In this study, we compare different machine learning approaches applied to acoustic resonance recognition of coins. Euro-cents and Euro-coins were classified by the sound emerging when throwing the coins onto a hard surface.The used dataset is a representative example of a small data which was collected in carefully prepared experiments.Due to the small number of coin specimens and the count of the collected observations, it was interesting to see whether deep learning methods can achieve similarly or maybe even better classification performances compared with more traditional methods.The results of the multi-class prediction of coin denominations are presented and compared in terms of balanced accuracy and Matthews Correlation Coefficient metrics. The feature analysis methods combined with the employed classifiers achieved acceptable results, despite the relatively small dataset.
In this paper, acoustic resonance testing on glass intravenous (IV) bottles is presented. Different machine learning methods were applied to distinguish acoustic observations of bottles with defects from the intact ones. Due to the very limited amount of available specimens, the question arises whether the deep learning methods can achieve similar or even better detection performance compared with traditional methods.
We present an approach for ultrasonic non-destructive testing of adhesive bonding employing unsupervised machine learning with autoencoders.The models are trained exclusively on the features derived from pulse-echo ultrasonic signals on a specimen with good adhesive bonding and tested on another specimen with artificially added defects.The resulting pseudo-probabilities indicating anomalies are visualized and presented along to the C-scan of the same specimen. As a result, we achieved improved representation of the defects, allowing their automatic and reliable detection.
The renaissance of artificial intelligence (AI) in the last decade can be credited to several factors, but chief among these is the ever-increasing availability and miniaturization of computational resources. This process has contributed to the rise of ubiquitous computing via popularizing smart devices and the Internet of Things in everyday life. In turn, this has resulted in the generation of increasingly enormous amounts of data. The tech giants are harvesting and storing data on their clients’ behavior and, at the same time, introducing concerns about data privacy and protection. Suddenly, such an abundance of data and computing power, which was unimaginable a few decades ago, has caused a revival of old and the invention of new machine learning paradigms, like Deep Learning. Artificial intelligence has undergone a technological breakthrough in various fields, achieving better than human performance in many areas (such as vision, board games etc.). More complex tasks require more sophisticated algorithms that need more and more data. It has often been said that data is becoming a resource that is "more valuable than oil"; however, not all data is equally available and obtainable. Big data can be described by using the "four Vs"; data with immense velocity, volume, variety, and low veracity. In contrast, small data do not possess any of those qualities; they are limited in size and nature and are observed or produced in a controlled manner. Big data, along with powerful computing and storage resources, allow “black box” AI algorithms for various problems previously deemed unsolvable. One could create AI applications even without the underlying expert knowledge, assuming there are enough data and the right tools available (e.g. end-to-end speech recognition and generation, image and object recognition). There are numerous fields in science, industry and everyday life where AI has vast potential. However, due to the lack of big data, application is not straightforward or even possible. A good example is AI in medicine, where an AI system is intended to assist physicians in diagnosing and treating rare or previously never observed conditions, and there is no or an insufficient amount of data for reliable AI deployment. Both big and small data concepts have limitations and prospects for different fields of application. This paper will try to identify and present them by giving real-world examples in various AI fields.
In this paper, we present a preliminary study for acoustic modeling in Upper Sorbian, where a model of German was used in cross-lingual transfer learning. At first, we define the grapheme and phoneme inventories and map the target phonemes from the most similar German source equivalents. Phonetically balanced sentences for the recording prompts were selected from a combination of general and domain-specific textual data. The speech corpora with a total duration of around 11 hours was collected in controlled recording sessions involving an equal number of females, males, and children. The baseline acoustic model was employed to force-align the speech corpora given the knowledge-based phoneme mappings. How well the mappings were, was evaluated by the phoneme confusions in free-phoneme recognition. The new derived data-driven model with a reduced phoneme set was included in the adaptation and evaluation along with the baseline acoustic model. The model adaptation performance was cross-validated with the “Leave One Group Out” strategy. We observed major improvements in phoneme error rates after adaptation for the knowledge-based and data-driven phoneme mappings. The study confirmed the feasibility of transfer learning for acoustic model adaptation in the case of Upper Sorbian, at the same time demonstrating practical usability with a small vocabulary speech recognition application (Smart Lamp).
Based on the fact that cogwheels are indispensable parts in manufacturing, we present the acoustic resonance testing (ART) of small data on sintered cogwheels for quality control in the context of non-destructive testing (NDT). Considering the lack of extensive studies on cogwheel data by means of ART in combination with machine learning (ML), we utilize time-frequency domain feature analysis and apply ML algorithms to the obtained feature sets in order to detect damaged samples in two ways: one-class and binary classification. In each case, despite small data, our approach delivers robust performance: All damaged test samples reflecting real-world scenarios are recognized in two one-class classifiers (also called detectors), and one intact test sample is misclassified in binary ones. This shows the usefulness of ML and time-frequency domain feature analysis in ART on a sintered cogwheel dataset.
We investigated and compared various algorithms in machine learning for anomaly assessment with different feature analyses on ultrasonic signals recorded by sensor networks. The following methods were used and compared in anomaly detection modeling: hidden Markov models (HMM), support vector machines (SVM), isolation forest (IF), and reconstruction autoencoders (AEC). They were trained exclusively on sensor signals of the intact state of structures commonly used in various industries, like aerospace and automotive. The signals obtained on artificially introduced damage states were used for performance evaluation. Anomaly assessment was evaluated and compared using various classifiers and feature analysis methods. We introduced novel methodologies for two processes. The first was the dataset preparation with anomalies. The second was the detection and damage severity assessment utilizing the intact object state exclusively. The experiments proved that robust anomaly detection is practically feasible. We were able to train accurate classifiers which had a considerable safety margin. Precise quantitative analysis of damage severity will also be possible when calibration data become available during exploitation or by using expert knowledge.
