108results about How to "Strong self-learning ability" patented technology

The invention provides an autonomous underwater vehicle (AUV) trajectory tracking control method based on deep reinforcement learning, belonging to the field of deep reinforcement learning and intelligent control. The autonomous underwater vehicle trajectory tracking control method based on deep reinforcement learning includes the steps: defining an AUV trajectory tracking control problem; establishing a Markov decision-making process model of the AUV trajectory tracking problem; constructing a hybrid policy-evaluation network which consists of multiple policy networks and evaluation networks;and finally, solving the target policy of AUV trajectory tracking control by the constructed hybrid policy-evaluation network, for the multiple evaluation networks, evaluating the performance of eachevaluation network by defining an expected Bellman absolute error and updating only one evaluation network with the lowest performance at each time step, and for the multiple policy networks, randomly selecting one policy network at each time step and using a deterministic policy gradient to update, so that the finally learned policy is the mean value of all the policy networks. The autonomous underwater vehicle trajectory tracking control method based on deep reinforcement learning is not easy to be influenced by the bad AUV historical tracking trajectory, and has high precision.

The invention relates the auto sorting faulty goods method on the high-speed automatic production line and unit. The method comprises the following steps: sorting faulty goods system receiving the sorting signal, starting the freedom degree linked conveying device 3, then positioning the mechanical arm connected with 3 on the faulty goods, the mechanical arm keeping the same speed with conveyer belt, gripping the faulty goods, lifting it and releasing it to appointed position. The unit comprises electrical control and mechanical part. The mechanical part comprises 3 and mechanical arm which comprises arm and hand. In the center of the mechanical hand there is camera, and the electrical control is connected with 3 and mechanical arm. The invention possesses high accuracy, low cost, good versatility, and no environmental pollution.

The invention provides a dynamic load balancing method based on self-adapting prediction of network flow. The scheme includes: observing historical data regularity of network load flowing into a load balancing switch or a switching software in a certain time cycle; obtaining a parameter value of self-adapting algorithm in a prediction program to form a computing formula; then substituting load observation value at present moment into the formula to predict load value at the next moment; and distributing flow for a rear-end server at real time according to prediction value, thereby enabling the network load to be regulated in advance, avoiding lag effect, constantly keeping the network load in a comparatively balancing state, greatly strengthening self-adapting self-regulating capability of network to load, and being adaptable to networks with a certain time cycle regularity at occasions such as regular-period network backup and the like.

The invention provides an intelligent arrhythmia diagnosis method based on multiple-lead and a convolutional neural network. The method includes the steps that 1, data samples are selected; 2, arrhythmia types are labeled; 3, led heartbeat signals are intercepted; 4, a normalized heartbeat set is obtained; 5, a concealed layer and an output layer are constructed; 6, a target function is set; 7, sample training is conducted; 8, arrhythmias classification is applied. According to the intelligent arrhythmia diagnosis method, network learning efficiency and precision of automatic arrhythmia diagnosis can be improved by training the convolutional neural network (CNN) through multiple-lead electrocardiogram data, a universal frame and a specific method for training the CNN through the multiple-lead electrocardiogram data with arrhythmia type labels are achieved, the arrhythmia types of electrocardiosignals to be diagnosed can be accurately judged, and the arrhythmia types can serve as diagnosis results or as reference of doctors.

A self-adaptive selection dynamic production scheduling control system, which is realized via a computer, is characterized in that: the system comprises a system emulator, a learning machine, a decision-making machine, a scheduling rules base, a scheduling knowledge base, a carrier, processing equipments and a buffer station thereof; the buffer station is provided with an optical grating, a sensor and a detection equipment; when a working piece reaches the buffer station and is processed, the learning machine detects the current system status for learning, so as to acquire dynamic scheduling knowledge about the system and update the knowledge in the scheduling knowledge base; when one processing equipment needs to be scheduled, the decision-making machine reads corresponding scheduling knowledge in the scheduling knowledge base according to the detected system status, acquires new scheduling knowledge through continuous interactive learning with the processing system, dynamically selects the scheduling rules based on the status of the processing equipments and the working piece in the system, and chooses the optimized scheduling rules to schedule the processing equipments. The invention can adapt to instable time-varying workshop dynamic production environments, obtain a better working-piece arrangement than prior rule-based scheduling technology, effectively reduce the process waiting time, and improve the fill rate of product delivery time.

