In unmanned aerial vehicle photographs, object detection algorithms encounter challenges in enhancing both speed and accuracy for objects of different sizes, primarily due to complex backgrounds and small objects. This study introduces the PDWT-YOLO algorithm, based on the YOLOv7-tiny model, to improve the effectiveness of object detection across all sizes. The proposed method enhances the detection of small objects by incorporating a dedicated small-object detection layer, while reducing the conflict between classification and regression tasks through the replacement of the YOLOv7-tiny model’s detection head (IDetect) with a decoupled head. Moreover, network convergence is accelerated, and regression accuracy is improved by replacing the Complete Intersection over Union (CIoU) loss function with a Wise Intersection over Union (WIoU) focusing mechanism in the loss function. To assess the proposed model’s effectiveness, it was trained and tested on the VisDrone-2019 dataset comprising images captured by various drones across diverse scenarios, weather conditions, and lighting conditions. The experiments show that mAP@0.5:0.95 and mAP@0.5 increased by 5% and 6.7%, respectively, with acceptable running speed compared with the original YOLOv7-tiny model. Furthermore, this method shows improvement over other datasets, confirming that PDWT-YOLO is effective for multiscale object detection.

With the exponential growth of remote sensing images in recent years, there has been a significant increase in demand for micro-target detection. Recently, effective detection methods for small targets have emerged; however, for micro-targets (even fewer pixels than small targets), most existing methods are not fully competent in feature extraction, target positioning, and rapid classification. This study proposes an enhanced detection method, especially for micro-targets, in which a combined loss function (consisting of NWD and CIOU) is used instead of a singular CIOU loss function. In addition, the lightweight Content-Aware Reassembly of Features (CARAFE) replaces the original bilinear interpolation upsampling algorithm, and a spatial pyramid structure is added into the network model’s small target layer. The proposed algorithm undergoes training and validation utilizing the benchmark dataset known as AI-TOD. Compared to speed-oriented YOLOv7-tiny, the mAP0.5 and mAP0.5:0.95 of our improved algorithm increased from 42.0% and 16.8% to 48.7% and 18.9%, representing improvements of 6.7% and 2.1%, respectively, while the detection speed was almost equal to that of YOLOv7-tiny. Furthermore, our method was also tested on a dataset of multi-scale targets, which contains small targets, medium targets, and large targets. The results demonstrated that mAP0.5:0.95 increased from “9.8%, 54.8%, and 68.2%” to “12.6%, 55.6%, and 70.1%” for detection across different scales, indicating improvements of 2.8%, 0.8%, and 1.9%, respectively. In summary, the presented method improves detection metrics for micro-targets in various scenarios while satisfying the requirements of detection speed in a real-time system.

In this paper, CoFe₂O₄@pyrolyzed bacterial cellulose (CFO@PBC) nanocomposites were synthesized by a hydrothermal process and an in situ composite technique as the cathode for rechargeable Li-O₂ batteries (LOB). CoFe₂O₄@Ketjenblack (CFO@KB) were prepared in the same way as the comparison. The as-prepared materials were characterized by SEM, XRD, and BET. The LOB with CFO@PBC-3 showed a maximum initial discharge capacity of 2660 mAh g⁻¹ (electrode) at the current density of 100 mAg⁻¹, while the mass ratio of CoFe₂O₄ and pyrolyzed bacterial cellulose was 3. The cyclic voltammetry of the CFO@PBC nanocomposite electrodes was investigated at 0.1 mVs⁻¹ in the voltage window of 2.0–4.5 V. With the limited capacity of 640 mAh g⁻¹ (electrode), the CFO@PBC-3 electrode showed the best cycle stability and higher cutoff voltage of discharging. The pyrolyzed bacterial cellulose, as the fibrous 3D backbone in CFO@PBC nanocomposite electrodes, produced better results than the CFO@KB.

