80results about How to "Fast reasoning" patented technology

The invention discloses an intelligent fault diagnosis system for an ICNI system, which can improve the maintenance efficiency, carry out intelligent and automatic diagnosis and is applicable to the ICNI system. According to the technical scheme of the invention, a knowledge base and a management module thereof carry out standardization research and mathematical modeling based on a fault tree, an SQL Server database software framework is adopted, a relational database is used for building a logic relation among a fault phenomenon, a fault mode, a detection method, a historical case and a fault tree internal event to form the knowledge base; and a diagnosis information acquisition module interacts with an automatic testing system via Ethernet to acquire diagnosis data from the ICNI system and a testing instrument, a reasoning machine module adopts CBR and RBR hybrid diagnostic reasoning, after comprehensive judgment is carried out on the fault phenomenon inputted by the user, the field knowledge stored by the knowledge base and the diagnosis data from the automatic testing system, a reasoning method is automatically selected to carry out reasoning diagnosis on the fault, a reasoning process and a reasoning result are outputted to an explanation machine module, and a diagnosis report is generated.

The present disclosure includes systems and techniques relating to representation and retrieval of data structures in databases. In general, embodiments of the invention feature a computer program product and a method including storing a generalized directed acyclic graph (DAG) in a database, wherein the storing includes encoding path information of the generalized DAG in entries of a path table in the database, the encoding includes converting the path information into text strings, and the entries of the path table correspond to paths in the generalized DAG from nodes of the generalized DAG to a root node of the generalized DAG; triggering generation of a lexical index of the path table using the text strings, wherein the lexical index separately lists tokens included in the entries; and retrieving one or more portions of the generalized DAG from the database for in-memory operations.

The invention discloses a transformer substation equipment sound fault detection and positioning method based on deep learning, and the method comprises the steps: obtaining the sound data of transformer substation equipment, carrying out the data labeling of the sound data of the transformer substation equipment, and generating a sample data set; carrying out data enhancement on the sample data set by adopting an audio noise adding mode, and carrying out data preprocessing on substation equipment sound data in the enhanced sample data set to generate a multi-channel spectrogram sequence; forthe multi-channel spectrogram sequence, constructing a substation equipment sound fault detection model based on convolutional neural network coding and long-term and short-term memory network detection; and training the substation equipment sound fault detection model according to the sample data set to obtain a trained substation equipment sound fault detection model, and performing fault detection and positioning on the substation equipment sound. According to the technical scheme, the problems of low detection efficiency, poor algorithm robustness and the like in an existing substation equipment sound fault detection technology are solved.

The invention discloses a global rank perception neural network model compression method based on a filter feature map, and solves the problems of high labor cost, low pruning efficiency and poor stability of a filter pruning model compression method. The method comprises the steps of obtaining and preprocessing image data; constructing a universal convolutional neural network, and setting hyper-parameters; selecting a loss function and an optimization algorithm; training and storing the pre-training model; acquiring an average rank of the feature map; adaptively learning a parameter matrix; representing and sequencing the importance of the filters; pruning and storing the pre-training model; and finely adjusting the network model to realize the global rank perception neural network model compression based on the filter feature map. According to the method, one-time learning and multi-time pruning are adopted, the filters are globally sequenced, pruning is unified, and ill-conditioned selection does not exist. The pruning effect and stability are better, the pruning efficiency of large-scale different-complexity networks is high, and the adaptability of different edge devices is good. The method is used for computing and storing resource-limited edge equipment.

The invention belongs to the technical field of RPA element recognition, and particularly relates to an RPA element pickup same-screen switching method and system. The method comprises the following steps: S1, selecting an element pickup mode; s2, moving the mouse to a specific software interface; s3, obtaining interface related information and interface screenshots, and transmitting and analyzing the interface related information and the interface screenshots; s4, forwarding the interface screenshot, and judging whether the specified time T is exceeded or not; s5, performing analysis processing on the interface screenshot algorithm to obtain an algorithm analysis result; s6, packaging an algorithm analysis result, and displaying the algorithm analysis result on a software interface; s7, judging an operation mode of the user; and S8, after the user selects the corresponding element, the picked element is highlighted again, and picking is completed after the user confirms that the element is correct. The method combines a service fragmentation mechanism, operation mode distinguishing and a front-end and rear-end processing strategy of floating window display, and has the advantages of being more excellent in pickup effect and wider in application range.

The invention relates to a human face pose estimation method and system based on knowledge distillation, which can compress a human face pose estimation model by applying a feature distillation method on the premise of ensuring the accuracy, the model parameter quantity after feature distillation is less, the reasoning speed of the human face pose is higher, the resource consumption is reduced, The method solves the problems of high resource consumption and low accuracy of a shallow network when a deep network structure is applied to face pose estimation, improves the pose estimation accuracy of a large-angle, fuzzy and mask-wearing face image according to the current application scene demand, improves the robustness of the model, improves the face pose angle prediction effect of the model, in practical application, a face pose angle prediction result in a complex scene is more accurate, and a prediction effect exceeding a ResNet50 network structure is realized by using a ResNet18 basic network.

