The knowledge-enhanced BERT model solves the problem of lacking knowledge in downstream tasks by injecting external expertize, and achieves higher accuracy compared with BERT model. However, owning to large-scale external knowledge is utilized into knowledge-enhanced BERT, some shortcomings comes such as information noise, lower accuracy and weak generalization ability, and so on. To solve this problem, a knowledge enhancement method BERTBooster which combines incremental learning and gradient optimization is proposed. BERTBooster disassembles the input text corpus into entity noun sets through entity noun recognition, and uses the incremental learning task denoising entity auto-encoder to create an incremental task set of entity nouns and external knowledge triples. Furthermore, BERTBooster introduces a new gradient optimization algorithm ChildTuningF into BERT model to improve the generalization ability. BERTBooster can effectively improve the factual knowledge cognition ability of CAGBERT model and improve the accuracy of the model in downstream tasks. Experiments are carried out on six public data sets such as Book_Review, LCQMC, XNLI, Law_QA, Insureace_QA, and NLPCC-DBQA. The experimental results show that the accuracy rate in downstream tasks is increased by 0.65% on average after using BERTBooster on CAGBERT.

To solve the problem of query restriction in supply-chain financial blockchain system, this paper proposes a privacy budget management and noise reusing method in multichain blockchain environment based on Hyperledger multichannel technology and community clustering algorithm. A historical record book is established to manage the privacy budget according to historical query types, and a differential privacy protection algorithm based on noise reusing is used to generate and reusing noise. In the experiments, we tested the privacy budget loss, data utility and system performance based on the business data-set from chemical supply chain. The experimental results show that the blockchain system with multichain structure based on community clustering algorithm reduces the system data storage space and decreases request processing time effectively. The privacy budget loss of the proposed method is about 1/3 of that of the pure Gaussian mechanism method. After 100 queries, the total amount of noise is about 8% less than that of pure Gaussian mechanism. Besides that, the noise reusing algorithm reduces the loss of the privacy budget and breaks through the query limitation caused by privacy budget wasting.

The real-world complex networks, such as biological, transportation, biomedical, web, and social networks, are usually dynamic and change over time. The communities which reflect the substructures hidden in the networks usually overlap each other, and detecting overlapping communities in the dynamic complex networks is a challenging task. Prior researchers have applied multiobjective optimization method to the detection of dynamic overlapping communities and achieved some excellent results. However, in terms of multiobjective processing, the prior studies all adopt the decomposition method based on weight parameters, and different weight parameters or different parameter values can easily affect the community detection results which further results in the uneven distribution of the detected results in the target space. To solve the above problems, a hybrid algorithm, that is, Collaborative Particle Swarm multiobjective Optimization-based Dynamic Overlapping Community Detection (CPSO-DOCD) algorithm is proposed in this paper. First, to improve the diversity of particles, the encoding/decoding of the particle and the cross inheritance and the variation of particle are redefined first based on label propagation. In each network snapshot, multiple particle swarms are initialized based on Community Overlap Propagation Algorithm (COPRA) to generate particles with uniform distribution. Multiple different objective functions are optimized using multiple particle swarms respectively to avoid the incorrect selection of weight parameters. In addition, a reference-point-based is adopted in the particle selecting stage to solve the uneven distribution of detected results in the target space. Second, a node label entropy-based particle swarm algorithm is proposed to improve the accuracy of community detection of current network snapshots. Finally, when one snapshot switches to another over time, a migration strategy based on COPRA local-search and clique generation is utilized to adjust the prior community detection results, which enables the former results can be adapted to the new network snapshots. The experiments are implemented based on four dynamic networks which are Cit-HepPh, Cit-HepTh, Emailed-EU-core-temporal, and CollegeMsg. The hypervolume value of the overlapping community detection result obtained by CPSO-DOCD is 0.5%–2% higher than MDOA, MCMOEA, SLPAD, and iLCD. Furthermore, CPSO-DOCD also performed better than MDOA, MCMOEA, SLPAD, and iLCD on C-metric values, and CPSO-DOCD can approach approximately to the Pareto frontier.

