This paper introduces an unsupervised defect detection system using a Convolutional Autoencoder (CAE) for brake caliper quality control in automotive manufacturing. A CAE is trained to reconstruct defect-free images, leveraging a Structural Similarity Index Measure (SSIM) loss to isolate anomalies between original and reconstructed images. Candidate defect regions are refined using Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering to distinguish true defects (e.g., deformities, scratches) from noise.Experiments conducted on a custom imaging workstation demonstrated strong performance in diverse regions of interest (ROIs) of the brake calipers. The system achieved F1-scores of 0.92, 0.95, and 0.76 on the logo, flat (2D), and non-flat (3D) ROIs, respectively. Data augmentation improved generalization, and clustering reduced false positives (FPs).By offering an alternative to traditional supervised methods, this CAE-based approach enables reliable defect detection, reducing manual dependence. Its flexible design supports integration into automotive production, leading to improved real-time monitoring and cost savings.

Topological learning is a wide research area aiming at uncovering the mutual spatial relationships between the elements of a set. Some of the most common and oldest approaches involve the use of unsupervised competitive neural networks. However, these methods are not based on gradient optimization which has been proven to provide striking results in feature extraction also in unsupervised learning. Unfortunately, by focusing mostly on algorithmic efficiency and accuracy, deep clustering techniques are composed of overly complex feature extractors, while using trivial algorithms in their top layer. The aim of this work is to present a novel comprehensive theory aspiring at bridging competitive learning with gradient-based learning, thus allowing the use of extremely powerful deep neural networks for feature extraction and projection combined with the remarkable flexibility and expressiveness of competitive learning. In this paper we fully demonstrate the theoretical equivalence of two novel gradient-based competitive layers. Preliminary experiments show how the dual approach, trained on the transpose of the input matrix i.e. XT, lead to faster convergence rate and higher training accuracy both in low and high-dimensional scenarios.

Despite all the progress made in biomedical field, the Electrocardiogram (ECG) is still one of the most commonly used signal in medical examinations. Over the years, the problem of ECG classification has been approached in many different ways, most of which rely on the extraction of features from the signal in the form of temporal or morphological characteristics. Although feature engineering can led to adequately good results, it mostly relies on human ability and experience in selecting the correct feature set. In the last decade, a growing class of techniques based on Convolutional Neural Network (CNN) has been proposed in opposition to feature engineering. The efficiency and accuracy of CNN-based approaches is indisputable, however their ability in extracting and using temporal features from raw signal is poorly understood. The main objective of this work was to uncover the differences and the relationships between CNN feature maps and human-curated temporal features, towards a deeper understanding of neural-based approaches for ECG. In fact, the proposed study succeeded in finding a similarity between the output stage of the first layers of a deep 1D-CNN with several temporal features, demonstrating that not only that the engineered features effectively works in ECG classification tasks, but also that CNN can improve those features by elaborating them towards an higher level of abstraction.

This paper presents a comparison between two recurrent neural networks (RNN) for arterial blood pressure (ABP) estimation. ABP is a parameter closely related to the cardiac activity, for this reason its monitoring implies decreasing the risk of heart disease. In order to predict the ABP values (both systolic and diastolic), electrocardiographic (ECG) and photoplethysmographic (PPG) signals are used, separately, as inputs of the networks. To train the artificial neural networks, the synchronized signals are extracted from the Physionet MIMIC database. The output-error Neural networks (NNOE) and the Long Short Term Memory (LSTM) architectures are compared in terms of RMSE and absolute error. NNOE neural network, with ECG signal as input, results the best configuration in terms of both the proposed metrics. The predicted ABP falls within the values of the normative ANSI/AAMI/ ISO 81060-2:2013 for sphygmomanometer certification.