The invention discloses an end-to-end semi-supervised target detection method based on improved yolov5, relates to the technical field of target detection, and solves the technical problems that the existing target detection is relatively complex and the precision improvement of a model is limited by a pseudo tag. The labeled data and the unlabeled data are loaded at the same time in the training process; taking the model updated by using the ema algorithm as a teacher model, and generating a pseudo tag by using a dynamic threshold mode; the loss of the yolov5 is modified, the loss is divided into supervised loss and unsupervised loss, and the unsupervised loss part is optimized by using a difficult case mining strategy. The whole process is simpler, the pseudo labels can be more and more accurate in the training process, the detector is helped to achieve better performance, the reasoning speed is higher, and the real-time requirement can be met.

The invention discloses a distributed retrieval resource library selection method based on a variational auto-encoder, and the method comprises the steps: building an encoder and decoder network structure by utilizing a deep neural network, learning implicit representation of a resource library text, and capturing deep semantic representation of the resource library text; reasoning the extended text of the query word through a model obtained by an unsupervised training method to obtain a hidden representation of the query word; obtaining the correlation ranking of the resource library by calculating the similarity between the query words and the implicit representation of the resource library. The model is unsupervised training, a resource library and a hidden representation vector of a text are automatically obtained, and the defect of text feature design in a supervised training method can be overcome. In addition, the network structure of the variational auto-encoder is simple, andthe calculation time consumption of variational reasoning is lower than that of an LDA topic model based on a Markov chain Monte Carlo reasoning method. And after model training is completed, the timeconsumption for resource library selection is low, and the resource library selection efficiency is high.

The invention provides a construction method of a mobile terminal flower recognition model. The construction method comprises the following steps of S10, creating a convolutional neural network modelof a floating point type trained by an ImageNet data set; S20, adding a quantization operation, i.e., after weight reading and activation output in an original floating point calculation model, inserting an analog quantization operation; S30, training the convolutional neural network model by using the flower data set until the model converges; S40, converting a floating point model into a 8-bit integer operation model, and obtaining a flower recognition model; and S50, compiling the flower recognition model into an APK installation package by using a Bazel construction tool. According to thepresent invention, the floating point operation convolutional neural network for mobile terminal flower recognition is converted into the efficient 8-bit integer operation convolutional neural network, so that the model size is reduced, the model prediction time is shortened, and the precision is reduced very low.

The invention discloses a cerebral function imaging diagnostic method based on a knowledge base. The cerebral function imaging diagnostic method based on the knowledge base comprises the following steps that the cerebral function imaging diagnostic problem is represented through knowledge, wherein tree-based representation and building of a cerebral function imaging diagnostic problem knowledge-based information model are included; a knowledge model is designed, the knowledge base is built, data of the knowledge base are managed, a data base is combined with the knowledge base, and imaging characteristics are extracted. According to the cerebral function imaging characteristics, knowledge representation theories of knowledge engineering are introduced, the problem is described according to the tree-based representation method based on knowledge, rules in auxiliary diagnosis are represented on this basis, and therefore the knowledge base of the problem is formed. By the use of the technology combining the knowledge base, the data base and medical image processing, a semi-automatic cerebral function imaging auxiliary diagnosis process according to the method of the knowledge engineering is designed, and the cerebral function imaging diagnostic method with some information having reference value can be provided for doctors.

The invention relates to a low-illumination image brightness enhancement and super-resolution method based on a double-channel codec, and belongs to the technical field of computer vision images. The method comprises the steps of 1, performing feature extraction on a dark light image through an encoder sharing parameters to obtain a set of feature maps; 2, sending the feature map to a super-resolution decoder for decoding to obtain a super-resolution feature map; 3, pooling the feature map output by the encoder in the step 1 and the super-resolution feature map in the step 2 respectively to obtain two feature vectors, performing weighted fusion on the feature vectors by using an attention mechanism, and then sending the feature vectors to a low-illumination decoder for decoding; and finally, carrying out post-processing on the outputs of the two decoders to obtain a corresponding image after super-resolution. The invention is reasonable in design, low-illumination enhancement and image super-resolution tasks are combined for the defects of an existing low-illumination enhancement method, the visual effect of the reconstructed image is improved, and a good effect is achieved on the low-illumination enhancement and super-resolution comprehensive tasks on the whole.