The invention relates to an optimized design method for modelling of an end wall of a high load fan/compressor. The method comprises eleven steps. As the stage pneumatic load is improved, internal flow of the fan/compressor is more severe, massive secondary flow is easily generated in an area close to the end wall, and a risk of flow stall exists. The modelling of the end wall has important influence on control over the secondary flow in an end area of the high load compressor, and becomes an important research direction. Firstly, an end wall modelling parameter defining method is researched; secondarily, based on an adaptive genetic algorithm and an artificial neural network response surface model, which are developed by the research group, an orthogonal experimental design, end wall parameter definition and a flow field value simulating technology are combined to realize automatic global optimization of the end wall of the compressor; thirdly, the effectiveness of the optimizing method is verified through different test functions; finally, a blade grid of the compressor is subjected to nonaxisymmetrical optimization based on the method. A result shows that the blade grid after modelling has good pneumatic performance.

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 invention discloses a robust controller of a permanent magnet synchronous motor based on a fuzzy-neural network generalized inverse and a construction method thereof. The construction method of the invention comprises the following steps of: combining an internal model controller and a fuzzy-neural network generalized inverse to form a compound controlled object; serially connecting two linear transfer functions and one integrator with the fuzzy-neural network with determined parameters and weight coefficients to form the fuzzy-neural network generalized inverse, serially connecting the fuzzy-neural network generalized inverse and the compound controlled object to form a generalized pseudo-linear system, linearizing a PMSM (permanent magnet synchronous motor), and decoupling and equalizing the linearized PMSM into a second-order speed pseudo-linear subsystem and a first-order current pseudo-linear subsystem; and respectively introducing an internal-model control method in the two pseudo-linear subsystems to construct the internal model controller. The robust controller of the invention has the advantages of overcoming the dependence and local convergence of the optimal gradient method on initial values and solving the problems of randomness and probability caused by using the simple genetic algorithm, obtaining the high performance control, anti-disturbance performance and adaptability of the motor and simplifying the control difficulty, along with simple structure and high system robustness.

The invention discloses a face authentication method based on a convolutional neural network and Bayesian decision. The face authentication method comprises the steps of 1), training the convolutional neural network and a Bayesian model by means of a face training database; 2), performing preprocessing such as face detection and face alignment on a testing database, and randomly combining test faces for obtaining 6000 pairs of faces; 3), extracting a characteristic vector of a testing face image pair by means of the convolutional neural network, and calculating similarity; and 4) after performing PCA dimension reduction on the characteristic vector, feeding the characteristic vector into a Bayesian network, calculating posterior probability according to the similarity, setting a threshold and determining whether each pair of faces belongs to one person. The face authentication method has advantages of improving robustness in face authentication and improving face authentication speed and face authentication accuracy. The face authentication method can be used in the field of identity authentication, public security, etc.

The invention provides a GBM multimodal magnetic resonance image segmentation method based on a deep neural network. The method comprises the following steps that firstly each slice image of the collected GBM multimodal magnetic resonance image is preprocessed and then all the slice images are divided into training samples and test samples, and the slice images of the training samples are marked;then the training sample image blocks are extracted and the mean and the variance are standardized, and a training data asset is formed after data amplification; then one deep neural network is constructed and the deep neural network is trained by using the training data set so as to obtain a deep neural network segmentation model; and finally the slice image to be segmented is preprocessed and the image blocks are extracted, and the voxels are classified and post-processed by using the deep neural network segmentation model so that GBM multimodal magnetic resonance image segmentation can be realized. The high requirements of automatic diagnosis, surgical planning and prognosis for the detection and locating accuracy of the abnormal brain tissues and the surrounding normal structures can be met.