Article Peng Wu ^1,2,*, and Ning Xiong ^{1 1} School of Innovation, Design and Engineering Mälardalen University, Västerås 721 23, Sweden ² Department of Computer Engineering, Taiyuan Institute of Technology, Taiyuan 030008, China ^* Correspondence: 14112078@bjtu.edu.cn Received: 31 March 2023 Accepted: 13 July 2023 Published: 26 September 2023 Abstract: The advancement in Internet of things and sensor technologies has enabled data to be continuously generated with a high rate, i.e., data streams. It is practically infeasible to store streaming data in a hard disk, and apply a traditional batch learning method to extract a relevant knowledge model from these data. This paper studies online incremental learning with data streams, in which one sample is processed at each time to update the existing model. For the learning target, the Bayesian classifier is adopted which is a computationally economical model of easy deployment for online processing in edges or devices. By using the individual new example, we first present an online learning algorithm to incrementally update classifier parameters in a way equivalent to the offline learning counterpart. In order to adapt to concept drifts in nonstationary environments, the proposed online learning algorithm is improved to enable recent examples to be more impactful during the sequential learning procedure. Preliminary simulation tests reveal that the improved online learning algorithm can lead to faster model adaption than the unimproved online algorithm when the data drift occurs. In case of presumed stationary data streams without drifts, the improved online algorithm is proved to be competent by performing at least as good as (sometimes, even better than) the unimproved algorithm. The advancement in the Internet of things (IoT) and sensor technologies has enabled an exponential growth of data which is generated and collected from a wide range of sources. Understanding big volumes of data [1] requires cutting-edge tools and techniques to extract useful knowledge and information for decision support. Data being continuously generated at a high rate is referred to as the streaming data (or the data stream) and is of high interest [2]. Data streams are prevalent in many industrial domains such as manufacturing and automatic control systems. As streaming data contains a huge number or even an unbounded sequence of samples, it is practically infeasible to store the full volume of data on a hard disk, and access the data multiple times to extract relevant knowledge. So far, the main stream of the machine learning methods has been focused on a batch-based and offline learning manner. A limitation of offline learning methods is that, with the arrival of a set of new examples, the learning algorithm has to be re-executed using all training data available. Hence, it can lead to an inefficient learning process without utilizing the knowledge learned previously. Particularly, offline learning is incapable of handling data streams in an online environment, in which fresh data is generated continuously with open ends. Incremental learning [3] is a promising technology to deal with online data streams, and aims to update an existing knowledge model based on learning from new data rather than scratches. This basic idea shows analogy to the cognitive behavior (of humans) which is used to acquire knowledge in an incremental fashion over time. One of the earliest incremental learning algorithms was proposed in [4] to solve data clustering problems. According to the definition given in [5], an incremental algorithm should be computationally efficient when incorporating the experience (training data) into models, and should not use any unreasonable space to store the already used experience. This implies that training examples have to be discarded immediately after being processed, leading to rather low memory demands during the learning procedure. Th every fast decision tree (VFDT) algorithm [6, 7] was developed for incremental learning of a decision tree based on streaming data. The tree is learned by recursively replacing leaves with decision nodes. The Hoeffding bounds are used in the VFDT to decide when to install a split-test at a leave. Later Rutkowski et al. [8] proposed a technique to use the Gaussian approximation to select the best splitting attribute in a considered node. It is proved that, by using this technique, the best selected attribute has a high probability to be identical to the attribute derived from the whole data stream. However, no pruning strategies are addressed in the above mentioned papers to prune the underlying tree and avoid over-fitting during the learning procedure. Wang, Minku and Yao developed two online stream mining methods, i.e. the oversampling based online bagging (OOB) and under-sampling based online bagging (UOB) [9], aiming to bias the learning to favor the minority class. Note that the adaptability of varying the imbalance status is supported by the incremental estimation of the occurrence probabilities of classes, and this provides basis to adjust the sampling rate in a dynamic process. A limitation of the OOB and UOB methods is that they can only handle data streams of two classes. Yavtukhovskyi et al. proposed an incremental learning method to construct fuzzy classification rules from data