+        # Normalize variance by voltage swing to make it scale-independent
+        normalized_variance = variance / (voltage_swing**2 + 1e-10)
+
+        # Digital features: threshold at midpoint
+        threshold = (max_val + min_val) / 2
+        digital = (signal > threshold).astype(int)
+        edges = np.diff(digital)
+        edge_count = int(np.count_nonzero(edges))
+        edge_density = edge_count / len(signal) if len(signal) > 0 else 0.0
+
+        # Duty cycle (fraction of time high)
+        duty_cycle = float(np.mean(digital))
+
+        return {
+            "mean": mean,
+            "variance": normalized_variance,
+            "min": min_val,
+            "max": max_val,
+            "voltage_swing": voltage_swing,
+            "duty_cycle": duty_cycle,
+            "edge_count": edge_count,
+            "edge_density": edge_density,
+        }
+
+    def _extract_frequency_features(
+        self,
+        signal: NDArray[np.floating[Any]],
+        sample_rate: float,
+    ) -> dict[str, float]:
+        """Extract frequency domain features via FFT.
+
+        Features:
+            - dominant_frequency: Frequency with highest power (Hz)
+            - bandwidth: Frequency range with >10% of peak power (Hz)
+            - spectral_centroid: Center of mass of spectrum (Hz)
+            - spectral_flatness: Ratio of geometric to arithmetic mean (0-1)
+
+        Args:
+            signal: Signal data array
+            sample_rate: Sample rate in Hz
+
+        Returns:
+            Dictionary of frequency domain features
+
+        Example:
+            >>> features = classifier._extract_frequency_features(signal, 1e6)
+            >>> print(features['dominant_frequency'])
+            115200.0
+        """
+        if len(signal) < 2:
+            return {
+                "dominant_frequency": 0.0,
+                "bandwidth": 0.0,
+                "spectral_centroid": 0.0,
+                "spectral_flatness": 0.0,
+            }
+
+        # Compute FFT
+        fft = np.fft.rfft(signal)
+        freqs = np.fft.rfftfreq(len(signal), 1.0 / sample_rate)
+        magnitude = np.abs(fft)
+
+        # Dominant frequency (skip DC component)
+        if len(magnitude) > 1:
+            dominant_idx = np.argmax(magnitude[1:]) + 1
+            dominant_frequency = float(freqs[dominant_idx])
+        else:
+            dominant_frequency = 0.0
+
+        # Bandwidth (frequencies with >10% of max power)
+        max_power = np.max(magnitude)
+        if max_power > 0:
+            threshold_power = 0.1 * max_power
+            active_freqs = freqs[magnitude > threshold_power]
+            bandwidth = float(active_freqs[-1] - active_freqs[0]) if len(active_freqs) > 1 else 0.0
+        else:
+            bandwidth = 0.0
+
+        # Spectral centroid (center of mass)
+        if np.sum(magnitude) > 0:
+            spectral_centroid = float(np.sum(freqs * magnitude) / np.sum(magnitude))
+        else:
+            spectral_centroid = 0.0
+
+        # Spectral flatness (measure of how noise-like the spectrum is)
+        # Geometric mean / arithmetic mean
+        # 0 = tonal (single frequency), 1 = noise-like (flat spectrum)
+        if len(magnitude) > 0 and np.all(magnitude > 0):
+            geometric_mean = np.exp(np.mean(np.log(magnitude + 1e-10)))
+            arithmetic_mean = np.mean(magnitude)
+            spectral_flatness = float(geometric_mean / (arithmetic_mean + 1e-10))
+        else:
+            spectral_flatness = 0.0
+
+        return {
+            "dominant_frequency": dominant_frequency,
+            "bandwidth": bandwidth,
+            "spectral_centroid": spectral_centroid,
+            "spectral_flatness": spectral_flatness,
+        }
+
+    def _detect_digital_patterns(
+        self,
+        signal: NDArray[np.floating[Any]],
+        sample_rate: float,
+    ) -> dict[str, float]:
+        """Detect protocol-specific patterns in signal.
