45results about How to "Suppress overfitting" patented technology

The invention discloses an automatic identification system of a number plate on the basis of a simplified convolutional neural network. The convolutional neural network comprises an input layer, a convolutional layer, a pooling layer, a hidden layer and a classification output layer and solves the problem of number plate identification under a daily background. The number plate identification comprises the following steps: positioning, segmenting and identifying. The invention puts forward a positioning method which extracts colorful edges by colorful edge information and colorful information. Since parameters in the method are set on the basis of color features, noise in the daily background can be effectively inhibited, and input images of different sizes can be subjected number plate extraction. The automatic identification system omits a front convolutional layer of a traditional depth convolutional neural network and only keeps one layer of convolutional layer and one hidden layer. As the supplementation of a missing convolutional layer and the strengthening of input features, a gray level edge image obtained by a Sobel operator is used as the input of a colorful image, i.e., coarsness features which are artificially extracted replace features extracted by multiple convolutional layers of the traditional convolutional neural network.

The invention discloses a static image crowd counting method based on joint learning, includes the following steps: pre-training stage: using 50-layer residual network to train the ImageNet 2012 classification data set to obtain the parameter initialization classifier network, classifying the image blocks into three categories by a Softmax, respectively corresponding to three regressors; In the training phase, every image block of the training data set is input to three regressors respectively, and the same image block will get different counting results. The regressor with the smallest counting error is used as a classification tag to mark the image block, and the three kinds of image blocks are used to fine-tune the regressors respectively. Classifier training phase: random sampling andensure that the number of labels of each category is consistent; Joint training phase: continuous iterative training of classifiers and regressors. The invention can carry out counting and density estimation in crowd dense scene, has certain scale self-adaptability, and improves counting precision and model generalization ability.

The invention relates to a rotating-machinery life-stage identification method based on a deep self-encoding learning network of noise enhancement samples. For the purpose that extraction and expression of rotating-machinery life features as well as life stage identification are automatically learned under the condition of a small sample size, noise enhancement are conducted on training samples; after a plurality of sparse self codes are stacked, classification layers are added to construct the deep sparse self-encoding learning network which can not only automatically learn extraction of the life features, but also intelligently identify the life stages. Stepwise non-supervision adaptive learning and supervision fine tuning are conducted on the the samples obtained after noise enhancement through multi-layer sparse self encoding, so as to inhibit deep-network over fitting and improve network robustness. Therefore, automatic extraction and expression of the rotating-machinery life features are achieved, and finally intelligent identification of the rotating-machinery life stages in the classification layers are completed. The rotating-machinery life-stage identification method can be applied in identifying rolling bearing life stages, and identifying results are good under the condition of a small sample size.

The invention provides an aircraft electromechanical system fault identification method based on a random convolutional neural network. The method includes the following steps that: firstly, the short-time Fourier transform is adopted to construct a vibration acceleration signal into a two-dimensional time-frequency map with good spatial correlation, and the problem of two-dimensional transformation of a one-dimensional signal is solved; secondly, the transformed two-dimensional time-frequency map is input into the random convolutional neural network, wherein the network adopts a random dropout mechanism to suppress the model overfitting and enhance the generalization ability, and adopts a momentum stochastic gradient descent algorithm to update network parameters and further complete theconstruction of an identification model; and finally, the random convolutional neural network is adopted to identify the fault of an aircraft electromechanical system. The scheme of the invention is good in identification effect and strong in practicability, is simple and practicable, and is suitable for fault identification of an aircraft system.

The invention provides an electromagnetic red information detection method based on cepstrum and a convolutional neural network. Firstly, samples of a plurality of electromagnetic leakage signals arecollected and downsampled; information leakage characteristics in the electromagnetic leakage signals are extracted through cepstrum analysis; electromagnetic red information feature representation based on cepstrum is formed; a large number of classification training are carried out on the extracted red information features through the convolutional neural network; a detection model about the electromagnetic red information is obtained; a to-be-detected electromagnetic leakage signal sample is input; downsampling and cepstrum electromagnetic red information feature extraction are also carriedout, finally, the trained electromagnetic red information detection model is used for carrying out identification judgment on the red information features, and the detection accuracy of the electromagnetic red information is evaluated by comparing a detection result with a priori label of the to-be-detected electromagnetic leakage signal sample. According to the electromagnetic red information detection method provided by the invention, the electromagnetic leakage signal can be detected in an environment with a low signal-to-noise ratio, the sensitivity is high, and the detection accuracy ishigher than that of a traditional method.

