Search

Industrial production processes are increasingly collecting data from machines in operation due to the cost reduction and popularization of sensor technologies. Valuable information is generated by using the right sensors and proper techniques on the machine’s current state in operation. This extraction allows detecting whether the machine is operating in a degraded state and then, if necessary, interrupts its operation before it goes into a broken state. The detection of the abnormal behavior of a machine is relevant since detection at the right time can reduce financial costs due to machine breakdown and production downtime. This work proposes a novel unsupervised method to detect anomalies in industrial machines and interrupt their operation before this machine goes into a state of breakdown. The proposed method receives as input a set of time series data from several sensors. Using a collection of statistical features, it calculates an indicator that characterizes the failure’s severity, issuing an alert to the machine operator if necessary. The method was evaluated in two benchmarks with known univariate data and two proprietary datasets with multivariate data. Conducted experiments revealed the low computational time spent on training and on evaluation. Overall results measured in Area Under the Receiver Operating Characteristic Curve (AUC) in Yahoo’s benchmark were 89.3%; in Numenta, it was 70.3%, and in the two multivariate datasets evaluated, it was 92.4% and 91.2%. These high AUC values reveal the potential of the proposed method applied in predictive maintenance in a large soybean oil production industry.

Uncontrolled environmental conditions often impact manufacturing processes and lead to product quality fluctuations. For machine tools, thermal influences are a major limitation to the volumetric performance. Climate controls for the shop floor, and machines, or thermally stable structural designs are economically not feasible, promoting control-based compensation as a possible solution. Since the relationship between disturbing quantities and effects are complex and specific to each machine, appropriate modelling is a critical requirement. The authors describe an approach for developing hybrid models, superposing white-box model knowledge, and machine learning. The overall effort can be optimized by combining and balancing different modelling methods, like designing the physical model part and training intelligent algorithms. A general model structure allows a continuous integration of different white-box and black-box model components. The authors integrate self-developed smart sensors into a demonstrator machine tool to test and validate the performance of the approach.

Industry 4.0 requires flexible and fast solutions to automatically identify and handle different parts during the production process. For such a multiclassification (MCC) task, usually feature detectors or machine learning approaches are used. However, increasing product variety and consequently rising number of classes impel the expansion of data set sizes, a mandatory procedure for accurate object identification. The acquisition of a sufficiently large number of sample images is a costly and time-consuming issue. This work presents an iterative reference data set creation method of small size image data sets for feature descriptor-based MCC. The novel method compares the shape of an object in a candidate image with the shapes from a growing reference image set consisting of all previously accepted candidates. An image is assigned to or rejected from the reference data set depending on the found shape similarity. Rejected images form a grid search set that is later used to optimize the feature descriptors’ hyperparameters and enable the addition of new classes. The benefits of this method are the small number of images to be acquired for MCC, the possibility of adding new parts without re-training, the little overhead for new applications, and its compatibility with most commonly used feature descriptors. When compared with a small control data set provided by an inexperienced user with an accuracy of 49% for a five classes MCC, the reference data set built by the novel method gains 20% on the accuracy (69%, 21 images in total) and can be performed by the same inexperienced user.

Shearography and thermography are two well-established nondestructive testing methods. Yet, both methods present high subjectivity in the interpretation of their results, which difficulties the automation. Many efforts go towards applying intelligent algorithms, drawing together many limitations, such as the need for a large data set and fine-tuning. In addition, the classification task to categorize defects is often used to check continuous parameters, such as impact energies, leading to the wrong interpretation of the results. In this work, we train stacked models for impact energy estimation in composites with images from both shearography and thermography. The base estimators feeding these stacked models are random forests trained with the standard deviation curves of the defect images and residual neural networks trained with the test images. Such images come from controlled impact tests performed on composite plates. The ability to estimate continuous impact energy values is also analyzed, as well as the capacity of these networks to make an out-of-distribution generalization, which provides error results in the order of 0.2 J mean absolute error for energies values never seen in the training progress. Choosing such regression models trained with the combination of two non-destructive tests enables the reduction of the dataset size and the monitoring of the defect progression by analyzing continuous energy values.

Composite materials have mechanical behavior comparable to metallic alloys, with the benefit of being lighter. However, due to their anisotropy, integrity characterization of impact damages remains a challenge. Non-Destructive Testing (NDT) methods are useful in this context, as they achieve success in evaluation while they avoid modifications in the characteristics of the piece. Shearography is an NDT that reveal changes on a surface in response to a load. Yet, shearography outputs carry several unwanted characteristics along with the defect, like background patterns, light changes, and noise. Image segmentation techniques can enhance the capability of automatic measurement of defective areas and also aid supervised methods and feature extractors, which rely on images as inputs. Yet, most of the times image pre-processing is required for better and useful results in segmentation. Principal Component Analysis (PCA) and Multi-dimensional Ensemble Empirical Mode Decomposition (MEEMD) allow this to be done at the same time that they deal with background and light changes, as they decompose the image. Using Matthews Coefficient as a metric for masks comparison, it is shown that MEEMD has better results than PCA, and with lower expanded uncertainty.