Aiming at the problems of secondary pollution and resource waste caused by inaccurate input of coal and ammonia in coal-fired power plant, an optimal controlling method of combustion and denitration coordinated operation based on Deep Deterministic Policy Gradient (DDPG) is proposed in this paper. First, the environmental model is constructed by the Stacking algorithm to predict the NOx emission concentration of the combustion and denitration system, which provides environmental state feedback for the optimal controlling model. Second, the optimization controlling model is constructed based on the DDPG algorithm within the standard limitation of denitration efficiency and NOx emission concentration. This model takes the minimization of comprehensive cost as its optimization objective to realize the optimal control of controllable variables in the cooperative operation process of combustion and denitration. The experimental results of real operational data from 1000 MW boiler unit in a power plant locating in south China show that the optimization results of coordinated operation for the combustion and denitration system are better than single-stage optimization results. In addition, the total cost is reduced by 1%–3% on average compared with before optimization.

Terahertz (THz) wave is an electromagnetic wave with a frequency between far infrared ray and millimeter wave, which is widely used in hazardous material detection for its waveband fingerprint spectroscopy. THz time-domain spectroscopy technology based on deep learning can be used for nondestructive detection of various hazardous materials by recognizing the fingerprint spectrum of substances. However, due to the high cost of collecting spectral data, training samples are not easy to obtain and scarce for classification models, which leads to poor training effectiveness and low accuracy of classification. To address this problem, a fully connected layer-based auxiliary classifier generative adversarial network (FC-ACGAN) data augmentation method is proposed in this paper, we realized the generator and discriminator with fully connected layers to fit original data distribution better and generate data with higher quality. First, THz time-domain spectral data from seven flammable liquids were augmented using Mixup and FC-ACGAN, and then we fed the generated data set and expanded data set into Residual Network (ResNet), convolutional neural network, fully convolutional network, and multilayer perceptron for training. It is demonstrated that our method can solve the overfitting of models because of insufficient data. Compared with direct training on original data set, the accuracy of models using augmented data set improved by 5.1325% on average, which is 3.15% higher than that using Mixup. Furthermore, we experimented on expanded data set with ResNet long short-term memory for classification, the final accuracy reaches 99.42% on average, which is 1.09% higher than that using the original data set.

Cross-modal retrieval has attracted great attention due to the increasing demand for tremendous amounts of multimodal data in recent years. These retrievals could either be text-to-image or image-to-text. To address the problem of inappropriate information included between images and texts, we propose two cross-modal recovery techniques established on a dual-branch neural network defined on a common subspace and the hashing learning method. First, a cross-modal recovery technique established on a multilabel information deep ranking model (MIDRM) is provided. In this method, we introduce a triplet-loss function into the dual-branch neural network model. This function takes advantage of the semantic information of the bimodal components, focusing on not only the similarities between similar images and text features but also the distances between dissimilar images and texts. Second, we establish a new cross-modal hashing technique said to be the deep regularized hashing constraint (DRHC). In this method, the regularized function is used to replace the binary constraint, and the discrete value is constrained to a certain numerical range so that the network can achieve end-to-end training. Overall, the time complexity is greatly improved, and the occupied storage space is also greatly reduced. Different experiments on our proposed MIDRM and DRHC models demonstrate their superior performance to those of the state-of-the-art methods on two widely used data sets. The experimental results show that our approach also increases the mean average precision of cross-modal recovery.

Aiming at the problems of complex structure, high components coupling, and difficultly monitoring of the whole health status with the industrial robot, a metric learning-based whole health indicator model is proposed. First, according to the more obvious degradation characteristics of industrial robots during accelerated operation, the accelerated signal is segmented and then the time-domain features are extracted. Second, the long-term and short-term memory (LSTM) network combined with the multihead attention is used to construct the network model, and the metric learning method is adopted to learn the similarity measurement method of the industrial robot monitoring data. Finally, the similarity measure method got from metric learning is used to construct the whole health indicator, which describes the whole degradation trend of the industrial robot. The experiments are based on the real accelerated aging data set from industrial robots. The results show that the proposed model can effectively construct the whole health indicator for industrial robots. The average trend of the proposed model reaches 0.9769. The average monotonicity reaches 0.5666, which is 0.1748, 0.1577, and 0.1492 higher than the similarity measurement method based on Euclidean distance, Markov distance, and LSTM.