The invention discloses an intelligent street lamp energy-saving control system based on an artificial neutral network. The intelligent street lamp energy-saving control system comprises a sensor system, an energy-saving control system and a power controller. An output signal of the sensor system is connected with an input terminal of the energy-saving control system; and an output signal of the energy-saving control system is connected with the power controller. Besides, the intelligent street lamp energy-saving control system is characterized in that: the sensor system includes environmental optical signal collection processing module, an infrared signal collection processing module and a sound signal collection processing module; the energy-saving control system consists of a DSP embedded system and an artificial neutral network module that is arranged in a DSP chip; and the artificial neutral network module is formed by a forward algorithm of an artificial neutral network BP algorithm and an artificial neuron that has been trained. According to the invention, according to peripheral environment situations of all street lamp illumination units, intelligent power control can be realized; and on the premise that an illumination requirement is met, energy consumption is effectively reduced.

The invention discloses a transformer fault detection method based on an SOM (Self Organizing Map) neural network. The method comprises the following steps: S100: selecting a transformer as a testing object, and acquiring vibration signals of the transformer in different states as sample data; S200: decomposing and extracting a characteristic vector by utilizing ensemble empirical mode decomposition in Hilbert-Huang transform; S300: inputting the characteristic vector into the SOM neural network; S400: calculating a distance between a weight of a mapping layer and an input vector; S500: adjusting weights of an efferent neuron and an adjacent neuron; S600: judging whether pre-set conditions are met or not, and finishing SOM neural network training to obtain a testing sample; and S700: inputting the testing sample, and outputting the transformer fault type corresponding to the testing sample according to the network, thereby realizing the technical effect of online detection of the transformer.

The invention discloses an NRIET quantitive precipation estimation method based on cloud classification and machine learning. The quantitive precipation estimation method based on cloud classificationand machine learning comprises the steps of first preprocessing radar data and rain gauge data, and matching a radar reflectivity with rain gauge precipation data based on a site; identifying different cloud systems such as strati and convective clouds according to the radar reflectivity intensity; then performing fitting training in real time using a machine learning regression algorithm to obtain a relationship model between cumulative precipation and radar combined reflectivity; and finally, applying the relationship model between cumulative precipation and radar combined reflectivity to radar combined reflectivity lattice field data in real time to obtain a quasi-real-time quantitive precipation estimation field.

The invention discloses a financial business risk assessment method, a risk control server side and a storage medium. The method comprises the steps of obtaining a preset number of risk assessment sample data of sample users; Performing clustering analysis on the risk assessment sample data by using a K-means algorithm to obtain K sample categories and sample data corresponding to the sample categories; Before a financial business handling request is received, risk assessment data of a user is obtained, and according to the risk assessment data and a sample type, the sample type to which the risk assessment data belongs is determined; Adding the risk assessment data into the sample data corresponding to the determined sample category to form new sample data; Calculating the new sample datathrough a decision tree algorithm to derive a decision condition; And updating the decision tree model according to the decision conditions, and evaluating the risk probability of the financial business through the updated decision tree model. According to the invention, the self-learning capability of the risk control model is improved, and the accuracy of risk control evaluation is improved.

The invention relates to a character recognition method and system based on an attention mechanism, and relates to a deep learning and image processing technology. According to the method, a convolutional neural network and a linguistic module based on an attention mechanism are used as the backbone of a deep learning model, a customized loss function is used to reinforce feature map extraction, the model is guided to learn to distinguish a foreground and a background during training, and forward and reverse bidirectional decoders are introduced to perform bidirectional decoding on characters.The method and the system are high in anti-interference capability, attention drifting can be reduced, and meanwhile, the situation that final recognition fails due to the fact that the first character is difficult to recognize during the forward decoding of the model can be avoided.

The invention relates to a Fourier descriptor and BP (Back Propagation) neural network-based garment style identification method. The method comprises the steps of preprocessing a garment image to obtain an outer contour of the garment; performing Fourier description of the outer contour of the garment, and performing data preprocessing; and performing BP neural network-based garment style identification. The preprocessing of the garment image refers to a process that the garment image is subjected to segmentation processing to obtain a garment region, and the garment image is subjected to edge detection to obtain a contour image of the garment; the Fourier description of the outer contour of the garment refers to a process that a standardized Fourier descriptor eigenvector of a contour shape of the garment is extracted, and the data preprocessing refers to normalization processing and principal component analysis performed on the standardized Fourier descriptor eigenvector; and the BP neural network-based garment style identification refers to garment style identification performed on a principal component matrix by using a three-layer BP neural network. The method can achieve the identification accuracy of 81%, is good in robustness and generalization ability, and can be suitable for identification of garment styles in garment images.