+
+        Computes scores for:
+            - uart_score: UART bit timing alignment
+            - spi_score: SPI clock consistency
+            - i2c_score: I2C pattern characteristics
+            - can_score: CAN encoding characteristics
+            - periodicity: Signal periodicity measure
+
+        Args:
+            signal: Signal data array
+            sample_rate: Sample rate in Hz
+
+        Returns:
+            Dictionary of pattern detection scores (0.0-1.0)
+
+        Example:
+            >>> patterns = classifier._detect_digital_patterns(signal, 1e6)
+            >>> print(patterns['uart_score'])
+            0.85
+        """
+        # Threshold signal to digital
+        threshold = (np.max(signal) + np.min(signal)) / 2
+        digital = (signal > threshold).astype(int)
+        edges = np.diff(digital)
+        edge_indices = np.where(np.abs(edges) > 0)[0]
+
+        if len(edge_indices) < 3:
+            return {
+                "uart_score": 0.0,
+                "spi_score": 0.0,
+                "i2c_score": 0.0,
+                "can_score": 0.0,
+                "periodicity": 0.0,
+            }
+
+        # Edge intervals
+        edge_intervals = np.diff(edge_indices)
+        if len(edge_intervals) == 0:
+            return {
+                "uart_score": 0.0,
+                "spi_score": 0.0,
+                "i2c_score": 0.0,
+                "can_score": 0.0,
+                "periodicity": 0.0,
+            }
+
+        # Periodicity score (coefficient of variation)
+        mean_interval = np.mean(edge_intervals)
+        std_interval = np.std(edge_intervals)
+        periodicity = 1.0 - min(1.0, std_interval / (mean_interval + 1e-10))
+
+        # UART score: check alignment with common baud rates
+        uart_score = self._compute_uart_score(edge_intervals, sample_rate)
+
+        # SPI score: high edge density + consistent timing
+        edge_density = len(edge_indices) / len(signal)
+        consistency = 1.0 - min(1.0, std_interval / (mean_interval + 1e-10))
+        spi_score = min(1.0, edge_density * 10 * consistency)
+
+        # I2C score: lower edge density than SPI, some irregularity (clock stretching)
+        # I2C typically has burst patterns with pauses
+        i2c_score = 0.5 if 0.05 < edge_density < 0.3 and periodicity < 0.9 else 0.0
+
+        # CAN score: similar to digital but with specific encoding patterns
+        # CAN uses bit stuffing - look for irregularity in bit timing
+        can_score = 0.5 if 0.7 < periodicity < 0.95 and edge_density > 0.1 else 0.0
+
+        return {
+            "uart_score": uart_score,
+            "spi_score": spi_score,
+            "i2c_score": i2c_score,
+            "can_score": can_score,
+            "periodicity": periodicity,
+        }
+
+    def _compute_uart_score(
+        self,
+        edge_intervals: NDArray[np.integer[Any]],
+        sample_rate: float,
+    ) -> float:
+        """Compute UART likelihood score based on baud rate alignment.
+
+        Checks if edge intervals align with common UART baud rates.
+
+        Args:
+            edge_intervals: Array of sample counts between edges
+            sample_rate: Sample rate in Hz
+
+        Returns:
+            UART score (0.0-1.0)
+        """
+        common_bauds = [9600, 19200, 38400, 57600, 115200, 230400, 460800, 921600]
+        baud_scores = []
+
+        for baud in common_bauds:
+            bit_period_samples = sample_rate / baud
+            # Count edges that align with this baud rate (within 20% tolerance)
+            tolerance = 0.2
+            aligned = np.sum(
+                np.abs(edge_intervals % bit_period_samples) < bit_period_samples * tolerance
+            )
+            score = aligned / len(edge_intervals) if len(edge_intervals) > 0 else 0.0
+            baud_scores.append(score)
+
+        return float(max(baud_scores)) if baud_scores else 0.0
+
+    def _classify_from_features(
+        self,
+        features: dict[str, float],
+        threshold: float,
+    ) -> tuple[SignalType, float, list[tuple[SignalType, float]]]:
+        """Make classification decision from extracted features.