streams [10]. This method not only generates new fuzzy rules to reflect information from new data, but also incrementally updates confidence in the old rules. When the existing knowledge is applied to classify a new instance, the low-confidence rules will not be used in the fuzzy rule based inference. Similar results were conducted using evolving fuzzy classifiers [11-15] that emerged as an own research line in recent years. However, both incremental and evolving fuzzy classifiers treat training samples equally in model construction, regardless of their arrival time. Hence, such classifiers are not well suitable for nonstationary environments, where recently generated data should weigh more than those data generated from the long past. Incremental learning is also exploited to increase the computational efficiency when learning from big data sets. The work in [16, 17] split the large volume of training data into multiple subsets, with each corresponding to a learning episode. In each episode, the knowledge acquired from the preceding stage is updated and refined based on the current data subset. Duda et al. [18] proposed a boosting based mini-batch training approach to accelerating the training process of deep neural networks. This approach differs from traditional learning methods in that 1) learning examples are randomly taken from the original training set to form a continuous data stream; 2) and the mini-batch is constituted by a group of recent elements from the stream for model update. Online learning [19], as a closely related concept to incremental learning, represents a family of machine learning methods attempting to learn from a sequence of samples one by one at a time. The goal of online learners is to optimize the performance for solving a sequence of prediction problems by utilizing known answers to previous tasks. According to [20], incremental learning can work in either an online or a batch mode. In the online mode, only one example is used to update the model each time, while in the batch mode, a batch of multiple examples is used for model update at each time. This paper focuses on incremental online learning of Bayesian classifiers which represent a class of computational cheap and probabilistic models for classification. Bayesian classifiers are appealing for employment in an online setting due to the following reasons. 1) The Bayesian classifiers involve no specification and tuning of the model architecture, which is yet common in many other machine learning models. Changing model architecture will terminate the sequential update procedure and request model training anew. 2) Online learning is often executed in edges or devices where the space capacity is limited. Fortunately, Bayesian classifiers are economic models with a low number of parameters, making their deployment easy on the edge or device levels. Therefore, we adopt the Bayesian classifier as the target model for online learning in this paper. We present an online learning algorithm that uses one sample (along the stream) at a time to update the parameters of the existing Bayesian classifier. It is also clearly shown that this online algorithm works equivalently to the offline learning counterpart that uses the entire data stream as the training data for once. Further, we consider nonstationary environments in which the data distribution and characteristics may evolve with time, i.e., concept drifts. Learning to adapt to concept drifts is one of the grand challenges and is crucial for many real-world applications [21]. To this end, we propose an improved online learning algorithm for Bayesian classifiers, which enables recent examples to be more impactful during the sequential learning procedure. Preliminary simulation tests reveal that the improved online learning algorithm can lead to faster model adaption than the unimproved online algorithm when such a drift occurs in data. In case of presumed stationary data streams without concept drifts, this improved online algorithm is still competent by performing at least as good as (sometimes, even better than) the unimproved algorithm. In summary, the novel contribution of this paper is highlighted from the following aspects. 1) We investigate online learning with Bayesian classifiers that are economic to be implemented in edge devices. 2) We propose the online learning algorithm (Algorithm 1), which is shown equivalent to the offline learning counterpart that uses the whole data stream for once. 3) We further propose the improved online learning algorithm (Algorithm 2), which enables faster model adaptation in nonstationary environments. The remainder of the paper is organized as follows. Section 2 presents the basic principle of the naive Bayesian classifier. The online learning algorithms for Bayesian classifiers are elaborated in Section 3. Section 4 gives the preliminary results of simulation tests to evaluate the proposed learning algorithm. Finally, concluding remarks are given in Section 5. Naive Bayesian classifiers are simple probabilistic models which can be used to solve many classification problems. The basic idea is to consider the posterior probability for each probable class H_k (k=1...K) when attempting to classify a new instance, which is represented by the feature vector (x₁, x₂, ···, x_n). According to the Bayes theorem, this