Yun-Wei Lin

Anthracnose diseases can severely affect strawberry plant stands and yields, making early detection essential. Manual identification of diseased plants is labor-intensive, prompting the development of IoT-based AI (AIoT) solutions for more efficient detection. While many AIoT methods rely on high-performance servers typically hosted in the cloud, edge solutions are preferable for commercial farms to reduce dependency on network connections. To implement an effective edge-based system, the computation hardware must feature a capable CPU for running the IoT engine and a GPU sufficient for YOLO-based detection, while maintaining code security. This article explores the StrawberryTalk-v2 solution using the NVIDIA Jetson Nano, secured with Winbond’s W77Q TRUSTME® Secure Serial Flash Memory, fulfilling these requirements. The deployment of the StrawberryTalk-v2 IoT solution on this hardware facilitates strawberry anthracnose detection. Choosing a YOLO version that balances edge device efficiency and high detection accuracy presents a significant challenge. While YOLO11 Nano’s computational complexity is well-suited for edge deployment, its baseline accuracy falls short of YOLO11 Extra Large. By incorporating WIoU and DySample into YOLO11 Nano, the detection accuracy exceeds that of the Extra Large version while retaining the Nano version’s low execution cost. These enhancements allow the system to achieve superior performance, with StrawberryTalk-v2 reaching a four-level detection accuracy of 95.5%.

This article introduces VoiceTalk, a no-code approach that develops voice-controlled smart home applications without requiring programming expertise. At its core, VoiceTalk utilizes IoTtalk, an IoT application development platform for managing a diverse range of IoT devices. IoTtalk employs a two-tier microservices architecture, enabling users to define and chain applications through an intuitive drag-and-drop line interface. Leveraging its microservice architecture, VoiceTalk integrates IoTtalk with Google Home, offering a no-code solution for voice-controlled applications. VoiceTalk leverages its understanding of smart appliances in the room/house to generate specific prompts. We have compared the translation accuracy of 7 Automatic Speech Recognition (ASR) systems. We make two contributions. First, the no-code VoiceTalk platform significantly simplifies the development of Google Home-like applications. Second, by integrating ASRs with a commercial LLM such as GPT, we dramatically reduce voice-to-text translation errors, for examples, from 5.13% to 0.54% for the Web Speech API and from 2.25% to zero for Whisper Medium. For small-sized open-source LLMs such as Llama 3.2 3B, the errors are reduced to 0.72% for the Web Speech API and to zero for Whisper Medium. Furthermore, Device LLM Agent of VoiceTalk can be easily extended to integrate IoTtalk with other voice platforms, such as AWS Alexa.

Traditional methods of orchid disease detection rely on manual inspection, which is both labor-intensive and inefficient. Modern smart solutions integrate Internet of Things (IoT) and Artificial Intelligence (AI) technologies to address these limitations in commercial greenhouses. Our approach leverages IoT for real-time monitoring of critical environmental factors, such as temperature and humidity, to enhance disease prediction accuracy, with a specific focus on Phalaenopsis orchids. We propose OrchidTalk-v2, a novel and dynamically adaptable system designed to improve disease risk monitoring. OrchidTalk-v2 integrates an intelligent spore germination sensor with Continuous Wavelet Transform (CWT) for feature extraction, feeding the data into a three-dimensional Convolution LSTM network model. The system achieves a precision rate exceeding 93% and provides early alerts for disease outbreaks with a recall rate above 92.75%. This marks a significant advancement in prediction accuracy for orchid disease detection within controlled greenhouse environments.

Existing non-image-based solutions for detecting Phalaenopsis orchid diseases rely on time-domain analysis of environmental conditions. These solutions are suitable for loosely controlled open farm fields. Conversely, the diseases are infrequent events in Internet of Things (IoT)-based smart orchid greenhouses, where environmental conditions are well-controlled. In tightly controlled settings, standard time domain methods fall short in detecting subtle changes in environmental conditions during rare orchid disease events. We observed that the influence of the time series of temperature and humidity on orchid fungus within the well-controlled greenhouse is not readily apparent, and the time-frequency coefficients of temperature and humidity must be used to detect orchid disease in this rare event case. In order to address this challenge, we propose OrchidTalk, an IoT-based deep learning platform, which employs a time-frequency domain analysis using the continuous wavelet transform (CWT) filter, for the purpose of detecting rare events of orchid diseases. Through a nontrivial process, we select the Gaussian wavelet with four derivatives to serve as the mother wavelet of the CWT in the data extraction phase, and invent the CWT IoT device to translate the time-domain IoT raw data to the time-frequency domain IoT data. The resulting CWT features are used as inputs to the Orchid-3D model (3-D ConvLSTM), which significantly outperforms the previously proposed Orchid-1D (CNN and LSTM) and Orchid-2D (2-D ConvLSTM) models. For the rare event detection of 0.2% of sick orchid plants, OrchidTalk achieves a recall of 0.902, precision of 0.92, accuracy of 0.99, and an F1-score of 0.911. To the best of our knowledge, this represents the highest prediction performance achieved for a well-controlled orchid greenhouse.