+
+        Applies classification rules and selects best match.
+
+        Args:
+            features: Dictionary of extracted features
+            threshold: Minimum confidence for primary classification
+
+        Returns:
+            Tuple of (signal_type, confidence, alternatives)
+                alternatives is list of (type, confidence) for secondary matches
+        """
+        # Evaluate all rules
+        scores: dict[SignalType, float] = {}
+
+        for rule in self.RULES:
+            score = self._evaluate_rule(rule, features)
+            scores[rule.signal_type] = score
+
+        # Sort by score
+        sorted_scores = sorted(scores.items(), key=lambda x: x[1], reverse=True)
+
+        # Best match
+        best_type, best_score = sorted_scores[0]
+
+        # If score is below threshold, mark as unknown
+        if best_score < threshold:
+            return "unknown", best_score, [(best_type, best_score)]
+
+        # Collect alternatives (other types with score >= threshold * 0.6)
+        alt_threshold = threshold * 0.6
+        alternatives = [
+            (sig_type, score) for sig_type, score in sorted_scores[1:] if score >= alt_threshold
+        ]
+
+        return best_type, best_score, alternatives
+
+    def _evaluate_rule(
+        self,
+        rule: ClassifierRule,
+        features: dict[str, float],
+    ) -> float:
+        """Evaluate classification rule against features.
+
+        Args:
+            rule: Classification rule to evaluate
+            features: Extracted features
+
+        Returns:
+            Match score (0.0-1.0), weighted by rule weight
+        """
+        matches = 0
+        total = len(rule.conditions)
+
+        for feature_name, (min_val, max_val) in rule.conditions.items():
+            if feature_name in features:
+                value = features[feature_name]
+                if min_val <= value <= max_val:
+                    matches += 1
+
+        # Base score is fraction of conditions met
+        base_score = matches / total if total > 0 else 0.0
+
+        # Apply rule weight
+        weighted_score = base_score * rule.weight
+
+        return min(1.0, weighted_score)
+
+    def _generate_reasoning(
+        self,
+        signal_type: SignalType,
+        features: dict[str, float],
+    ) -> str:
+        """Generate human-readable explanation of classification.
+
+        Args:
+            signal_type: Classified signal type
+            features: Features used for classification
+
+        Returns:
+            Reasoning string
+        """
+        if signal_type == "digital":
+            return (
+                f"Digital signal detected: high variance ({features.get('variance', 0):.2f}), "
+                f"edge density {features.get('edge_density', 0):.3f}"
+            )
+        elif signal_type == "analog":
+            return (
+                f"Analog signal detected: low edge density ({features.get('edge_density', 0):.3f}), "
+                f"continuous values"
+            )
+        elif signal_type == "pwm":
+            return (
+                f"PWM signal detected: periodic pattern (periodicity {features.get('periodicity', 0):.2f}), "
+                f"duty cycle {features.get('duty_cycle', 0):.2f}"
+            )
+        elif signal_type == "uart":
+            return (
+                f"UART signal detected: baud rate alignment score {features.get('uart_score', 0):.2f}, "
+                f"edge density {features.get('edge_density', 0):.3f}"
+            )
+        elif signal_type == "spi":
+            return (
+                f"SPI signal detected: high edge density ({features.get('edge_density', 0):.3f}), "
+                f"consistent timing"
+            )
+        elif signal_type == "i2c":
+            return f"I2C signal detected: characteristic patterns (score {features.get('i2c_score', 0):.2f})"
+        elif signal_type == "can":
+            return (
+                f"CAN signal detected: encoding patterns (score {features.get('can_score', 0):.2f})"
+            )
+        else:
+            return "Signal type unclear: low confidence in all classifications"
+
+
+__all__ = [
+    "ClassificationResult",
+    "ClassifierRule",
+    "SignalClassifier",
+    "SignalType",
+]