Reliability, punctuality and efficiency are some of the distinctive characteristics of successful businesses which, in order to maintain high standards, adopt advanced techniques for quality management. Using AI allows you to have more control over processes and anticipate critical issues, with machine learning you can then tackle complex tasks, such as image recognition and the analysis of large amounts of data
Quality management is a central aspect in the context of growing competitiveness and becomes a pivotal point of challenge and opportunity in the management and organization of the entire physical and information flow of business processes. Quality management is considered an important strategic management tool that involves the application of quality principles and practices at all organizational levels. A philosophy aimed at quality entails a management of production and processes aimed at guaranteeing high standards; in fact, customers are increasingly demanding and looking for products and services in line with their expectations and this decisively influences the customer satisfaction and corporate image; not only that, better quality management also improves internal company performance, ensuring greater safety and better execution of activities.
Quality management focuses on a broad and varied set of activities ranging from supply, therefore from the selection of suppliers to the evaluation of raw materials entering the warehouses, to delivery to the final consumer and to the after-sales assistance. Furthermore, quality management is a continuous process of small improvements in search of the level of excellence to achieve efficiency, sustainability and competitiveness.
Quality as a competitive advantage
The process of globalization of economies has led to a significant increase in competitiveness between organizations. Michael Porter, father of strategic marketing, has identified three different competitive advantages, that is, three characteristics of the product or service that the company must keep under control so that customers do not pass over to the competition. The three types of competitive advantage can be summarised as follows:
- Cost: the corporate cost structure does not burden the company, which can therefore offer its products and services at competitive prices or implement discount and promotion strategies;
- Differentiation: products or services have distinctive characteristics for which they are preferred by customers, such as technological or image characteristics;
- Niche: the company’s market segment is specific and limited to a limited number of customers.
The introduction or updating of aquality management system must be targeted and customised according to the company and its core business. In view of a transition process towards a quality-oriented philosophy, it is therefore advisable to always keep in mind the critical success factors of the company, in other words, the relevant aspects that support competitive advantage, and the systems for measuring performance and monitoring these factors; having critical factors and performance under control makes it possible to evaluate and direct the improvement choices of quality management systems without wasting resources and energy.
Machine learning is a branch of artificial intelligence that collects methods developed by different scientific sectors, such as statistics, physics and computer science, which are used to write programs for machine learning of rules and patterns without the definition of explicit instructions. The learning process, as defined by Tom Mitchell, is based on three elements: the task to be learned, the experience and performance in solving the task. A system is able to learn if performance increases as the experience available increases. In general, there are three categories of learning tasks:
- Supervised: the system has available an input data and the annotation of the data; for example, in the fabric industry it is possible to have images of fabrics in excellent condition and images of damaged fabrics and teach a recognition system to automatically identify and report parts of damaged fabric;
- Unsupervised: the system has only the data available, but has no information on the annotation of the data; often these techniques are applied in the identification of anomalies and in contexts where it is not possible to annotate the data;
- With reinforcement: the system has rewards or penalties available depending on the behavior and tries to implement actions to earn more rewards and not incur penalties; this type of techniques is typically adopted in dynamic, uncertain and competitive contexts, such as games and robots;
Machine learning provides new tools to tackle complex tasks, such as image recognition and the analysis of large amounts of data, and for this reason, it is having a significant impact on the quality management systems of companies. A commonly used approach is the use of a supervised classifier, such as: K-Nearest Neighbor, Support Vector Machines, Naive Bayes Classifier, Decision Tree Classifier, Adaboost and neural networks. While preferable and capable of achieving comparable, sometimes superior, performance of human abilities, supervised learning techniques are not always available options. In fact, the collection of properly annotated data is a process that requires time, energy and resources; in some cases it may not even be possible to annotate the data and it is necessary to resort to techniques of data augmentation or generation of synthetic data. In these cases, unsupervised techniques are used, such as: clustering algorithms, K-Means and the like, One-Class Support Vector Machines and neural networks of the Autoencoder and Generative Adversarial Networks type.
Some examples of machine learning adoption areas for quality management
Companies and the scientific community have dedicated attention and resources to the study and development of applications of machine learning techniques in the field of quality management.
Ming et al (2020) presents a review of the literature in the field of defect detection in 3C glasses always used in theComputer, Communication and Consumer Electronics industry. Glass defects, such as scratches, cracks, holes and bubbles, cause many problems in electronic instruments equipped with screens and touch screens. The defect identification process requires advanced imaging techniques to acquire, process and segment screen images, algorithms for the extraction of relevant attributes and classification algorithms to identify problems.