The invention relates to evaluation system and method for a virtual machine. The evaluation system comprises a test template module, an automatic acquisition module for system software-hardware configuration, an automatic system test module, a self-learning, analysis and recommendation module for the system and a system evaluation information center. The test template module is used for a test template for evaluation. The automatic acquisition module for system software-hardware configuration acquires configuration information of the tested system. The automatic system test system is used for relevant testing according to the test template after receiving the configuration information. The self-learning, analysis and recommendation module for the system is used for analyzing test results, generating an optimal virtual-machine quantity deployment scheme for the evaluated system, and generating corresponding performance evaluation results. The system evaluation information center is used for storing the optimal virtual-machine quantity deployment scheme and the performance evaluation results.

The invention discloses a CFA algorithm and BP neural network-based invasion detection method. The method comprises the following steps of: encoding operation parameters of a BP neural network as cell individuals in a CFA algorithm; taking an error function as an adaptive value function of the CFA algorithm; after carrying out iteration for multiple times, selecting the parameter with the optimum fitness as an initial weight value and threshold value of the BP neural network to carry out training; and finally applied the trained BP neural network into a classifier for invasion detection. According to the method, the characteristics of global search and high convergence speed of the CFA algorithm, the initial operation parameters of the BP neural network are optimized, and then the classifier which can be applied to network invasion detection is constructed. Through improving the disadvantages of local minimum and low convergence speed, caused by initial parameter randomization, of the BP neural network, the detection correctness of the BP neural network in the network invasion detection is improved.

The invention discloses a real-time-power-amount smart-prediction method and device of an unmanned vehicle. According to the method, research is carried out for main power consumption aspects such asaspects of air conditioning, wind resistance and the like of vehicle running, complex road environments are combined to establish a power consumption amount prediction model, the selected and used model is obtained on the basis of training of a neural network, self-learning performance is high, and accuracy is good; and power amount prediction of the model is independent of a battery, and is trained in real time, use situations of power amounts in the different road condition environments can be identified, prediction result accuracy is high, timeliness is good, and the problem of inaccuracy caused by calculating power consumption amounts only for basic working principles and charging/discharging characteristics of batteries in the prior art is better avoided.

The invention provides a motor control method based on self-learning of a rotating speed-current two-dimensional fuzzy model. The method comprises the following steps that (1) the dual closed-loop feedback control process is performed to obtain a feedback duty ratio db(t); (2) the feedforward control process of the fuzzy model is performed, wherein mapping of a current grid point p (including speed and current) on a fuzzy curve model S is performed, information of four top points is obtained, and the feedforward duty ratio corresponding to the point p is obtained according to membership and a gravity method; (3) the self-learning process is performed, wherein the feedforward duty ratio of the grid point p at the time (t-1) is amended according to the speed error at the time t, and the information of the four top points around the point p is amended according to the membership and a counter-gravity method. The method effectively gives consideration to stability and rapidness, and is good in self-learning capacity.

The invention relates to the field of air quality monitoring, and particularly relates to a calibration method and system based on a miniature air quality monitoring instrument. The method comprises the following steps: carrying out delivery calibration, in a set range of the monitoring instrument, when a measured concentration C and an electric signal value x are in linear correlation, acquiringa slope k and an intercept b; inspecting and testing delivery consistency: placing the n monitoring instruments near a standard station, in a space range with the same concentration, obtaining a datapair of the measured concentration C and the electric signal value x, and performing unary linear regression analysis on the data pair to obtain a delivery linear equation; acquiring a parameter value: acquiring a correction coefficient between the monitoring instruments; and carrying out field intelligent calibration: placing the reference monitoring instrument near the standard station of a to-be-detected area, generating calibration parameters, calculating data of each monitoring instrument according to a linear equation of the corrected parameters to obtain a result, and re-establishing alinear relationship between the electric signal and the detected concentration. The method is simple, programming is convenient, self-learning is supported, and a calculation speed is high.