To solve the problems such as low accuracy and low retrieval performance in feature extraction of environmental sound data from Internet consumer finance scenario, a novel method for environmental sound filtering and feature extraction (VF-EFENet) is proposed. First, the Conv-TasNet speech separation model is clipped and migrated to filter foreground voice. Second, an environmental sound feature extraction model is established based on the improved VGGish, and pretraining weight is used to improve the feature extraction accuracy. Finally, metric learning is used to optimize the distance function to improve retrieval accuracy. Metric learning can make the same kind of audio feature space cohesive and the different types of audio feature space away. The experiments are implemented based on AISHELL-1 and ESC-50 data sets to test voice filter performance, average classification accuracy and average retrieval accuracy. The experimental results show that VF-EFENet can effectively filter the voice in mixed audio and the SI-SNR reaches 12.51 db. When sampling rate is 8 kHz, the average classification accuracy is improved by 8.3% after voice filtering using VF-EFENet. When Top30 samples are retrieved, the average retrieval accuracy of VF-EFENet is 7.37% higher than that of ESResNetAttention.

Aiming at the problems of high manual cost, low efficiency, and low precision of the mechanical axis health management in industrial robot applications, this paper proposes a health assessment and state prediction algorithm based on hidden Markov model (HMM) and temporal convolutional networks (TCN). First, the MPdist similarity comparison algorithm is used to construct the mechanical axis health index. Then the hidden Markov model is trained with observable sensor data. After that, the temporal convolution neural network is used to predict state transition time iteratively, and the predicted results are decoded by HMM. The experimental results show that the HMM-TCN model can accurately assess the health state of the mechanical axis and predict the state transition in real-time. The prediction accuracy of this method reaches 87.5%, and the error interval locates in [−3,9] time steps. The accuracy, early/late prediction indicators are better than HMM-RNN, HMM-LSTM, and HMM-GRU.

Since the new crown epidemic, mask-wearing has become a new normal in people's work and life. The inspection mechanism for mask-wearing at the entrance and exit of public places is seriously insufficient. The phenomenon of “pick-up on entry” has led to the severe formalization of mask-wearing inspection. Manual detection of mask-wearing in an open and dynamic crowded environment is unrealistic, which is not only time-consuming and labor-intensive but also cannot achieve early warning throughout the entire process. In response to this problem, this paper proposes a real-time recognition and early warning method for mask-wearing in an open, dynamic, complex environment based on improved YOLOv5 R6.1. First, replacing the first Conv structure of the backbone network in the YOLOv5 R6.1 model with an improved Stem structure to minimize the computational overhead while improving the performance. Then by normalizing the data, the random erasure data expansion technique is used to enhance the antiocclusion robustness of the algorithm. Finally, according to the mask-wearing specification in the training data set, optimizing and adjusting the anchor box parameters of the YOLOv5 R6.1 model to improve the model's ability to recognize small targets. The experiments are based on open data sets, and the results show that the mean precision (mAP), precision, and recall of this method reach 92.9%, 94.1%, and 88.5% on average, and the average frames per second (FPS) reaches 117. Moreover, the mAP and FPS are improved by an average of 6.5% and 474% compared with algorithms based on RetinaNet, Attention-Retina, Single Shot multibox Detector, Fast-RCNN, YOLOv4, and YOLOv5.

In this paper, a non-negative matrix factorization feature expansion (NMFFE) approach was proposed to overcome the feature-sparsity issue when expanding features of short-text. First, we took the internal relationships of short texts and words into account when segmenting words from texts and constructing their relationship matrix. Second, we utilized the Dual regularization non-negative matrix tri-factorization (DNMTF) algorithm to obtain the words clustering indicator matrix, which was used to get the feature space by dimensionality reduction methods. Thirdly, words with close relationship were selected out from the feature space and added into the short-text to solve the sparsity issue. The experimental results showed that the accuracy of short text classification of our NMFFE algorithm increased 25.77%, 10.89%, and 1.79% on three data sets: Web snippets, Twitter sports, and AGnews, respectively compared with the Word2Vec algorithm and Char-CNN algorithm. It indicated that the NMFFE algorithm was better than the BOW algorithm and the Char-CNN algorithm in terms of classification accuracy and algorithm robustness.