Farmers faced challenges with pest control, leading to lower crop yields and financial losses due to limited access to accurate information. To address this, the study introduced pestaway, a chatbot for Filipino farmers that uses Large Language Models (LLMs) to provide interactive and speech-enabled solutions. An 80,000-item data set divided between English and Tagalog, was generated using GPT-4 and credible agricultural sources. The models were fine-tuned using different training approaches: Supervised Fine-Tuning (SFT), SFT with Direct Preference Optimization (DPO), and Odds Ratio Preference Optimization (ORPO). Among them, the ORPO Trainer produced the best results, achieving a METEOR score of 0.3271, a BERTScore of 0.7316, and an average perplexity of 1.5185. Additionally, word error rate (WER) and character error rate (CER) were used to evaluate the speech recognition capabilities of the chatbot. The WER ranged from 0% to 28.57%, which showed that some transcriptions exceeded the acceptable threshold, while the CER ranged from 0% to 3.43%, which indicated more accurate transcriptions due to minimal character-level errors. These findings demonstrated that the chatbot was able to recognize voice queries in English and Tagalog with reliable precision. Furthermore, a user satisfaction survey involving farmers, agricultural students, and individuals interested in agricultural practices yielded an average rating of more than 4.1 out of 5, with 97.73% giving positive feedback and 100% recommending the chatbot for use. The participants highlighted the usefulness, precision, and bilingual capability of the chatbot and suggested additional enhancements such as faster responses, visual aids, and a mobile version.

Existing data acquisition systems for lower-limb assessment typically rely on single-sensor modalities and devicespecific protocols, which limit accurate measurement of both knee angles and peak muscle activations during exercises like standing leg curls. To address these limitations, we developed a wireless device integrating IMU and sEMG sensors, synchronizing data via timestamps and dynamic time warping. The Kalman filter is used to obtain the IMU signals, while an adaptive filter is used to denoise the sEMG signals. From the combined IMU and sEMG recordings, clear trends emerge in peak leg curl power as a function of knee flexion angle; to visualize these patterns, we applied K-means clustering (k$=3)$to knee flexion angle$\left(40^{\circ}-115^{\circ}\right)$versus peak leg curl power ($0.2-1.0 ~\mathrm{W}$) across all repetitions, which revealed three regimes, low ($\approx 60^{\circ}-80^{\circ}, 0.5-0.8 ~\mathrm{W}$), medium ($\approx 80^{\circ}-95^{\circ}$,$0.7-0.9 ~\mathrm{W}$), and high ($\approx 95^{\circ}-115^{\circ}, 0.7-1.0 ~\mathrm{W}$), highlighting typical performance zones. Analysis of the resulting scatter plot showed peak muscle activation during concentric phases at flexion angles of$\mathbf{9 7}^{\boldsymbol{\circ}} \boldsymbol{-} \mathbf{1 1 3}^{\boldsymbol{\circ}}$. Prototype knee-angle measurements achieved RMSE 7.37°, MAE 4.76${ }^{\circ}$, ICC 0.98, and Spearman's$\rho=0.9765$against ground truth, confirming the system's reliability in data acquisition. Overall, the integrated system provided reliable measurements of muscle activation and knee angles with acceptable error margins and consistency.

Indoor navigation systems play a crucial role in guiding users through complex environments such as airports, shopping malls, hospitals, and university campuses. With the advancement of Artificial Intelligence (AI), researchers have explored dialogue-based approaches to enhance user interaction within these systems. However, existing models face limitations in accurately processing diverse user inputs, which poses challenges in efficiency and usability. To address this, the researchers developed a virtual assistant using a Large Language Model (LLM) to facilitate a dialogue-driven navigation experience. The study was conducted in three phases: (1) generation of datasets, (2) development of models, and (3) virtual assistant performance assessment through a user satisfaction survey. The dataset, consisting of 200,000 entries, was split into 70% for training and 30% for testing, designed to train the navigation model on real-world scenarios. Due to hardware limitations, specifically the 4GB of VRAM on NVIDIA GeForce GTX 1650 and RTX 3050 GPUs, the GPT-2 model was selected as the base model. Despite being an outdated and less capable model for handling complex language tasks, GPT-2 was chosen due to its compatibility with limited hardware resources, such as GPUs with only 4GB of VRAM. Among the training methods tested, the General Purpose Trainer achieved the best performance, with a BERT score of 0.89, a METEOR score of 0.84, and an average perplexity of 2.82, indicating moderate success in understanding and responding to user input. A user satisfaction survey further validated the practicality of the system, with an average rating of 87.44%. Although the model shows promise, its effectiveness in interpreting highly unstructured or unconventional queries could be improved using more advanced LLMs.

