پایش هوشمند سلامت گاو شیری با پردازش سیگنال مبتنی بر هوش مصنوعی: راهکاری یکپارچه در کشاورزی دقیق و سیستم‌های اینترنت اشیاء (IoT)

Precision livestock farming has emerged as an effective approach for improving animal health, welfare, and production efficiency through continuous behavioral monitoring. Among the various behavioral indicators, feeding and rumination are recognized as two of the most informative physiological activities because changes in these behaviors are closely associated with metabolic disorders, digestive diseases, and overall animal health. Conventional monitoring techniques, including visual observation and camera-based systems, are labor-intensive, costly, and difficult to implement in large-scale dairy farms. Consequently, the development of low-cost intelligent monitoring systems capable of real-time behavioral analysis has become an important research objective.
Introduction
This study presents a comparative investigation of two signal processing approaches for dairy cattle behavior recognition using a low-power piezoelectric sensing system integrated with artificial intelligence techniques. The primary objective was to evaluate the effectiveness of conventional time-domain feature extraction and Discrete Wavelet Transform (DWT)-based feature extraction for identifying feeding and rumination behaviors while considering both classification accuracy and computational efficiency. The proposed framework combines wearable sensing, wireless data acquisition, signal preprocessing, deep learning, and Internet of Things (IoT) communication into a unified monitoring platform suitable for precision livestock farming.
Materials and Methods
Data acquisition was performed at the dairy farm of the University of Tabriz using a piezoelectric sensor mounted on the animal's halter to capture jaw movement vibrations during different activities. An ESP32 microcontroller was employed for signal acquisition and Bluetooth communication with a mobile application, enabling wireless data collection and storage. The recorded datasets were transferred to a cloud storage platform for subsequent processing in Python using the Google Colab environment. Signal preprocessing included digital noise filtering, Min–Max normalization, and segmentation into fixed-length windows containing 20–30 samples. Because of the imbalance between behavioral classes, particularly the limited number of feeding samples, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to generate balanced training data.
 
Results and Discussion
Two independent feature extraction strategies were investigated. In the first approach, statistical time-domain descriptors, including the mean, standard deviation, minimum, and maximum values, were extracted from each signal segment. In the second approach, Discrete Wavelet Transform (DWT) was employed to obtain time-frequency representations capable of capturing transient characteristics of jaw movements. Both feature sets were subsequently used to train identical one-dimensional Convolutional Neural Network (1D-CNN) models implemented in TensorFlow/Keras. The network architecture consisted of successive convolutional, batch normalization, activation, pooling, dropout, and fully connected layers. Model optimization was performed using the Adam optimizer with categorical cross-entropy loss, while Early Stopping and adaptive learning-rate reduction were incorporated to improve convergence and prevent overfitting. Model performance was evaluated using validation accuracy, precision, recall, F1-score, confusion matrices, and learning curves.
The experimental results demonstrated clear differences between the two signal processing strategies. The time-domain approach achieved the highest overall performance, reaching a validation accuracy of 78.86%, while requiring only 21.68 minutes for model training and approximately 0.17 million trainable parameters. In contrast, the DWT-based approach achieved a validation accuracy of approximately 76%, required 23.44 minutes of training time, and increased the model complexity to approximately 0.19 million parameters. These findings indicate that eliminating wavelet transformation significantly reduces computational complexity while maintaining superior classification performance, making the proposed approach particularly suitable for resource-constrained intelligent monitoring systems.
A detailed analysis of individual behavioral classes revealed different strengths for the two approaches. For the Feeding class, the DWT-based method achieved a slightly higher F1-score (0.7627) than the time-domain approach (0.752), suggesting that time-frequency decomposition can better represent the transient and irregular characteristics of feeding behavior. Conversely, the time-domain method produced superior performance for Rumination, achieving an F1-score of approximately 0.9903, compared with 0.9888 obtained by the DWT-based model. Since rumination exhibits highly repetitive and periodic jaw movements, its essential characteristics can be effectively captured using simple statistical features without requiring computationally expensive spectral analysis.
The comparative evaluation indicates that the selection of signal processing techniques should be guided by the intended operational objectives of intelligent livestock monitoring systems. For large-scale dairy farms where computational resources, energy consumption, and implementation costs are important considerations, the conventional time-domain approach provides an excellent balance between classification accuracy, processing speed, and model simplicity. In contrast, DWT-based analysis may be advantageous in applications requiring enhanced sensitivity for detecting feeding behavior, despite its higher computational cost.
Conclusion
Overall, this study demonstrates that accurate behavioral monitoring of dairy cattle can be achieved using a low-cost piezoelectric sensing platform combined with deep learning techniques. The proposed framework offers a practical solution for IoT-enabled precision livestock farming by integrating wearable sensing, wireless communication, intelligent signal processing, and automated behavioral classification. The results further suggest that computationally efficient time-domain feature extraction represents the most suitable strategy for real-time deployment in practical livestock monitoring systems, while wavelet-based analysis remains a promising complementary technique for applications requiring more detailed characterization of complex behavioral patterns. These findings contribute to the development of intelligent animal health monitoring systems capable of supporting sustainable dairy farming through continuous, automated, and data-driven behavioral assessment.