Soft sensors are mathematical models of industrial processes that are often used for monitoring and control. Data-driven techniques based on artificial intelligence are generally used for identifying these models. Data scarcity is a challenging problem that occurs when the variable to be estimated must be measured by laboratory analysis. Here, a difference-based neural network called Δ -Net is proposed to develop dynamic nonlinear soft sensors when only a few hundred labeled data are available. The Δ -Net, which exploits the case difference heuristic approach, is based on two parallel neural networks responsible for processing pairs of input samples. The Δ -Net is trained on an augmented dataset obtained by pairwise ordering the small original dataset. The outputs of the subnets are then combined to estimate the difference between the corresponding outputs of the pairs. The reconstruction of the output sample is computed as the mean of the distances from the sample to a set of anchor points, to increase the robustness of the prediction. The proposed approach has been applied to well-known industrial benchmarking datasets. The obtained results show superior performance compared to other data augmentation approaches, including bootstrap resampling, variational autoencoders and Wasserstein generative adversarial networks.

Spiking Neural Networks (SNNs) represent a novel class of models that are becoming increasingly popular. Thanks to their unique way to process discrete and asynchronous signals, these type of networks demonstrated to be computationally and energetically more efficient than traditional Deep Neural Networks (DNNs), making them suitable for devices with limited hardware capabilities. Moreover, their ability to continuously adapt to external inputs changes enables an online evolution of network weights which is ideal for real-time applications scenarios. As a downside, one of the main challenges when working with SNNs consists in the impossibility of using traditional learning rules such as backpropagation (or its variations) for the online training of these architectures. In such a context emerges the need to define online and biologically plausible training solutions that would allow the hardware implementation of these networks on neuromorphic devices. To overcome this issue, in this paper we propose a learning rule for the training of vanilla Spiking Autoencoders (Spiking-AEs) which exploits synaptic trace activities to enable an online weights adaptation. We also propose a novel modular tied Spiking-AE to address reconstruction and anomaly detection tasks. In this sense, the tied connections make the training less complex, drastically reducing the number of trainable parameters and model memory footprint. Experimental results conducted on the MNIST and MFPT datasets demonstrate the effectiveness of the proposed solution in terms of reconstruction and anomaly detection capabilities, respectively.

In the field of robotics, ensuring reliable and efficient performance is crucial, especially when robots are entrusted with critical tasks. Anomaly detection systems play a crucial role in maintaining this reliability by detecting deviations from normal behavior and taking timely interventions. Traditional model-and knowledge-based approaches, while effective in controlled environments, reach their limits in dynamic and resource-constrained settings due to their reliance on predefined models, expert knowledge and high computational requirements. While data-driven methods, especially those using deep learning, offer better adaptability, they also bring challenges in terms of energy consumption and hardware limitations. To address these issues, this paper proposes a hybrid anomaly detection system that utilizes Spiking Neural Networks (SNNs) and Convolutional Neural Networks (CNNs). The SNN provides low-power processing of input features for anomaly detection, while the CNN efficiently extracts spatial features from high-resolution images to realize an accurate, high-performance classifier that can be used to perform an online fine-tuning of the SNN. A large dataset for robot navigation is used as a testbed. The obtained results show that this hybrid approach increases the speed and accuracy of anomaly detection while significantly reducing energy consumption, making it well-suited for applications in resource-constrained robotic systems.

Anomaly detection systems are crucial for identifying deviations from normal behavior in various domains, ranging from cybersecurity to industrial system monitoring. Traditional approaches often rely on predefined rules or supervised learning algorithms, which may struggle to adapt to complex and evolving patterns. As a result, deep learning techniques have been employed to address these challenges, but they usually require significant computational resources and memory for training and inference. On the contrary, Spiking Neural Networks (SNNs), inspired by the brain neural architecture, excel in effectively capturing complex temporal patterns present in real-world data through unsupervised learning, potentially leading to more robust performance while reducing energy consumption. In this paper, we employ the homeostatic plasticity characteristic that governs the synapses behaviour in SNN to implement an anomaly detection system designed for processing raw vibration data. Initially, we generate spike trains to represent the input data. These are then fed into a Balanced Spiking Neural Network (BSNN) to detect suspicious anomalies. The proposed SNN learning algorithm dynamically adjusts synaptic weights based on the precise timing of spikes between neurons, and introduces a combination of Spike Timing Dependent Plasticity (STDP) and reverse STDP techniques. Experimental results conducted on a public available vibration dataset demonstrate the effectiveness of the proposed approach in accurately identifying anomalies while maintaining low false positive rates.