Quality management is of extreme importance not only in the electronics industry but also in the agro-food industry. Automatic image recognition systems can help with many process tasks: water management, disease detection, crop planning and much more.
In the textile industry, it is possible to identify more than 70 types of defects and generally the intervention of a human inspector is required to recognize the problem and intervene (Hanbay et al, 2016). The introduction of automatic recognition systems allows industries to interrupt the production process the instant the defects appear and thus reduce the waste of raw materials, time and guarantee the quality of the fabric. In recent years, alongside statistical and signal analysis methods, machine learning techniques have been introduced that exploit images captured by cameras such as Naive Bayes classifiers, feed-forward neural networks, convolutional neural networks, Support Vector Machines and Genetic Algorithms.
Krummenacher et al (2020) present two learning-based approaches for the detection of defects in railway wagon wheels. Maintenance is extremely important in the transport sector, both to guarantee quality and punctuality in the service, and to avoid situations of danger and damage to people or infrastructures. To identify the problems of the wheels, time series of data collected by sensors placed on the wheels and analyzed by means of convolutional neural networks and Support Vector Machines were used.
The key information in the decision-making process on quality management in products is contained in theopinion of the final consumer. Consumer satisfaction is one of the main aspects and very often it is expressed in posts, comments and reactions expressed on social media such as Facebook, Twitter, TripAdvisor, Amazon and many others. Text mining, a branch of data mining, is a set of techniques and methods specialized in extracting information from large amounts of textual data that, in recent years, has dealt with the analysis of text on social media (Tang et al, 2014). For example, Yussupova et al (2016) used Decision Tree Classifier to analyze opinions from 635824 reviews of Russian hotels and resorts.
Deep Learning and synthetic data: application of GANs for the identification of defects
The screens of smartphones and tablets are very affected by imperfections and defects that can appear on the surface of the glass; for this reason the electronics industry invests a lot of resources in quality management systems (Yuan et al, 2018). In general, the techniques for identifying defects on glass require the analysis of specific characteristics of the images and the knowledge of sector experts is essential to detect the most significant attributes to circumscribe the defect. Learning techniques based on deep learning, like deep neural networks, are appropriate for this type of task and are able to achieve excellent performance even in the absence of expert support; to their detriment, deep learning techniques require larger amounts of data and the use of synthetic data is often necessary.
An interesting approach from Generative Adversarial Networks, GAN, (Goodfellow et al, 2014) was presented for the detection of defects in smartphone screens (Yuan et al, 2018). The GANs are made up of two deep neural networks that compete with each other in a zero-sum game, that is, a game in which victory or loss is balanced. One network is called generator G and produces “bogus” data, the other network is called discriminator D and aims to distinguish the “bogus” data produced by the generator from real data. The goal of the generator is to learn how to create “bogus” data very similar to real data in order to mislead the discriminator. The generator is often compared to a forger whose goal is to print counterfeit banknotes to deceive the financier, i.e. the discriminator: every time the financier learns to recognize counterfeit banknotes, the forger will try to find new strategies to deceive him.
Figure 1: Schematic of a GAN. Generator G produces bogus data G (z) in an attempt to confuse discriminator D in recognizing real data x.
As the ability of generator G increases, its function and that of the training set overlap, causing D to no longer be able to discern the origin of the data. In statistical terms, D is trained in such a way as to maximize log (D (x), the probability that a data comes from the dataset, while G must minimize the log (1-D (G (z))) where z is the representation in latent space. Competition is defined as a zero-sum minmax game, and is described by the following function:
Quality management is a key aspect of competition and business success. Companies have at their disposal many tools, technological, organizational and information, to improve their quality management systems; Recent developments in the machine learning sector have broadened the range of possibilities and companies are initiating innovation and improvement processes to reap the benefits. In the process of innovation of quality management systems, companies must pay attention to their own critical success factors, develop measurement systems and metrics suitable for controlling and monitoring performance to coordinate improvement interventions. Machine learning techniques are proving to be able to support companies in the development of innovative quality management systems. The approach to the problem with supervised learning allows for excellent performance, but it is often a viable path in business contexts that have already embarked on a process of digitization and have correctly annotated data. In contexts in the digitalization phase, in anticipation of the installation of sensors for process monitoring and feasibility studies, it is sometimes necessary to use tools for the generation of synthetic data or unsupervised learning techniques.