Hamstring muscles are commonly injured in sports involving strenuous activities. Evaluating the muscle condition is done through physical examinations like ultrasonography and MRI, which provide accurate assessments but are costly. An alternative method in obtaining information from muscle activity is through Surface Electromyography (sEMG) signals, which is recorded through electrical signals. Particularly, the nonlinear features are assessed to monitor muscle activity in the hamstring since these can determine intrinsic changes in muscle behavior. These nonlinear features bring about high dimensionality to the signal, which requires manifold learning to reduce its dimensionality and extract its underlying nonlinear features and relationships. In this paper, the researchers propose a solution that provides an automated classification of HSI based on nonlinear features of sEMG signals using machine learning algorithms such as Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LoR), along with manifold learning techniques such as t-distributed Stochastic Neighborhood Embedding (t-SNE) and Locally Linear Embedding (LLE). The study results showed that SVM was considered the best overall predictive model for classification, with an accuracy of 88.51% when applied on t-SNE and 75.40% when applied on LLE.

Inertial measurement unit (IMU) sensors are widely used in sports science to acquire motion-related data, such as joint angle measurements, and provide valuable insights into the performance of the athletes. These sensors can be used for specific purposes that needs regular calibration to compensate for drift since assessment of muscle stress rely on these devices. However, sensor-based systems may not always provide accurate joint angle measurements. This paper aims to mitigate the drift of the IMU data by pre-processing it with various sensor fusion algorithms: Complementary filter, Kalman filter, and Madgwick filter. The output of the calibrated IMU sensor device was compared with the data collected from the VICON motion capture system to validate its accuracy. Finally, the calibrated IMU sensor output is used to measure the joint angles. As a result, the complementary filter resulted in a better accuracy and reliability performance than the Kalman and Magdwick filter. The Kalman filter obtained a higher reliability and precision due to its lower variability and standard deviation of 37.68. Moreover, it also indicated that the complementary filter often yielded low error values, suggesting better performance as seen in its Root- Mean-Square Error (RMSE) of 9.75. An Analysis of Variance (ANOVA) test was done, indicating that the RMSE accepts the null hypothesis, H0, that there are no significant differences among the means of the metrics.

Surface electromyography (sEMG) is crucial in sports science, offering insights into muscle activation patterns. However, standard devices like the Delsys Trigno Wireless System are relatively costly. This study presents a prototype that simultaneously obtains sEMG and IMU signals to measure muscle activity. The presence of noise necessitates signal processing techniques to be applied to the raw signal, including feed-forward comb, adaptive, and wavelet filters. Evaluation against a reference signal revealed that the wavelet filter performed best, exhibiting the lowest MSE and highest SNR scores. Specifically, it achieved MSE scores of 2.62E-10 in the time domain, 6.25E-22 in PSD, and SNR scores 13.96 for squats. These findings underscore the effectiveness of the wavelet filter in reducing sEMG signal noise.

The hamstring muscle is frequently susceptible to injury among athletes across various sports. This paper investigates the use of multidimensional scaling (MDS) to project data points derived from IMU and sEMG data into a lower dimensional space to characterize hamstring muscle activation during leg squats. The research aims to assemble a wearable sensor system for hamstring data acquisition and to characterize the extracted features, including the knee angle, thigh Euler angles, and sEMG signals, analyzing trends and behaviors in the manifold representation. Then, recurrence quantification analysis (RQA) was used to generate recurrence plots, revealing relevant insights into the underlying dynamics, predictability, and complexity of hamstring muscle activity. A t-test on the RQA parameters between rest and active states showed significant differences, indicating distinct behaviors between these states.

Philippines is a tropical country and every year, the country is experiencing typhoons, thunderstorms and excessive rainfalls because of climate change. Since then, the government continuously provides efforts to mitigate natural disasters through numerously growing researches that are inclined with the meteorological processes happening in our country. There are many researches with regards to the methods of quantifying the amount of rainfall but based on those studies, the acoustic disdrometer is rendered useless because of its inability to classify ambient noise from rain types. With this, the main purpose of this study is to develop an improved acoustic disdrometer by adding a capability in which it will categorize the intensity of the amount of rainfall from ambient noise using machine learning algorithm. The proposed methodology is applied by developing a prototype with four piezoelectric sensors, Arduino microcontroller and ZigBee transmitter. Also, the K-nearest neighbors (KNN) predictive model will be established. The obtained results show that the accuracy of the predictive model is 89.95%.