In the evolving landscape of healthcare, Machine Learning (ML) and Deep Learning (DL) have revolutionized medical practises, particularly in Smart Health. This paper explores their application in Intelligent Cyber Physical Systems (iCPSs), focusing on surface Electromyography (sEMG) signal classification for prosthetic control. We conduct a detailed quan-titative study, comparing the performances and memory efficiency of Recurrent Neural Networks (RNNs) with Long Short-Term Memory (LSTM) cells and Convolutional Neural Networks (CNNs) with Temporal Convolutional Networks (TCNs) layers. Using the most popular quantization frameworks, TensorFlow Lite (TFLite) and Quantized Keras (QKeras), we quantify the impact on the proposed models of both INT8 Quantization and Deep Quantization. The aim is to undertake a comparison among the proposed structures and prove that their Deeply Quantized version lead to minimal performance degradation related to memory footprint's substantial reduction. Our contributions include a custom TCN layer, compliant with QKeras deep quantization, a comparative analysis between LSTM and TCN and evaluation of quantization frameworks. We offer insights into model performance and efficiency, informing their potential in real-world medical contexts.

Deliver intelligence into low-cost hardware e.g., Microcontroller Units (MCUs) for the realization of low-power tailored applications nowadays is an emerging research area. However, the training of deep learning models on embedded systems is still challenging mainly due to their low amount of memory, available energy, and computing power which significantly limit the complexity of the tasks that can be executed, thus making impossible use of traditional training algorithms such as backpropagation (BP). During these years techniques such as weights compression and quantization have emerged as solutions, but they only address the inference phase. Forward-Forward (FF) is a novel training algorithm that has been recently proposed as a possible alternative to BP when the available resources are limited. This is achieved by training the layers of a neural network separately, thus reducing the required energy and memory. In this paper, we propose µ-FF, a variation of the original FF which tackles the training process with a multivariate Ridge regression approach and allows to find closed-form solution by using the Mean Squared Error (MSE) as loss function. Such an approach does not use BP and does not need to compute gradients, thus saving memory and computing resources to enable the on-device training directly on MCUs of the STM32 family. Experimental results conducted on the Fashion-MNIST dataset demonstrate the effectiveness of the proposed approach in terms of memory and accuracy.

The industrial Internet of Things technology had a very strong impact on the realization of smart frameworks for detecting anomalous behaviors that could be potentially dangerous to a system. In this regard, most of the existing solutions involve the use of artificial intelligence models running on edge devices, such as intelligent cyber physical systems typically equipped with sensing and actuating capabilities. However, the hardware restrictions of these devices make the implementation of an effective anomaly detection algorithm quite challenging. Considering an industrial scenario, where signals in the form of multivariate time-series should be analyzed to perform a diagnosis, echo state networks (ESNs) are a valid solution to bring the power of neural networks into low complexity models meeting the resource constraints. On the other hand, the use of such a technique has some limitations when applied in unsupervised contexts. In this article, we propose a novel model that combines ESNs and autoencoders (ESN-AE) for the detection of anomalies in industrial systems. Unlike the ESN-AE models presented in the literature, our approach decouples the encoding and decoding steps and allows the optimization of both the processes while performing the dimensionality reduction. Experiments demonstrate that our solution outperforms other machine learning approaches and techniques we found in the literature resulting also in the best tradeoff in terms of memory footprint and inference time.

Vehicular Social Networks (VSNs) is an emerging communication paradigm, derived by merging the concepts of Online Social Networks (OSNs) and Vehicular Ad-hoc Networks (VANETs). Due to the lack of robust authentication mechanisms, social-based vehicular applications are vulnerable to numerous attacks including the generation of sybil entities in the networks. We address this important issue in vehicular crowdsourcing campaigns where sybils are usually employed to increase their influence and worsen the functioning of the system. In particular, we propose a novel User Recruitment Policy (URP) that, after extracting the participants within the event radius of a crowdsourcing campaign, detects and filters out the sybil vehicles by using a novel sybil detection approach, called SybilDriver. This technique combines the advantages of VANETs and OSNs by means of an innovative concept of proximity graph obtained from the physical vehicular network, in conjunction with a community detection and Random Forest techniques adopted in the OSN domain. Detailed experimental evaluations demonstrate the effectiveness of our approach and also show that it outperforms existing state-of-the-art methods typically used in the OSNs.1