Smart agriculture utilizes Internet of Things (IoT) technologies to enable low-cost electrical conductivity (EC) sensors to support farming intelligence. Due to aging and changes in weather and soil conditions, EC sensors are prone to long-term drift over years of operation. Therefore, regular recalibration is necessary to ensure data accuracy. In most existing solutions, an EC sensor is calibrated by using the standard sensor to build the calibration table. This paper proposes SensorTalk3, an ensemble approach of machine learning models including XGBOOST and Random Forest, which can be executed at an edge device (e.g., Raspberry Pi) without GPU acceleration. Our study indicates that the soil information (both temperature and moisture sensor data) plays an important role in SensorTalk3, which significantly outperforms the existing calibration approaches. The MAPE of SensorTalk3 can be as low as 1.738%, compared to the 7.792% error of the original sensor. Our study indicates that when the errors of uncalibrated moisture and temperature sensors are not larger than 8.3%, SensorTalk3 can accurately calibrate EC. SensorTalk3 can perform model training during data collection at the edge node. When all training data are collected, AI training is also finished at the edge node. Such an AI training approach has not been found in existing edge AI approaches. We also proposed the dual-sensor detection solution to determine when to conduct recalibration. The overhead of this solution is less than twice the optimal detection scenario (which cannot be achieved practically). If the two non-standard sensors are homogeneous and stable, then the optimal detection scenario can be approached. Conventional methods require training calibration AI models in the cloud. However, SensorTalk3 introduces a significant advancement by enabling on-site transfer learning in the edge node. Given the abundance of farming sensors deployed in the fields, performing local transfer learning using low-cost edge nodes proves to be a more cost-effective solution for farmers.

Deep learning models are often trained with a large amount of labeled data to improve the accuracy for moving object detection in new fields. However, the model may not be robust enough due to insufficient training data in the new field, resulting in some moving objects not being successfully detected. Training with data that is not successfully detected by the pre-trained deep learning model can effectively improve the accuracy for the new field, but it is costly to retrieve the image data containing the moving objects from millions of images per day to train the model. Therefore, we propose FenceTalk, a moving object detection system, which compares the difference between the current frame and the background image based on the structural similarity index measure (SSIM). FenceTalk automatically selects suspicious images with moving objects that are not successfully detected by the Yolo model, so that the training data can be selected at a lower labor cost. FenceTalk can effectively define and update the background image in the field, reducing the misjudgment caused by changes in light and shadow, and selecting images containing moving objects with an optimal threshold. Our study has demonstrated its performance and generality using real data from different fields. For example, compared with the pre-trained Yolo model using the MS COCO dataset, the overall recall of FenceTalk increased from 72.36% to 98.39% for the model trained with the data picked out by SSIM. The recall of FenceTalk, combined with Yolo and SSIM, can reach more than 99%.

Electric energy prediction is an important issue and has been studied for many years. The prediction approaches have evolved from traditional statistical methods, conventional machine learning methods, deep learning (DL) methods, and then hybrid deep learning methods. This article proposes ElectricityTalk, an Internet of Things (IoT) platform for smart farms, which integrates the artificial intelligence (AI) mechanism with farming IoT devices for electric energy prediction and anomaly detection. The AI mechanism called AItalk is designed with modified convolution neural network (CNN) and long short-term memory models. Traditional electric energy prediction approaches only consider the information provided by smart meters. This article shows that with the extra IoT switch status information in the smart farm and postprocessing with a simple yet novel random walk model, the performance of ElectricityTalk is significantly improved (by 34.5%) as compared with the AI mechanism without the farming IoT switch information. We show that the mean absolute percentage error of AItalk is 8.62% (for the UCI dataset) and 1.53% (for the Bao farm dataset), which outperforms the previous solutions. We also show that ElectricityTalk detects all anomalies in real farm operations, and can achieve recall of 1 and precision larger than 0.994, which also outperforms the previous solutions. In particular, our mechanism can detect all anomalies in three minutes, which has not been reported in previous studies.

The deployment of a client–server-based distributed intelligent system involves application development in both the network domain and the device domain. In the network domain, an application server (typically in the cloud) is deployed to execute the network applications. In the device domain, several Internet of Things (IoT) devices may be configured as, for example, wireless sensor networks (WSNs), and interact with each other through the application server. Developing the network and the device applications are tedious tasks that are the major costs for building a distributed intelligent system. To resolve this issue, a low-code or no-code (LCNC) approach has been purposed to automate code generation. As traditional LCNC solutions are highly generic, they tend to generate excess code and instructions, which will lack efficiency in terms of storage and processing. Fortunately, optimization of automated code generation can be achieved for IoT by taking advantage of the IoT characteristics. An IoT-based distributed intelligent system consists of the device domain (IoT devices) and the network domain (IoT server). The software of an IoT device in the device domain consists of the Device Application (DA) and the Sensor Application (SA). Most IoT LCNC approaches provide code generation in the network domain. Very few approaches automatically generate the DA code. To our knowledge, no approach supports the SA code generation. In this paper, we propose DeviceTalk, an LCNC environment for the DA and the SA code development. DeviceTalk automatically generates the code for IoT devices to speed up the software development in the device domain for a distributed intelligent system. We propose the DeviceTalk architecture, design and implementation of the code generation mechanism for the IoT devices. Then, we show how a developer can use the DeviceTalk Graphical User Interface (GUI) to exercise LCNC development of the device software.