posterior probability is expressed by Assuming that all features are independent of each other, Equation (1) can be further written as: The rationale of this independence assumption of features can be found in [22]. Since the denominator in Equation (2) is a constant given the feature values, we can decide the class y* that maximizes the numerator, i.e., The likelihood in Equation (3) needs to be derived from a predefined probability density function provided that feature takes continuous values. This paper uses the Gaussian function as the density function to express the distribution of a feature under a given class. It was explained in [23] that the Gaussian function is able to approximate a wide range of distributions such as Gamma, Binnominal and Poisson distributions. Thereby, the likelihood required by Equation (3) is computed as where and stand for the mean and standard deviation of the Gaussian function, respectively. Let the training data set be formulated as with representing the feature vector of the ith training example of class . denotes the number of training examples of class . refers to the jth element (the value of feature ) in the feature vector . The prior probability of class and the mean and standard deviation of feature (j= n) under class can be obtained from the training data set, which are formulated as follows: which are considered as the parameters of the Bayesian classifier. This section discusses incremental online learning for Bayesian classifiers. In Subsection 3.1, we present the online learning algorithm that weighs all the samples in the streams equally. This algorithm is then improved in Subsection 3.2 to deal with the concept drift of nonstationary data streams in time-varying environments. 3.1. Online Learning with Stationary Data Stream Online learning is performed to update the Bayesian classifier with each new training example. First, let’s consider stationary data streams in which the data characteristics and distribution remain unchanged with time. Hence, it is required to realize incremental calculation of the classifier parameters defined by Equations (5)–(7). This means that, given a new training example, the old classifier parameters have to be updated to new values which are the same as those obtained using all the old and new training samples together. Let be the prior probability of class after incorporating the first i training samples. With the arrival of the next sample , the class probability is updated by where is a logic variable returning 1 when is of class , and 0 otherwise. Likewise, the mean of a feature under a class can be incrementally updated using examples of that class. Let be the mean of feature under class , which is estimated based on the first i samples of that class. Now, suppose that the next example of class is available and its value on the jth feature is denoted by . We have the new estimate of the mean parameter by updating the old one as follows: Relying on the relation in Equation (9), we further have Equation (10) concerning the update of the standard deviation of features under a class. The standard deviation feature under class is Therefore, we have Based on Equations (8)–(10) and the recursive calculation of the Bayesian classifier parameters, the online BC learning algorithm is presented as follows: The outcomes of Algorithm 1 are the prior probabilities of the class, the mean and the standard deviation of class features. These parameters of the Bayesian classifier can be used in Equations (3) and (4) for online classification of new instances, and are continuously updated during the execution of the algorithm. 3.2. Adapting to Concept Drifts The online BC learning algorithm presented in Subsection 3.1 is based on the recursive calculation of the class probabilities, and the mean and standard deviation of features under classes. For the sake of preserving the knowledge learned in the past, all historical samples are weighed equally when the total learning process is completed. Nonetheless, in a dynamic environment where data distributions evolve with time, the data that arrives recently is of more importance than the data that arrives long before. In order to adapt to concept drifts, a challenging issue is how to learn from more recent data. This subsection tackles this issue by proposing the improved online BC learning algorithm. Revisiting Equations (8)–(10), it can be seen that the inverse of (i+1) is an important coefficient to decide the step size, where the old parameters of the Bayesian classifier are updated by a new example and i+1 is the number of examples that have been received so far. Obviously, the step size diminishes quickly with the increase of i, which results in the declining influence of new samples along with time. This is, unfortunately, contradictory to our expectation of learning with drifting data streams that, more recent samples should gain more importance in the learning procedure in order to not only dilute outdated knowledge, but also adapt to the concept drift. Under this consideration, we replace by a constant coefficient , which is termed as the learning rate in Equations (8)–(10). This leads to the new online incremental learning rules for the Bayesian classifier parameters given below: The advantage of this revision is that, with the time-independent