The Industry 4.0 paradigm has changed the way industrial systems with hundreds of sensor-actuator enabled devices, including industrial internet of things (IIoT), cooperate and communicate with the physical and human worlds. Given the intricacy, the diagnostics of such systems is extremely important. While anomaly detection is a valid approach to avoid unplanned maintenance or even complete breakdown, its effective realization in IIoT requires the design and implementation of frameworks for efficient monitoring, data collection, and analysis. Most of the existing anomaly detection techniques provide only a diagnosis of the fault without taking into account the uncertainty. Moreover, the lack of ground truth data (which is a typical problem in the industrial context), make their implementation even more challenging. This paper proposes an anomaly detection technique built on top of an industrial framework for the data collection and monitoring. Specifically, we address the lack of labeled data by designing a semi-supervised anomaly detection algorithm that exploits Bayesian Gaussian Mixtures to assess the working condition of the plant while measuring the uncertainty during the diagnosis process and we implement the proposed framework on a real-life IIoT testbed, namely a scale replica assembly plant. Experimental results demonstrate that our anomaly detection algorithm is able to detect the plant working conditions with 99.8% of accuracy, and the semi-supervised approach performs better than a supervised one.

With the advent of smart environments the requirements for diagnostics and prognostics techniques gained a lot of interest. Lately, the Industry 4.0 paradigm is starting to expand also in the smart road, referred to in this context as "maintenance 4.0". In such a context, the possibility to predict the conditions of a road network allows to deliver a preventive maintenance that can strongly reduce the costs and avoid severe consequences. However, considering the actual pavement management state of the art, it is evident the huge amount of heterogeneous data necessary to perform this challenging tasks. Leveraging the Cloud and Edge technologies, this paper proposes a web Geographical Information System (GIS) platform capable to dynamically collect and analyze several type of sensor data for the management of road pavements. On top of that, the platform is able to compute advanced indexes and metrics that are used to produce a maintenance proposal scheme. Experimental results present a preliminary case study on a real motorway and demonstrate the effectiveness of the proposed platform as a support tool during the maintenance process.

Diagnosis of plant health conditions is gaining significant attention in smart agriculture. Timely recognition of early symptoms of a disease can help avoid the spread of epidemics on the plantations. In this regard, most of the existing solutions use some AI techniques on smart edge devices (IoTs or intelligent Cyber Physical Systems), typically equipped with a hardware like sensors and actuators. However, the resource constraints on such devices like energy (power), memory and computation capability, make the execution of complex operations and AI algorithms (neural network models) for disease detection quite challenging. To this end, compression and quantization techniques offer viable solutions to reduce the memory footprint of neural networks while maximizing performance on the constrained devices. In this paper, we realized a real intelligent CPS on top of which we implemented an AI application, called Deep Leaf running on a microcontroller of the STM32 family, to detect coffee plant diseases with the help of a Quantized Convolutional Neural Network (Q-CNN) model. We present a quantitative analysis of Deep Leaf by comparing five different deep learning models: a 32-bit floating point model, a compressed model, and three different types of quantized models exhibiting differences in terms of accuracy, memory utilization, average inference time, and energy consumption. Experimental results show that the proposed Deep Leaf detector is able to correctly classify the plant health condition with an accuracy of 96%, thus demonstrating the feasibility of our approach on a Smart Edge platform.

Aspects related to the maintenance scheduling have become a crucial problem especially in those sectors where the fault of a component can compromise the operation of the entire system, or the life of a human being. Current systems have the ability to warn only when the failure has occurred causing, in the worst case, an offline period that can cost a lot in terms of money, time, and security. Recently, new ways to address the problem have been proposed thanks to the support of machine learning techniques, with the aim to predict the Remaining Useful Life (RUL) of a system by correlating the data coming from a set of sensors attached to several components. In this paper, we present a machine learning approach by using LSTM networks in order to demonstrate that they can be considered a feasible technique to analyze the "history" of a system in order to predict the RUL. Moreover, we propose a technique for the tuning of LSTM networks hyperparameters. In order to train the models, we used a dataset provided by NASA containing a set of sensors measurements of jet engines. Finally, we show the results and make comparisons with other machine learning techniques and models we found in the literature.

Nowadays, smart environments are becoming an integral part of our everyday lives. Objects are becoming smarter and the number of applications where they are involved increases day by day. In such a context, indoor localization is a key aspect for the development of smart services which are strictly related to the user position inside an environment. In this paper, we present a deep learning approach to estimate the indoor user location starting from its Wi-Fi fingerprint composed by those signals perceived in the environment. We show some experimental results that demonstrate the feasibility of the proposed approach.