Using unified automatic speech recognition and synthesis system for cross-language acoustic modeling
(2012)
Artificial intelligence experienced a technological breakthrough in science, industry, and everyday life in the recent few decades. The advancements can be credited to the ever-increasing availability and miniaturization of computational resources that resulted in exponential data growth. However, because of the insufficient amount of data in some cases, employing machine learning in solving complex tasks is not straightforward or even possible. As a result, machine learning with small data experiences rising importance in data science and application in several fields. The authors focus on interpreting the general term of "small data" and their engineering and industrial application role. They give a brief overview of the most important industrial applications of machine learning and small data. Small data is defined in terms of various characteristics compared to big data, and a machine learning formalism was introduced. Five critical challenges of machine learning with small data in industrial applications are presented: unlabeled data, imbalanced data, missing data, insufficient data, and rare events. Based on those definitions, an overview of the considerations in domain representation and data acquisition is given along with a taxonomy of machine learning approaches in the context of small data.
In this paper we show how word class based language modeling can support the integration of a small language in modern applications of speech technology. The methods described in this paper can be applied for any language. We demonstrate the methods on Upper Sorbian. The word classes model the semantic expressions of numerals, date and time of day. The implementation of the created grammars was realized in the form of finite-state-transducers (FSTs) and minimalists grammars (MGs). We practically demonstrate the usage of the FSTs in a simple smart-home speech application, that is able to set wake-up alarms and appointments expressed in a variety of spontaneous and natural sentences. While the created MGs are not integrated in an application for practical use yet, they provide evidence that MGs could potentially work more efficient than FSTs in built-on applications. In particular, MGs can work with a significantly smaller lexicon size, since their more complex structure lets them generate more expressions with less items, while still avoiding wrong expressions.
We present a data-driven approach for the quantitative analysis of glottal stops before word-initial vowels in Upper Sorbian, a West Slavic minority language spoken in Germany. Glottal stops are word-boundary markers and their detection can improve the performance of automatic speech recognition and speech synthesis systems.
We employed cross-language transfer using an acoustic model in German to develop a forced-alignment method for the phonetic segmentation of a read-speech corpus in Upper Sorbian. The missing phonemic units were created by combining the existing phoneme models. In the forced-alignment procedure, the glottal stops were considered optional in front of word-initial vowels.
To investigate the influence of speaker type (males, females, and children) and vowel on the occurrence of glottal stops, binomial regression analysis with a generalized linear mixed model was performed. Results show that children glottalize word-initial vowels more frequently than adults, and that glottal stop occurrences are influenced by vowel quality.
We present a method for extracting health indicators from piezoelectric sensors applied in the case of microfluidic valves. Convolutional autoencoders were used to train a model on the normal operating conditions and tested on signals of different valves. The results of the model performance evaluation, as well as, the qualitative presentation of the indicator plots for each tested component, showed that the used approach is capable of detecting features that correspond to increasing component degradation. The extracted health indicators are the prerequisite and input for reliable remaining useful life prediction.
Künstliche Intelligenz zieht derzeit in alle Bereiche der Gesellschaft und des Lebens ein. Aber welchen Stellenwert hat sie momentan auf dem Gebiet der zerstörungsfreien Prüfung? Was kann KI leisten? Welche Herausforderungen müssen erfolgreich bewältigt werden? Gibt es das eine KI-Verfahren, welches prinzipiell für ZfP geeignet ist? Bei der Bauteil- und Materialprüfung während und unmittelbar nach der Herstellung, der Überwachung von Verschleißteilen in Maschinen und Anlagen oder der Schadensdetektion an Bauteilen und Komponenten liefern ZfP-Verfahren Daten, die bewertet werden müssen. Obwohl inzwischen sehr leistungsfähige Toolkits verfügbar sind, erfordert der optimale Einsatz der KI für ein ZfP-Verfahren oftmals mehr. Die meisten Kunden möchten nicht nur eine Lösung ihres Problems; sie wollen verstehen, warum die KI so und nicht anders entschieden hat, warum der Klassifikator das Bauteil einer bestimmten Klasse (z. B. gut/schlecht oder neuwertig/verschlissen/defekt) zugewiesen hat. Abhängig von der Klassifikationsaufgabe sowie der Art und der Anzahl der vorliegenden Daten kann ein geeignetes Verfahren bestimmt werden. Mit Methoden des maschinellen Lernens werden Modelle gebildet, welche die Basis für die KI-Verfahren zur Klassifikation bilden.
Der Beitrag liefert einen Überblick über KI-Verfahren und deren Anwendungen in der zerstörungsfreien Prüfung. Zahlreiche Beispiele und Ergebnisse werden vorgestellt, um die Mannigfaltigkeit des Einsatzes in der ZfP und der bestehenden Möglichkeiten zu demonstrieren.
Softness is one of the essential properties of hygiene tissue products. Reliably measuring it is of utmost importance to ensure the balance between customer expectations and cost-effective tissue production. This study presents a method for assessing softness by analyzing acoustic emissions produced while tearing a tissue specimen. The aim was to train neural network models using the corrected results of human panel tests as the ground truth labels and to predict the tissue softness in two- and three-class recognition tasks. We also investigate the possibility of predicting some production parameters related to the softness property. The results proved that tissue softness and production parameters could be reliably estimated only by the tearing noise.