learning rate, the influence of training examples (that arrive later) will not decrease with time. We have the opportunity of defining a suitable value for to achieve timely reactions (of the model) to the streaming drift. In principle, the value of is related to the speed of the data evolution. The faster the data evolves, the larger has to be. On the other hand, the learning rate must not be too large in order to avoid quickly forgetting previously acquired useful knowledge, i.e. the risk of catastrophic forgetting [23]. Based upon the new learning rules of Equations (12)–(14) , the improved online BC learning algorithm is given in Algorithm 2. This section presents the simulation tests conducted on some benchmark data sets, and evaluates the efficacy of the online learning algorithms. Subsection 4.1 briefly introduces the experimental setup. Subsequently, the results of simulation tests are given and discussed in Subsection 4.2. 4.1. Experimental Setup Eight benchmark datasets from the UCI machine learning repository [24] are used for simulation tests in this paper. All these data sets are used for classification, see Table 1. Each data set is treated as a data stream in the simulation tests. This means that each training example is processed once and then discarded during the execution of the online learning algorithms. Table 1. Summary of the eight benchmark data sets. The K-fold cross-validation method [25] is used in the evaluation to make a fair assessment of the generalization ability of the learning algorithms. The concrete value of K is set to be 10 in the experiments. Hence, each data set is divided into 10 equal parts with one part being used as test data, and the other nine parts together being used as the training data (in each of the 10 trials). The final classification accuracy is the mean of the accuracy based on the test data of the 10 trials. Additionally, the learning rate for the improved online algorithm is set to be 0.01 when learning with all the data sets. 4.2. Results and Discussions We apply the traditional Bayesian classifier (offline BC learning) and the two online learning algorithms (online BC learning and improved online BC learning) to the eight benchmark data sets. The mean and variance of accuracy obtained by the three algorithms in the 10-fold cross-validation are shown in Table 2. It can be seen from the table that the online BC learning algorithm always has the same results as the offline BC learning algorithm on all the eight data sets. This is fully consistent with our theoretic recognition that, the online BC learning algorithm is equivalent to the offline learning counterpart which uses the entire data stream for once. In the comparison of the two online algorithms, we find that the improved online BC learning algorithm performs similarly as well as the online BC learning algorithm and, on the Australian data set, performs even slightly better. This is an interesting observation given that the used data sets are static containing no drifts inside. This indicates that, although proposed for nonstationary streams, the improved online BC learning algorithm can work well in stationary environments. The reason is that, if the data stream is time invariant, the new samples will still follow the current data distribution that has been learned so far. Therefore, increasing the influence of recent samples, as implemented in the improved online BC algorithm, will not lead to significant changes in the...

Herein, a three-step approach toward a multi-layered porous PBC/graphene sandwich has been developed, in which the chemical bonding interactions have been successfully enhanced via esterification between the layers of pyrolyzed bacterial cellulose (PBC) and graphene. Such a chemically induced compatible interface has been demonstrated to contribute significantly to the mass transfer efficiency when the PBC/graphene sandwich is deployed as electrode material for both supercapacitors and lithium–sulfur batteries. The high specific capacitance of the supercapacitors has been increased by three times, to 393 F g⁻¹ at 0.1 A g⁻¹. A high initial discharge specific capacity (~1100 mAhg⁻¹) and high coulombic efficiency (99% after 300 cycles) of the rPG/S-based lithium–sulfur batteries have been achieved.

Decentralized power systems are commonly used in high-speed trains. However, many parameters in decentralized power systems are uncertain and inevitably have errors. We present a reasoning method based on the interval numbers for decentralized power systems in high-speed trains. Uncertain parameters and their unavoidable errors are quantitatively described by interval numbers. We also define generalized linear equations with interval numbers (LAIs), which can be used to describe the movement of the train. Furthermore, it is proven that the zero sets of LAIs are convex. Therefore, the inside of the fault-tolerance area can be formed by their vertexes and edges and represented by linear inequalities. Consequently, we can judge whether the system is working properly by verifying that the current system state is in the fault-tolerance area. Finally, a fault-tolerance area is obtained, which can be determined by linear equations with an interval number, and we test the correctness of the fault-tolerance area through large-scale random test cases.