The invention relates to an intelligent lamp pole crowd behavior recognition method based on human body joint point coordinates. The method specifically comprises the following steps: S1, constructingan image training set for human body behavior recognition and a human body joint point coordinate data set corresponding to human body behaviors; S2, acquiring crowd behavior information and extracting human skeleton behaviors by a camera of the intelligent lamp post, and interpolating the human skeleton behaviors without articulation points in the crowd behavior information according to the human articulation point coordinate data set; S3, setting a time attention layer for human skeleton behaviors, extracting features of the human skeleton behaviors, then constructing a deep learning network, and setting corresponding hyper-parameters; and S4, training the deep learning network through the deep learning open source framework to obtain a crowd behavior recognition model, and recognizingnewly obtained crowd behavior information through the crowd behavior recognition model. Compared with the prior art, the method has the advantages of improving the human body behavior recognition rate, reducing the influence of shielding objects on the recognition result and the like.

The invention relates to an unsupervised hyperspectral video target tracking method based on spatial-spectral feature fusion. The hyperspectral target tracking method based on deep learning is designed in combination with a cyclic consistency theory method, a hyperspectral target tracking deep learning model can be trained in an unsupervised manner, and the cost of manual labeling is saved. On the basis of a Siamese tracking framework, an RGB branch (space branch) and a hyperspectral branch are designed; RGB video data is used for training a space branch, a trained RGB model is loaded into network fixed parameters, meanwhile, a hyperspectral branch is trained, and the fused features with higher robustness and discrimination capability are obtained; and finally, the fused features are input into a correlation filter (DCF) to obtain a tracking result. According to the method, the problem of manual labeling of the hyperspectral video data and the problem of few hyperspectral training samples for deep learning model training can be solved, and the precision and speed of a hyperspectral video tracking model can be effectively improved.

The invention discloses a hyperspectral image classification method and system based on lightweight neural architecture search. A hypernetwork is built through modular lightweight candidate operations, the serialization of the discrete candidate operations is carried out on the edge through a weighted mixing operation,double-layer optimization of the hypernetwork is carried out in a gradient optimization mode, and meanwhile, the model search speed is increased through subnet weight sharing. Then, in the optimization process, a greedy decision is utilized to select an undispersed edge, an operation with the maximum framework parameter on the edge is reserved, other operations in the edge are deleted, the remaining network forms a new super-network, the new super-network is iteratively optimized in the mode, the super-network is continuously simplified along with continuous dispersion of the edge and deletion of the operations on the edge, and finally a lightweight deep neural network architecture for hyperspectral image classification is obtained. The lightweight module is fully utilized to construct the super network, the neural architecture search method based on sequential greedy is realized, and the network architecture with less parameter quantity and higher classification precision can be automatically generated.

The invention relates to a remote sensing image target detection method based on dense connection and feature enhancement, and the method comprises the following steps: building a remote sensing image data set, and inputting the remote sensing image data set into a remote sensing image detection model for training, inputting a remote sensing image to be detected into a trained remote sensing image detection model to obtain a target detection result, wherein the remote sensing image detection model comprises a feature extraction unit, a feature enhancement unit, a feature pyramid unit and a predictor. An input image of the remote sensing image detection model is sequentially processed by the feature extraction unit, the feature enhancement unit, the feature pyramid unit and the predictor to obtain a target detection result. Compared with the prior art, the method has the advantages that the feature extraction capability of the network is improved, the resolution of the input image is increased, and low-latitude feature information is reserved while parameters are reduced so as to adapt to detection of a model on a remote sensing image target.

The invention provides a remote sensing image multi-class target detection method based on sample reweighting. The method comprises: firstly, performing image data augmentation processing and scale zooming preprocessing; then, constructing a target detection network, the target detection network comprising a feature extraction module, a feature enhancement module and a detection head module, and performing feature enhancement operation on part of feature hierarchies in order to achieve significance expression of features; then, performing a network end-to-end training process, and guiding thetraining network to pay more attention to more target samples with large aspect ratio difference by adopting a sample reweighting strategy so as to optimize the training model; and finally, realizinga target detection process, inputting a to-be-detected remote sensing image into the trained target detection network to obtain a category prediction value and a coordinate offset of each priori frame, and filtering out a detection result with a relatively high overlapping rate for the same target by using non-maximum suppression. The method has high remote sensing image target detection precisionand speed.

The invention relates to a pet identification method, device and equipment and a computer readable storage medium. The pet identification method comprises the steps: carrying out the enhancement of a pet marking image according to a generative adversarial network, and obtaining an enhanced training image; training the enhanced training image to obtain a lightweight target detection model YOLO and an image classification model based on an addition network AdderNet; detecting the pet in the acquired image according to the YOLO model to obtain a target image only containing the pet; and inputting the target image into an AdderNet-based image classification model to obtain the category of the pet in the target image, thereby reducing the training cost of the neural network model, reducing the hardware cost through YOLO and AdderNet, accelerating the reasoning speed, and ensuring the real-time performance and accuracy of smart community pet identification.