40results about How to "Increase training rate" patented technology

The invention discloses an insulator category detection method based on deep transfer learning, comprising: 1. Preprocessing the aerial insulator image; 2. Expanding the preprocessed aerial insulator image and classifying different types of aerial insulator images ;3. Use the YOLO algorithm to initially locate the aerial insulator images with complex backgrounds, and perform normalization processing on the positioned insulators; 4. Construct an Inception deep learning network with a multi-level differential adaptive architecture; 5. Construct test samples 6. Construct an insulator state cognitive feedback adjustment mechanism based on semantic error entropy. Through the method of deep transfer learning, the present invention can realize the self-optimization adjustment and reconstruction of the multi-level differentiated feature space of the insulator state and its classification criteria, thereby improving the state detection rate of the self-explosion of the aerial image of the insulator under different backgrounds, and meeting the requirements of accuracy and speed actual needs.

The invention discloses a method for identifying deformation faults of a skirt plate grille of a railway motor car, which solves the problem of low detection efficiency of the existing method for identifying deformation faults of the skirt plate grille of a railway motor car, and belongs to the technical field of fault identification of a railway motor car. The invention includes: constructing a skirt plate grid deformation sample set of a railway motor car; using the skirt plate grid deformation sample set to train a deep learning target detection network Faster R-CNN to obtain a Faster R-CNN detection model and weight; using the Faster R-CNN ‑CNN detection model and weights identify the side image of the railway train to be detected, and determine whether the skirt grille in the side image is deformed and the position of the deformation. The feature extraction network of the present invention includes: replacing the 3x3 convolution kernel in the Bottleneck block in the Resnet-50 network with a 3x1 convolution kernel connected in series with a 1x3 convolution kernel, and replacing the ReLU activation in the feature extraction network with the Swish activation function function. The invention recognizes and detects deformation faults of train skirt plate grids, and effectively avoids recognition errors caused by fatigue and individual judgment differences during manual detection.

The invention discloses an intelligent rotating machinery fault depth network feature identification method. By arranging a vibration sensor at the part of the rotating machinery to be detected in the train rolling bearing, the original vibration sequence of the rolling bearing is collected, and then the original vibration sequence is decomposed by a singular spectrum analysis method. Reconstruct and extract the root mean square value, standard deviation, skewness index and peak value of the reconstructed vibration sequence, use the rotating machinery fault location diagnosis model trained by the support vector machine to judge the fault location, and then assemble the reconstructed vibration sequence Empirical mode decomposition calculates the permutation entropy values ​​of a set of decomposed intrinsic modal components, uses the permutation and combination of permutation entropy values ​​as detection features, and uses the rotating machinery fault type diagnosis model trained by support vector machine to judge the fault type. The invention can detect the fault position and fault type of the rotary machine in time, and improves the accuracy and reliability of fault diagnosis.

The invention discloses a scented electrospinning film and a preparation method thereof. The method comprises the following steps: (1) adopting a Na2CO3 water solution to degum raw silk, then dissolving, and carrying out dialysis and freeze-drying treatment to obtain regenerated silk protein; (2) taking a formic acid solution of spice-poly(2-ethyl-2-oxazoline) as a core layer spinning solution and taking a formic acid solution of the regenerated silk protein as a shell layer spinning solution for carrying out coaxial electrospinning so as to obtain electrospun fibers; (3) adopting methanol to carry out aftertreatment on the electrospun fibers obtained in the step (2) so as to obtain the scented electrospinning film. The scented electrospinning film prepared by the method has higher spice loading capacity and a better slow release property.

The embodiments of the present specification provide a method and device for asking a rhetorical question based on a dialogue system, the dialogue system includes a reinforcement learning model, and the method includes: acquiring a first dialogue round based on the dialogue system, the first dialogue round A dialogue round includes T rounds of dialogue; acquire T states corresponding to the first dialogue round; acquire a plurality of feedback score vectors corresponding to multiple rhetorical questions included in the T round of dialogue, the feedback score The multiple dimensions included in the vector correspond to multiple predetermined features respectively; based on the T states, the multiple rhetorical questions, and the multiple feedback score vectors, the reinforcement learning model is trained to be used based on the reinforcement The learning model asks rhetorical questions to user questions.

The invention relates to a semi-supervised ultra-limit learning machine classification method based on graph balance regularization. The present invention balances the adjacency graph based on label consistency and the adjacency graph based on information structure consistency through non-negative weights, so as to achieve graph balance, and can obtain the Laplacian regular term of the optimal graph to constrain the model, And it is considered that when the information consistency map cannot describe the structural information of the sample set well, the weight of the label consistency map should be increased, otherwise the corresponding proportion should be reduced. The present invention first constructs the adjacency supervised graph between the training samples through the label consistency of the samples, combines it with the semi-supervised graph based on the consistency of the sample information to constrain the output of the model, and changes the description data by reasonably adjusting the proportion of the graph The ability of the distribution, after obtaining the optimal adjacency graph, the best output weight vector is obtained. The invention has broad application prospects in brain-electrical signal processing and brain-computer interface systems.

The present application provides a people counting method and device, the method is applied to a server equipped with a people counting model including an image feature extraction sub-model based on a convolutional neural network and an SSD classification regression sub-model, specifically including: inputting an image frame into an image The feature extraction sub-model generates an image feature map; based on the SSD classification regression sub-model, a default box is generated for each pixel in the image feature map, and the position coordinates and probability scores of each default box are obtained, and the maximum probability score is used as the primary confidence; The top K default boxes with the highest confidence are used as target candidate boxes; based on the position coordinates and probability scores of each target candidate box, bounding box regression analysis and softxmax classification are performed to obtain the coordinate position and final confidence of each target candidate box; based on non- The maximum value suppression algorithm obtains the target frame, and counts the number of people in the monitoring area based on the number of target frames. Using this method can effectively improve the real-time performance of people counting.