For a higher classification accuracy of disturbance signals of power quality, a disturbance classification method for power quality based on gram angle field and multiple transfer learning is proposed in this paper. Firstly, the one-dimensional disturbance signal of power quality is transformed into a Gramian angular field (GAF) coded image by using the gram angle field, and then three ResNet networks are constructed. The disturbance signals with representative signal-to-noise ratios of 0 dB, 20 dB and 40 dB are selected as the input of the sub-model to train the three sub-models, respectively. During this period, the training weights of the sub-models are transferred in turn by using the method of multiple transfer learning. The pre-training weight of the latter model is inherited from the training weight of the previous model, and the weight processing methods of partial freezing and partial fine-tuning are adopted to ensure the optimal training effect of the model. Finally, the features of the three sub-models are fused to train the classifier with a full connection layer, and a disturbance classification model for power quality is obtained. The simulation results show that the method has higher classification accuracy and better anti-noise performance, and the proposed model has good robustness and generalization.

Acquiring labels in anomaly detection tasks is expensive and challenging. Therefore, as an effective way to improve efficiency, pretraining is widely used in anomaly detection models， which enriches the model's representation capabilities, thereby enhancing both performance and efficiency in anomaly detection. In most pretraining methods, the decoder is typically randomly initialized. Drawing inspiration from the diffusion model, this paper proposed to use denoising as a task to pretrain the decoder in anomaly detection, which is trained to reconstruct the original noise-free input. Denoising requires the model to learn the structure, patterns, and related features of the data, particularly when training samples are limited. This paper explored two approaches on anomaly detection: simultaneous denoising pretraining for encoder and decoder, denoising pretraining for only decoder. Experimental results demonstrate the effectiveness of this method on improving model’s performance. Particularly, when the number of samples is limited, the improvement is more pronounced.

Specific emitter identification (SEI) is the process of identifying individual emitters by analyzing the radio frequency emissions, based on the fact that each device contains unique hardware imperfections. While the majority of previous research focuses on obtaining features that are discriminative, the reliability of the features is rarely considered. For example, since device characteristics of the same emitter vary when it is operating at different carrier frequencies, the performance of SEI approaches may degrade when the training data and the test data are collected from the same emitters with different frequencies. To improve performance of SEI under varying frequency, we propose an approach based on continuous wavelet transform (CWT) and domain adversarial neural network (DANN). The proposed approach exploits unlabeled test data in addition to labeled training data, in order to learn representations that are discriminative for individual emitters and invariant for varying frequencies. Experiments are conducted on received signals of five emitters under three carrier frequencies. The results demonstrate the superior performance of the proposed approach when the carrier frequencies of the training data and the test data differ.

Large concrete structures such as buildings, bridges, and tunnels are aging. In Japan and many other countries, those built during economic reconstruction after World War II are about 60 to 70 years old, and flacking and other problems are becoming more noticeable. Periodic inspections were made mandatory by government and ministerial ordinance during the 2013-2014 fiscal year, and inspections based on the new standards have just begun. There are various methods to check the soundness of concrete, but the hammering test is widely used because it does not require special equipment. However, long experience is required to master the hammering test. Therefore, mechanization is desired. Although the difference between the sound of a defective part and a normal part is very small, we have shown that neural network is useful in our research. To use this technology in the actual field, it is necessary to meet the forms of concrete structures in various conditions. For example, flacking in concrete exists at various depths, and it is impossible to learn about flacking in all cases. This paper presents the results of a study of the possibility of finding flacking at different depths with a single inspection learning model and an idea to increase the accuracy of a learning model when we use a rolling hammer.

In recent years, deep learning and especially deep neural networks (DNN) have obtained amazing performance on a variety of problems, in particular in classification or pattern recognition. Among many kinds of DNNs, the convolutional neural networks (CNN) are most commonly used. However, due to their complexity, there are many problems related but not limited to optimizing network parameters, avoiding overfitting and ensuring good generalization abilities. Therefore, a number of methods have been proposed by the researchers to deal with these problems. In this paper, we present the results of applying different, recently developed methods to improve deep neural network training and operating. We decided to focus on the most popular CNN structures, namely on VGG based neural networks: VGG16, VGG11 and proposed by us VGG8. The tests were conducted on a real and very important problem of skin cancer detection. A publicly available dataset of skin lesions was used as a benchmark. We analyzed the influence of applying: dropout, batch normalization, model ensembling, and transfer learning. Moreover, the influence of the type of activation function was checked. In order to increase the objectivity of the results, each of the tested models was trained 6 times and their results were averaged. In addition, in order to mitigate the impact of the selection of learning, test and validation sets, k-fold validation was applied.

The article presents research on animal detection in thermal images using the YOLOv5 architecture. The goal of the study was to obtain a model with high performance in detecting animals in this type of images, and to see how changes in hyperparameters affect learning curves and final results. This manifested itself in testing different values of learning rate, momentum and optimizer types in relation to the model’s learning performance. Two methods of tuning hyperparameters were used in the study: grid search and evolutionary algorithms. The model was trained and tested on an in-house dataset containing images with deer and wild boars. After the experiments, the trained architecture achieved the highest score for Mean Average Precision (mAP) of 83%. These results are promising and indicate that the YOLO model can be used for automatic animal detection in various applications, such as wildlife monitoring, environmental protection or security systems.

In the domain of affective computing different emotional expressions play an important role. To convey the emotional state of human emotions, facial expressions or visual cues are used as an important and primary cue. The facial expressions convey humans affective state more convincingly than any other cues. With the advancement in the deep learning techniques, the convolutional neural network (CNN) can be used to automatically extract the features from the visual cues; however variable sized and biased datasets are a vital challenge to be dealt with as far as implementation of deep models is concerned. Also, the dataset used for training the model plays a significant role in the retrieved results. In this paper, we have proposed a multi-model hybrid ensemble weighted adaptive approach with decision level fusion for personalized affect recognition based on the visual cues. We have used a CNN and pre-trained ResNet-50 model for the transfer learning. VGGFace model’s weights are used to initialize weights of ResNet50 for fine-tuning the model. The proposed system shows significant improvement in test accuracy in affective state recognition compared to the singleton CNN model developed from scratch or transfer learned model. The proposed methodology is validated on The Karolinska Directed Emotional Faces (KDEF) dataset with 77.85% accuracy. The obtained results are promising compared to the existing state of the art methods.

Variation in powertrain parameters caused by dimensioning, manufacturing and assembly inaccuracies may prevent model-based virtual sensors from representing physical powertrains accurately. Data-driven virtual sensors employing machine learning models offer a solution for including variations in the powertrain parameters. These variations can be efficiently included in the training of the virtual sensor through simulation. The trained model can then be theoretically applied to real systems via transfer learning, allowing a data-driven virtual sensor to be trained without the notoriously labour-intensive step of gathering data from a real powertrain. This research presents a training procedure for a data-driven virtual sensor. The virtual sensor was made for a powertrain consisting of multiple shafts, couplings and gears. The training procedure generalizes the virtual sensor for a single powertrain with variations corresponding to the aforementioned inaccuracies. The training procedure includes parameter randomization and random excitation. That is, the data-driven virtual sensor was trained using data from multiple different powertrain instances, representing roughly the same powertrain. The virtual sensor trained using multiple instances of a simulated powertrain was accurate at estimating rotating speeds and torque of the loaded shaft of multiple simulated test powertrains. The estimates were computed from the rotating speeds and torque at the motor shaft of the powertrain. This research gives excellent grounds for further studies towards simulation-to-reality transfer learning, in which a virtual sensor is trained with simulated data and then applied to a real system.