The invention discloses a small data cross-domain action recognition method based on a double-chain deep double-flow network. Efficient action recognition of other small-scale data sets based on a source domain data set is achieved. Meanwhile, a data set CDSAR used on the cross-domain action recognition task is provided. The method comprises the following specific steps: (1) preprocessing video; (2) building a model based on a double-chain deep double-flow network; (3) constructing a target function based on the double-chain deep double-flow network; (4) adopting a small data cross-domain action recognition method based on a double-chain deep double-flow network. Efficient action recognition can be carried out on other data sets only with a small number of samples based on the known data set, the problems that the data size of the target data set is small and data distribution among different data sets is inconsistent can be effectively solved, and the convergence speed of the method is high.

A method for extracting near-infrared brain function signals based on a least squares support vector machine, the invention relates to a method for extracting near-infrared brain function signals. The purpose of the present invention is to solve the problem of low detection accuracy of near-infrared brain function activity signals. The specific process is called: 1: Obtain the time signal of the optical density change of two near-infrared lights of different wavelengths at different distances; 2: Obtain the time signal of the concentration change of oxygenated hemoglobin and the time signal of the concentration change of reduced hemoglobin; 3: Obtain the time signal of the change in the concentration of oxygenated hemoglobin Brain function activity signal; four: get W of W(t) ^* (t); Five: Get the brain function activity signal E(t): Six: Get the brain function activity signal after eliminating the error interference; Seven: Construct the Lagrangian function to get the linear equation system; Eight: Solve the linear equation system to get La The numerical solution of the Grangian multiplier vector α and the offset b, at this time, the brain function activity signal processed by the least squares support vector machine regression function is expressed as E ^* (t). The invention is used for brain function signal extraction.

The invention discloses a supporting-vector-machine-based industrial melt index soft measuring meter and a method. By performing the optimal optimizing on parameters of a fuzzy neural network through the supporting vector machine, the problems of parameter set of the existing fuzzy neural network are solved, and meanwhile, due to the adoption of the method, the structure of the whole fuzzy neural network is subjected to self-adaption updating so as to adapt the change of input data. According to the invention, an on-site intelligent meter and a control station are connected with a DCS (distributed control system) database; a soft measurement value display device comprises a supporting-vector-machine-based industrial melt index soft measuring model; the DCS database is connected with the input end of the soft measuring model; and the output end of the supporting-vector-machine-based industrial melt index soft measuring model is connected with the melt index soft measurement value display device; the measuring meter and the method disclosed by the invention have the characteristics of self-adoption optimized model structure, high noise resisting property and preferable popularization property.

The invention provides a ContenseNet algorithm fused with an attention selection mechanism, and the algorithm comprises the steps: carrying out the feature extraction on data in a convolutional neuralnetwork through m network structure blocks, wherein each structure block comprises n groups of complete feature transformation layers, enabling the data to obtain a corresponding feature matrix through each feature transformation layer, cascading the m network structure blocks, performing feature extraction on data of the network structure blocks to obtain a final feature matrix, calculating a loss value of network training according to the obtained feature matrix, calculating an error term and a weight gradient of each layer, and judging whether the network is converged or not according to the loss value, if not, adjusting the convolutional neural network initialization parameter according to the weight gradient to perform training again, and if so, outputting a network training result.According to the CondenseNet algorithm fused with the attention selection mechanism, multi-dimensional feature information is efficiently utilized, the learning and expression capacity of a deep network is enhanced, and the classification accuracy is improved.

The invention discloses an industrial melt index soft measurement method through combination of a fuzzy neural network and group intelligent optimization. The method is applied to online prediction for an industrial melt index in the propylene polymerization production process. A data preprocessing module, a fuzzy neural network module, a dragonfly optimizing algorithm module, a sliding window module and a fuzzy rule optimization module are involved. According to a sliding window method, input of the fuzzy neural network every moment is provided, by using an improved dragonfly optimizing algorithm, a fuzzy neural network parameter is intelligently optimized, and the setting problem of the fuzzy neural network parameter in the prior art is solved. According to the method, global and local information in the optimizing process is considered at the same time, a network parameter can be stably and online corrected, and the adaptivity of the system to input data fluctuation is improved. Themethod has the advantages of stable optimized model structure, good generalization performance and high anti-interference capacity.