Non-stationary topological representation can be addressed in two ways, according to the application: life-long modeling or by forgetting the past. Life-long learning requires neural networks equipped with a tool for judging if a neuron has to be created for tracking the input distribution. It is always implemented as an isotropic criterion (a hypersphere centered at the winner weight vector represents the domain of the neuron). Instead, the G-EXIN neural network, presented here, uses an anisotropic convex polytope, which, models the shape of the neuron neighborhood. This idea allows to consider the boundaries of the Voronoi sets of data and controls the extent of the extrapolation. It also employs a novel kind of edge, called bridge, which carries information on the extent of the distribution time change. Indeed, the analysis of bridges, mainly their density, yields a deeper insight to the kind of non-stationarity. Both artificial and real examples are given of the advantages of this approach with regard to the ESOINN neural network, which is the best existing approach to life-long modeling.

The electrocardiogram (ECG) is becoming a promising technology for biometric human identification. Usually ECG is used for health measurements and this is useful for biometric applications to state that the subject under analysis is live. But an individual identification shouldn't require a classical ECG clinical analysis where several contacts are applied to the person to be identified. In literature, ECG biometric recognition is usually studied for the recognition of a subject within a group of known subjects. In this paper, a new approach is considered. The aim of our embedded wearable controller is to authorize a subject or to reject him, labeling as an intruder unknown to the system. The study used 40 healthy subjects: two authorized and 38 intruders. A one-lead ECG trace has been recorded from the wrists of subjects, features have been extracted using a combination of Autocorrelation and Discrete Cosine Transform (AC/DCT) and then classified using a Multilayer Perceptron. Results show that intruder recognition can be performed with a success rate equal to 100%.

In the food industry, the importance of automatic detection and selection of raw food ingredients is increasing. In this paper, a method for real time automatic detection, segmentation and classification of hazelnuts using x-ray images is presented. Automatic extraction of independent nut images is made using image processing techniques. To extract meaningful features, moment invariants and texture properties are calculated on global level as well as from co-occurrence matrices. Principal component analysis is applied on features to achieve orthogonality in addition to dimensionality reduction. An anomaly detection algorithm is used for classification. Multivariate Gaussian distributions are calculated for model estimation using training data. Results are calculated on test data by using the threshold value obtained from best validation outcome. The classifier showed 98.6% correct classification rate for negative examples with 0% false negative rate.

Quality assessment and standardization of the property of the final product is fundamental in food industry. Coffee particle granulometry and density are continuously monitored during coffee beans grinding and grinders are controlled by operators in order to keep coffee particle granulometry within specific thresholds. In this work, a neural system is used to learn how to control two grinders used for coffee production at LAVAZZA factory, obtaining average control error of the order of a few μm. The results appear promising for the future development of an automatic decision support system.

The issue of air quality is now a major concern for many citizens worldwide. Local air quality forecasting can be made on the basis of meteorological variables and air pollutants concentration time series. We propose an adaptive filter technique based on an artificial neural network (ANN) to make 24-hours maximal daily ozone-concentrations forecasts.

The aim of our project is to develop a mobile real-time reader device for blind people. It uses Artificial Neural Networks (ANNs) for the core character recognition engine. The hardware constraints led us to develop a cross-platform framework to design and evaluate such subsystem in order to find a good trade off between run-time performances and accuracy of results.

The research activity described in this paper concerns the study of the phenomena responsible for the urban and suburban air pollution. The analysis carries on the work already developed by the NeMeFo (neural meteo forecasting) research project for meteorological data short-term forecasting. The study analyzed the air pollution principal causes and identified the best subset of features (meteorological data and air pollutants concentrations) for each air pollutant in order to predict its medium-term concentration (in particular for the particulate matter with an aerodynamic diameter of up to 10 mum called PM10). The selection of the best subset of features was implemented by means of a backward selection algorithm which is based on the information theory notion of relative entropy. The final aim of the research is the implementation of a prognostic tool able to reduce the risk for the air pollutants concentrations to be above the alarm thresholds fixed by the law. The implementation of this tool will be carried out using data-driven models based on some of the most wide-spread statistical data-learning techniques (artificial neural networks and support vector machines).