Carbon fiber reinforced plastics (CFRP) are composite materials which are an interesting alternative to metal alloys in fields such as oil, aerospace, automotive, since CFRP have mechanical properties like metals but with a fraction of their weights. However, these materials have typically a highly anisotropic behavior, which may hinder the characterization of their integrity for example when subjected to an impact, because of its stochastic nature. Non-destructive testing (NDT) methods are interesting for integrity assessment, as they can evaluate the damage extension without affecting any part characteristics. Optical lock-in thermography (OLT) is an convenient NDT inspection alternative since it is a depth-wise method in which one can set different loading frequencies, leading to different scan depths. Pre-processing techniques like Principal Component Analysis (PCA) and Empirical Mode Decomposition (EMD) can be used to more accurately evaluate the damaged area. Their dimensionality reduction capability is highly desired as OLT images of CFRP laminates do not only show the defect, but also undesired information such as changes of background radiation, noise and the disposition of the fiber tissue. Traditional feature extraction methods must be highly tuned to obtain useful results. PCA and EMD methods may be considered non-supervised approaches, making them useful for a wide variety of inputs. However, PCA and EMD have their own natural limitations when being applied to images. While PCA may require high computational effort because of mathematical manipulation of matrices due to the mathematical manipulation of large matrices, problems due to mode superposition may occur if EMD original method is applied on images. In this sense, in this work it has been performed a comparison between PCA, with a new input vector architecture used to mitigate its problem with matrix dimensions, and a derivation of EMD method, called Multi-dimensional Ensemble Empirical Mode Decomposition (MEEMD), applied to prevent the previously mentioned superposition but without a higher computational effort, when segmenting OLT phase images. Outputs in both cases are binary masks with an estimation of the defect region, which were compared to a ground truth manually defined by a specialist. Matthews Correlation Coefficient (MCC) was chosen as segmentation comparison metric instead of F-score, since the last one does not take true negatives into account. The main difference during the application of PCA and MEEMD is that MEEMD yields two kinds of results, one for each frequency image and one average for the whole frequency set, whereas PCA gives only results in relation to the average. This study shown that PCA has better performance than MEEMD when the average results are compared, but MEEMD provides better results if onlt the best image per frequency set is used. Both PCA and MEEMD can yet provide interesting results for three-dimensional reconstruction with OLT images, thus further investigation of such techniques is desirable.

Electric motors are subjected to many different quality control tests during their manufacture. Some of these tests are typically performed by human operators. It is well known in the literature that these operators are not reliable for repetitive inspections due to factors such as subjectivity and fatigue. Vision systems come as an alternative to perform visual tests for quality control. The authors have already proposed a vision system based on edge-detection tools to identify defects in electric motors characterized by one or more coil segments of the winding that are not properly fastened to the other coils and are placed in the projection of the orifice where the rotor is inserted. In this paper, a comparison between an improved version of this first algorithm and a convolutional neural network is done. Data augmentation is used to enhance the image data set, improving the reliability of network training. This dataset was also extrapolated to emulate the results of a manufacturing line. For a test dataset, neural networks presented better results than the edge detection algorithm, but their performance was similar for extrapolated images. For large production volumes, it is recommended the use of neural networks with proper training, but for small datasets the edge detection algorithm with proper parametrization is still the best choice.

High subjectivity, lack of attention and fatigue are factors inherent to human analysis in inspection activities such as shearography, a non-destructive optical method. In order to minimize the probability of human error, a study was conducted in which a binary classification from 256 shearography test samples obtained from pipes repaired with glass fiber patches was performed. The dataset was split into major and minor defects and used to train two convolutional neural networks architectures, - a specific artificial neural network well known for its application on image classification. Architecture A achieved a maximum accuracy of 73% on major defect detection, while architecture B, slightly more complex, led to better results. Posterior studies on architecture B led to the conclusion that a combination of double layer filters and dropout layers are the best setup for this type of classification problem. It is possible that other architectures might lead to better results, but no grid search was performed to confirm this assumption. An accuracy of 79% was achieved with Architecture B, therefore is reasonable to say that convolutional neural networks are able to learn from parameters which are difficult to correctly process, such as the fringe patterns obtained from shearography test samples.