The invention discloses a soft industrial melt index measurement instrument and method for an optimal fuzzy network. According to the method, a previous fuzzy neural network is optimized by introducing a support vector machine so that the problem that parameters are difficult to set in the construction process of the fuzzy neural network is solved. According to the soft industrial melt index measurement instrument and method for the optimal fuzzy network, a field intelligent instrument and a control station are connected with a DCS (Data Communication System) database; a soft measurement value displayer comprises a soft industrial melt index measurement model for the optimal fuzzy network; the DCS database is connected with the input end of the soft measurement model; the output end of the soft industrial melt index measurement model for the optimal fuzzy network is connected with the soft melt index measurement value displayer. Finally, the soft industrial melt index measurement instrument and method for the optimal fuzzy network and the soft measurement have the characteristics of on-line measurement, high calculation speed, strong noise immunity and good popularization performance.

The invention relates to an imaged liquor recognition method applicable to multiple electronic nose platforms. The method comprises the steps that liquor sample original data collected by different sensors is acquired and subjected to normalization processing; sensor response values are converted into an angle cosine and matrix through an anti-cosine function; a half-angle sine and matrix Hs and ahalf-angle cosine and matrix Hc are obtained; Markov transfer probabilities are calculated after binning operation is performed on sensor responses, and then a Markov domain matrix MTF is calculated;sensor images are constructed, piecewise aggregate approximation is utilized to determine image dimensions, and multiple sensor images are combined for each sample to obtain a combination image describing the single sample; a convolutional neural network model is established; and training is performed.

The invention provides a pedestrian re-identification method and device. The pedestrian re-identification method includes the steps: measuring the similar distance of each test sample pair in a test set one by one based on a metric space; and sorting the similar distance measurement results of all test sample pairs in the test set to obtain a distance measurement matrix. The pedestrian re-identification method obtains a group of stable combined metric space through iterative learning, calculates the combined similar distance of the test sample pairs in the multi-metric space, overcomes the over fitting problem of an existing metric learning method in the high-dimensional small sample problem, and has higher robustness for the complex feature change in the pedestrian re-identification problem.

The invention provides an intelligent fault diagnosis method based on a stack pruning sparse denoising automatic encoder and a convolutional neural network, which is called sPSDAE-CNN for short. According to the method, original input data is processed by using the stack denoising automatic encoder, and more training data is obtained by using a data enhancement method. The stack sparse pruning and noise reduction self-encoder comprises a full-connection automatic encoding network, and the characteristics extracted at the front layer of the network are used for performing the operation of the subsequent layer, which means that some new connections appear between the front and rear layers of networks, so that the information loss is reduced, and more effective characteristics are obtained; meanwhile, pruning operation is introduced, so that the training efficiency and precision of the network are improved, higher training speed and high adaptability to noise signals are achieved, and the overfitting problem of the convolutional neural network is suppressed to a certain extent; according to the method, the flight data of the quad-rotor unmanned aerial vehicle are input into the model, and high fault diagnosis accuracy is obtained under the condition of high noise interference.

The invention discloses an accent classification method based on a deep neural network. The method comprises the steps: extracting frame-level frequency domain features of an original audio, and constructing a 2D speech spectrum as a network input X; constructing a multi-task weight shared CRNNs-based front-end encoder to extract local sequence descriptors {P1,..., PT'} of the spectrum X; in the training process, additionally arranging a speech recognition task branch network behind the front-end encoder to inhibit an overfitting phenomenon in accent recognition; constructing a core branch network for an accent recognition task, and firstly, integrating all local sequence descriptors into a global accent feature; secondly, introducing a discriminative loss function in a prediction process; and finally, classifying the global accent features through a softmax-based classification layer to realize accent prediction. The invention also discloses a highly discriminatable accent classification model based on the deep neural network, and reliable accent prediction can be provided for speakers from different regional groups.