COMPREHENSIVE ANALYSIS AND CLASSIFICATION OF SIGNALS IN MILK PASTEURIZATION PROCESS USING MACHINE LEARNING ALGORITHMS

Аннотация

This article investigates the classification and quality evaluation of process signals in milk pasteurization using machine learning methods. The temperature signal, which is vital for assessing the pasteurization process efficiency, is analyzed in detail. Through statistical feature extraction and intelligent classification, the quality of signals is determined. Algorithms such as Random Forest and Autoencoder are applied for classification and anomaly detection, respectively. Results demonstrate that combining both approaches improves robustness and reliability. The method is implemented using Python and tested with Raspberry Pi sensor inputs, simulating real-time industrial conditions. Keywords: pasteurization, temperature signal, machine learning, Random Forest, Autoencoder, anomaly detection, signal quality. 1.Introduction Milk pasteurization ensures microbiological safety by heating milk to a specific temperature and maintaining it for a determined period. Signal analysis is crucial to confirm that the thermal process is executed effectively. The most important signal in this process is the temperature signal, which is captured via digital or analog sensors. These signals may vary due to noise, sensor degradation, or external interferences. Hence, reliable signal classification is necessary. The purpose of this study is to develop a hybrid methodology using machine learning to distinguish between high-quality and poor-quality signals. This would reduce human error, increase automation, and improve milksafety. 2.SignalTypes andTheirCharacteristics Pasteurizationsystems generatethefollowing types of signals:  Temperature Signals: Directly related to microbial safety, usually continuous and analog.  Flow Rate Signals: Indicates milk movement; varies depending on valve and pump settings.  PressureSignals: Reflect fluid resistance; affected bycloggingor mechanical failures.  SensorStatus Signals:Binary oranalog indicators showing sensorhealth. http://www.internationaljournal.co.in/index.php/jasass 801 Volume 15Issue 06, June 2025 Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023: 6.995, 2024 7.75 Each signal type has distinct characteristics. Temperature signals tend to be smooth and periodic. In contrast, flow and pressure signals may show sudden changes. Low-quality signals typically exhibit characteristics such as irregular spikes, flat-lining, or erratic oscillation. Accurate modelingof each signal’s normal behavior is crucial forcomparison andclassification. 3.Differentiating Quality andNon-quality Signals To evaluate signal quality, several statistical and frequency-domain features are extracted:  Statistical Features:Mean, standard deviation, skewness, kurtosis.  Frequency Features: Spectral entropy, FFT (Fast Fourier Transform)peak amplitudes.  Time-Domain Features: Gradient changes, numberofzero crossings, signal energy. These features are used as input for a supervised classifier. High-quality signals have consistent metrics and align well with physical models. Low-quality signals exhibit abrupt transitions, high noise-to-signal ratios, or deviate from expected patterns. A key part of this work is the creation ofalabeled dataset based onexpert knowledge andsyntheticnoiseinjection. 4.Review ofExistingWork Prior research in industrial automation has utilized machine learning for signal quality detection. Notablemethods include:  RandomForest (Breiman, 2001): Used forclassification tasks with highinterpretability.  Autoencoders (Goodfellow et al., 2016): Neural networks trained to reconstruct input data; useful for anomaly detection.  LSTMNetworks: Applied in time-series forecasting andanomaly prediction. These methods have shown success in areas such as vibration monitoring, power quality assessment, and sensor drift detection. However, few studies have focused specifically on milk pasteurization. Moreover, combining classification with reconstruction error analysis remains underexplored. 5.Proposed AlgorithmandMethodology Wepropose atwo-phase hybrid approach: Phase 1:Signal Classification  Signal is preprocessed using Savitzky-Golay filter.  Features are extracted: {mean, std, kurtosis, FFT peak, entropy}.  Atrained RandomForest classifier determines thequality label. http://www.internationaljournal.co.in/index.php/jasass 802 Volume 15Issue 06, June 2025 Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023: 6.995, 2024 7.75 Phase 2:Anomaly Detection  Thesame signal is fed intoatrained Autoencoder.  If reconstruction error exceeds athreshold, itis flagged as “non-quality”. Algorithm1:SignalQuality Detection Input:Temperature signal S(t) Output:Quality label ∈{High, Low} 1.Apply smoothingfilter: S_filtered ←SG_Filter(S(t)) 2.Extract features: F←Feature_Extraction(S_filtered) 3.Label ←RandomForest_Classifier(F) 4.Recon ←Autoencoder(S_filtered) 5.Error ←MSE(S_filtered, Recon) 6.If Error >ε,then Label ←Low Return Label This approach combines supervise