Most machine learning proposals in the Internet of Things (IoT) are designed and evaluated on preprocessed data sets, where data acquisition and cleaning steps are often considered a black box. In addition, IoT environments have numerous challenges related to acquiring data from sensors, where sensitive data can be threatened by malicious users who seek to interfere with the communication channel or storage. Additionally, sensor data can also be affected by noise. Therefore, differentiating the type of threat/anomaly requires additional energy and computational resources. We propose to carry out data cleaning/anomaly techniques on the IoT device itself, not in the cloud servers but closer to the data source. Therefore, the IoT device sends trusted data to the Cloud. Among the benefits of this is the considerable reduction in the cost of implementation due to less movement of data between IoT devices and the Cloud. Consequently, we define three anomaly detection steps using smoothing filters, unsupervised learning, and deep learning techniques (i.e., hybrid model) to detect different variations of anomalies and threats while focusing on a small computational/memory footprint. The deployment of the hybrid model on AVR, Tensilica, and ARM microcontrollers showed that the last ones are an adequate target to implement the model because they best satisfy the necessary hardware requirements. The proposed model consumes 50 kB of Flash and 12 kB of RAM and processes data locally, achieving a bandwidth reduction of 60%. Finally, the hybrid model was tested in external data sets.

We present an intelligent electronic glove system able to detect numbers of sign language in order to automate the process of communication between a deaf-mute person and others. This is done by translating the hands move sign language into an oral language. The system is inside to a glove with flex sensors in each finger that we are used to collect data that are analyzed through a methodology involving the following stages: (i) Data balancing with the Kennard-Stone (KS), (ii) Comparison of prototypes selection between CHC evolutionary Algorithm and Decremental Reduction Optimization Procedure 3 (DROP3) to define the best one. Subsequently, the K-Nearest Neighbors (kNN) as classifier (iii) is implemented. As a result, the amount of data reduced from stage (i) from storage within the system is 98%. Also, a classification performance of 85% is achieved with CHC evolutionary algorithm.

Nowadays, increasing air-pollution levels are a public health concern that affects all living beings, with the most polluting gases being present in urban environments. For this reason, this research presents portable Internet of Things (IoT) environmental monitoring devices that can be installed in vehicles and that send message queuing telemetry transport (MQTT) messages to a server, with a time series database allocated in edge computing. The visualization stage is performed in cloud computing to determine the city air-pollution concentration using three different labels: low, normal, and high. To determine the environmental conditions in Ibarra, Ecuador, a data analysis scheme is used with outlier detection and supervised classification stages. In terms of relevant results, the performance percentage of the IoT nodes used to infer air quality was greater than $90\%$. In addition, the memory consumption was 14 Kbytes in a flash and 3 Kbytes in a RAM, reducing the power consumption and bandwidth needed in traditional air-pollution measuring stations.

Agricultural production is on most countries’ national agenda because climate change affects crops, fruits, vegetables, and insect infestation. Therefore, achieving maximum production results is a challenge faced by professional growers, who have seen greenhouses as a very good option to guarantee these results. By using new technologies inside greenhouses, farmers can reduce the damaging effect of insects on plants and improve indoor cultivation through climate control. However, to efficiently manage agricultural fields and greenhouses today, farmers have to apply technologies in line Industry 4.0, such as: robots, Internet of Things devices, machine learning applications, and so on. In this context, deploying sensors plays a key role in collecting data and finding information supporting the farmer’s decision-making. As a feasible solution for small farms, this paper presents an autonomous robot that moves through greenhouse crop paths with previously-planned routes and can collect environmental data provided by a wireless sensor network, where the farmer does not have previous information about the crop. Here, an unsupervised learning algorithm is implemented to cluster the optimal, standard, and deficient sectors of a greenhouse to determine inappropriate growth patterns in crops. Finally, a user interface is designed to help farmers plan both the route and distance to be traveled by the robot while collecting information from the sensors to observe crop conditions.

Internet of Things (IoT) and edge devices have grown in their application fields due to machine learning (ML) models and their capacity to classify images into previously known labels, working close to the end-user. However, the model might be trained with several convolutional neural network (CNN) architectures that can affect its performance when developed in hardware-constrained environments, such as edge devices. In addition, new training trends suggest using transfer learning techniques to get an excellent feature extractor obtained from one domain and use it in a new domain, which has not enough images to train the whole model. In light of these trends, this work benchmarks the most representative CNN architectures on emerging edge devices, some of which have hardware accelerators. The ML models were trained and optimized using a small set of images obtained in IoT environments and using transfer learning. Our results show that unfreezing until the last 20 layers of the model’s architecture can be fine-tuned correctly to the new set of IoT images depending on the CNN architecture. Additionally, quantization is a suitable optimization technique to shrink $2\times $ or $3\times $ times the model leading to a lighter memory footprint, lower execution time, and battery consumption. Finally, the Coral Dev Board can boost $100\times $ the inference process, and the EfficientNet model architecture keeps the same classification accuracy even when the model is adopted to a hardware-constrained environment.

Car drivers under the influence of alcohol is one of the most common causes of road traffic accidents. To tackle this issue, an emerging, suitable alternative is the use of intelligent systems -traditionally based on either sensor networks or artificial vision- that are aimed to prevent starting the car when drunk status on the car driver is detected. In such vein, this paper introduces a system whose main objective is identifying a person having alcohol in the blood through supervised classification of sensor-generated and computer-vision-based data. To do so, some drunk-status criteria are considered, namely: the concentration of alcohol in the car environment, the facial temperature of the driver and the pupil width. Specifically, for data acquisition purposes, the proposed system incorporates a gas sensor, temperature sensor and a digital camera. Acquired data are analyzed into a two-stages machine learning system consisting of feature selection and supervised classification algorithms. Both acquisition and analysis stages are to be performed into a embedded system, and therefore all procedures and algorithms are designed to work at low-computational resources. As a remarkable outcome, due mainly to the incorporation of feature selection and relevance analysis stages, proposed approach reaches a classification performance of 98% while ensures adequate operation conditions for the embedded system.