STIPER is an Italian national project whose aim is the study of devices to help blind people in daily activities. The core of the portable reader device is based on Artificial Neural Networks, used to recognize characters in real time beyond the fingers of blind people flowing on labels, restaurant menus and other printed objects. Neural Nets outputs are used to drive a Braille matrix, stimulating the fingertips of a blind person, and to speak by means of a common PDA device.

This paper presents a reliable ice detection and forecasting system based on a special road weather information system and a data processing algorithm for ice event forecasting. The ice detection system is mainly based on two innovative sensors, developed ad hoc for this particular application; namely an ice sensor for the reliable detection of accumulation of ice or water on the sensor and a salinity sensor for the purpose of freezing point calculations. These data are then used by a statistical data processing algorithm or a neural network for ice event forecasting.

Features selection to assess the accuracy of a protein three-dimensional model, when only the protein sequence is known, is a challenging task because it is not clear which features are most important and how they should best be combined. We present the results of an information theory-based approach to select an optimal subset of features for the prediction of protein model quality. The optimal subset of features was calculated by means of a backward selection procedure, starting from a set of structural features belonging to the following three categories: atomic interactions, solvent accessibility, and secondary structure. Three statistical-learning approaches were evaluated to predict the quality of a protein model starting from an optimum subset of features. The performances of a probabilistic classifier modeled by means of a kernel probability density estimation method (KPDE) were compared with those of a feed-forward artificial neural network (ANN) and a support vector machine (SVM).

In this paper a committee of "intelligent systems" evaluates the occurrence of meteorological phenomena. Rain and fog are the events which are considered. The forecast system is based on a multinetwork approach which evaluates data coming from electronic sensors and from satellite observations. More data and more engines are used to increase the reliability of the event prediction. The increased complexity of the global system requires more data coming from different sources but gives a good reliability.

Nowadays the most powerful tool to evaluate complex production processes' performance is simulation software. This approach is computationally slow. So one alternative could be an approximation of the original system's model able to detect the relationships between input and output, yet it's computationally more efficient than simulation. During training, a neural network is presented with several input/output pairs, and it learns the functional relationship between input and outputs of the simulation model. The network can guess the output for inputs other than the ones presented during training. Usually it's unknown what are the most significant variables for the output of a system. For this problem, it can be used as a data mining preprocess analysis to see the influence of each input parameter on the performance of the production process. In this paper it is shown a statistical nonparametric method, based on 8Parzen window, to make a data mining analysis and then a multilayer perceptron to approximate the original system.

Artificial neural networks are extensively applied to several applications where data-driven methods are requested. This work describes a neural architecture, which controls an "inverted pendulum" in a very flexible manner with a reusability perspective. The project was implemented through a "digital core" constituted of a FPGA, a microcontroller and an SRAM block, which co-operate to the neural computation. The FPGA was programmed in VHDL to implement the neural architecture. The core was written in a recursive manner to permit the reconfigurability of the network and its reusability to all the systems, which can be modelled through a similar neural network. Through these parameters the system combines the configurability (typical of a sw project) with the velocity guaranteed by the hw implementation of the mathematical algorithms. Experimental results validated the effectiveness of the proposed approach; the network was able to balance a mechanical inverted pendulum above the middle of the slide guides.

We developed a real-time wideband speech codec adopting a wavelet packet based methodology. The transform domain coefficients were first quantized by using a psycho-acoustic model and then encoded with an arithmetic coding. The arithmetic coding was carried out by adapting the probability model of the quantized coefficients frame by frame by means of a competitive neural network, which was trained to detect regularities in the distribution of the wavelet packet coefficients. The weight matrix of the neural network is periodically updated during the compression in order to model better the speech characteristics of the current speakers. The coding/decoding algorithm was first written in C and then optimized on the TMS320C6000 DSP platform in a QoS-compliant fashion.

Pharmacokinetics (PK) studies the absorption and the elimination of drugs in the human body. Most PK models are parametric. Once it is identified, a PK model can estimate e.g. plasma concentrations over time for a given individual. Population PK aims at making sensible predictions even when measurements are so scarce to prevent model identification. In fact those predictions integrate past experimental evidence concerning some reference population. Again, integration typically relies on a (meta)-parametric model of the distribution of individual PK parameters in the population. In this paper we propose a fully data-driven, model-independent, non-parametric approach to population PK that makes use of Probabilistic Neural Networks (PNNs). The method incorporates regularized optimization of the kernels' width and is able to deal with missing data. It was validated on a problem that is refractory to traditional PK modeling techniques. We carried out a clinical study where a single oral dose of racemic ibuprofen was administered to 18 healthy volunteers. In each subject plasma concentrations of the R- and S-ibuprofen enantiomers were measured at predefined times after administration, together with some co-variates (e.g. age, body mass). Estimates on each single individual were obtained by taking the remaining ones as the reference population. Results show the ability to infer the whole concentration profile of either enantiomer from few concentrations of the other one. In conclusion, our approach does not require any explicit model to take advantage of any statistical dependencies among (perhaps heterogeneous) quantities

We present an interactive educational DSP platform for realtime subband audio coding. The tool integrates a DSP course on advanced audio coding and focuses on wavelets, subband coding, and signal quantization. The package consists of a Windows control application running on a PC and real-time DSP code implemented on a Texas Instruments DSP Starter Kit (DSK), which performs the actual audio encoding and decoding process. The tool is interactive since students can customize the codec by changing some encoding parameters through the PC interface and then verify the effect on the coding process. The DSP codec accepts audio samples in real-time via the DSK audio analog input port and sends the reconstructed samples to the output port. The graphical user interface allows to choose the type of filter-bank and scalar quantization; the availability of the full source code and of a DSK makes possible to experience first-hand the main aspects of real-world DSP development, e.g., real-time programming, arithmetic precision and algorithmic delay. The software is available at http://multimedia.polito.it/dspwavelab.

In this paper we show how the familiar concept of gradient descent can be extended in presence of binary neurons. The procedure we devised formally operates on generic feedforward networks of logistic-like neurons whose activations are re-scaled by an arbitrarily large gauge. Whereas the gradient decays exponentially with increasing values of the gauge, the sign of each component becomes definitely equal to a constant value. Those values are actually computed by means of a "twin" network of binary neurons. This allows the application of any "Manhattan" training algorithm such as resilient propagation.

Analog artificial neural networks (ANN) could be the core of an "intelligent" signal processor, with the today used digital processing replaced by a raw "data driven" methodology. Characteristic of this approach is: analog input/output: signals don't need A/D and D/A converters; speed: an analog system can be faster than a digital or a mixed-mode one; effectiveness: ANN already demonstrated their power in many applications; low power: analog circuits can save power. NESP2 is a neural signal processor which offers the above characteristics and is based on a traditional, available and not expensive double polysilicon double metal commercial VLSI process.

A keypoint of the control of a power system is the forecast of the short term load. The paper presents a dynamic model for short term load forecasting (STLF) which uses a recurrent neural network. This network can be used to build empirical models for the load of a dynamic system. We investigate this problem applying a basic neural network with feedback connections which is unfolded in time and becomes a general feedforward network with weights sharing. The main advantage of this model consists in unfolding the network in time, which becomes a non fully connected feedforward network and facilitates the training stage. At the same time our model provides a one day ahead prediction.

The authors briefly summarize the main lines of the convergent and chaotic-bifurcative approaches in neural networks, and present a general model founded on an informational use of a chaotic dynamics. It exploits the inner fine structure of unstable periodic orbits of a chaotic dynamics to perform invariant extractions and reconstruction tasks in a dynamic way from a complex time-varying (at least chaotic) input. The neurophysiological background (i.e. synchronization behavior and functional segregation in the sensory cortex) is discussed. The proposed approach suggests that there exists a strict relationship in chaotic systems between dynamic reconstruction, optimization, and stabilization intended as a relaxation process in as much as they are all functions of an inner self-correlation process. This may depend on the fact that chaos, owing to its ultimate deterministic nature, is an intelligent noise. In the fine structure of its invariants, it retains a memory of its evolution