45 results about "Sequence learning" patented technology

The invention discloses a multi-scale time-frequency network sleep stage classification method based on single-channel electroencephalogram, and belongs to the technical field of medical signal processing. Aiming at the problems of capturing multi-scale time-frequency characteristics, processing variability among subjects and modeling long-range time dependence of a sleep staging method, the method comprises the following steps: firstly, acquiring electroencephalogram signal data, and performing time sequence segmentation, sleep stage category label determination and standardized preprocessing by taking a single-channel electroencephalogram signal as an analysis object; performing data set division by adopting nested cross validation, and constructing a multi-scale time-frequency network model; the model comprises a feature extraction module and a sequence learning module, wherein the feature extraction module comprises a time domain branch and a frequency domain branch; a subject adaptive feature calibration module is proposed to dynamically compensate the influence brought by individual difference and signal quality fluctuation; respectively training a feature extraction module and a sequence learning module by adopting a component type training strategy; and inputting to-be-classified electroencephalogram signal data into the trained model, and outputting a corresponding sleep stage classification result.

The invention relates to the technical field of physiological signal processing and sleep monitoring, in particular to a sleep staging method and a sleep staging system for collecting multiple physiological signals, and the sleep staging method and the sleep staging system for collecting the multiple physiological signals synchronously collect auditory meatus photoelectric volume pulse waves, temperature and head micro-motion signals through an in-ear sensor array. According to the method, the signal quality index is calculated, time domain, frequency domain and nonlinear features are extracted, a dynamic weighted fusion mechanism is adopted, feature weights are adjusted according to the signal quality index, a hierarchical depth time sequence learning model is input for sleep staging, and the sleep staging accuracy is improved to 89% or above and is improved by 15-20% compared with a single brain wave method. A sleep state evaluation report and a personalized feedback intervention strategy generated by the system are beneficial for improving sleep quality, a dynamic weighted fusion mechanism enhances system robustness, adapts to different signal qualities and ensures stable performance, and the invention provides an efficient and accurate new method for the field of sleep monitoring.

A long-term continuous learning method based on task core memory management and consolidation aims to enable a model to sequentially learn from a large number of task sequences, new knowledge is obtained, information of previous learning is reserved, and the method is similar to a human learning mode. The method comprises the following steps: 1) task input and instruction fine tuning; 2) performing difference analysis on the model parameters of the current task and the previous task, identifying a task core memory unit, calculating an adaptive weight based on task prototype similarity, and dynamically updating the memory unit; 3) constructing an experience playback buffer area through a difficult sample selection strategy and a difference sample selection strategy; and 4) utilizing the joint loss function training model to keep the memory of the historical tasks while learning the new tasks. According to the method, the problem of disastrous forgetting in long-term continuous learning is mainly solved, and the performance of the model in a long-term sequence task is remarkably improved.

The invention is suitable for the technical field of battery safety monitoring, and provides a sodium ion battery energy storage system health state dynamic prediction method based on simulation digital twinning, and the method comprises the steps: constructing a digital twinning body comprising a physical layer, a virtual layer and a data interaction layer; performing health state estimation based on the physical layer and the virtual layer, and outputting a health state estimation value; comparing the health state estimated value with the measured value, calculating a residual error, and updating key aging parameters in the simulation model; inputting the updated key aging parameters into a pre-trained time sequence neural network model, carrying out sequence learning, and outputting a capacity attenuation trend and a residual life prediction result of the battery; generating a full life cycle optimization strategy based on the prediction result, and generating an optimal charge-discharge power table by simulating the influence of different charge-discharge strategies on the life; the problems of life attenuation and high operation and maintenance cost caused by insufficient prediction precision in the prior art are effectively solved.

The invention discloses a multi-vehicle cooperative three-dimensional target detection method for a heterogeneous intelligent network connection vehicle group, and the method comprises the following steps: carrying out the preprocessing of obtained multi-source heterogeneous vehicle-mounted sensor data, and extracting the features of multi-mode heterogeneous data; constructing a dynamic pose compensation network, learning and predicting a pose compensation amount by using a time sequence, correcting a pose error in multi-vehicle cooperative perception, and ensuring a space reference of multi-vehicle features; performing cross-modal feature fusion on the multi-vehicle multi-modal heterogeneous data features through a global-local cross-modal attention mechanism; designing a multi-scale feature distillation strategy, and reducing the semantic fusion gap of the cross-modal features; an end-to-end combined training framework is constructed to integrate dynamic pose compensation, cross-modal feature fusion and multi-scale feature distillation strategies, and a multi-vehicle collaborative three-dimensional target detection model is optimized through an adaptive weighted loss function. The technical problems of multi-vehicle heterogeneous sensor fusion, pose drift correction and cross-modal semantic alignment are solved.

The invention provides a multi-modal sentiment analysis method based on graph-attention collaborative optimization cross-modal recombination, and relates to the technical field of multi-modal sentiment analysis. The method comprises the following steps: firstly, designing a modal self-adaptive multi-modal graph construction module, constructing a local sparse graph based on KNN-RBF for a language modal, and adopting a low-rank representation method combined with nuclear norm regularization for an audio and video modal; secondly, the processed modal features are transmitted into a graph attention network to realize high-order feature aggregation; then, a language-guided hierarchical cross-modal interaction mechanism is constructed, and multi-granularity semantics are accumulated in combination with an advanced multi-modal feature container module; and finally, designing an advanced feature recombination strategy based on dynamic matching, and realizing feature alignment by taking a language feature container as an anchor point. According to the method, graph learning and sequence learning are unified in a collaborative framework through a graph-attention collaborative optimization cross-modal recombination model, so that the problems of cross-modal attention noise interference, modal imbalance, insufficient cross-modal feature alignment efficiency and the like can be effectively solved.

The embodiment of the invention provides an ATE embedded wafer defect detection method and device based on a dynamic electrical sequence RNN, and the method comprises the steps: obtaining original electrical parameter data from an ATE test process, and carrying out the preprocessing of the original electrical parameter data, and determining quality parameter data; constructing sequence data based on the quality parameter data; constructing an initial depth sequence learning model, training the initial depth sequence learning model based on the sequence data, and determining a target depth sequence learning model; wherein the deep sequence learning model is an RNN architecture; and acquiring current wafer test data, and determining defect information based on the current wafer test data and the target depth sequence learning model. Through the deep dynamic analysis of the electrical parameter sequence, the potential failure risk and process abnormality can be identified earlier and more sensitively before the defect feature of the Die is not completely shown or is judged to be qualified in the traditional test, and the capture rate of early defects is improved.

The invention relates to the field of electric power engineering, and particularly discloses a port shore power load double-flow LSTM prediction method, which comprises the following steps: S1, selecting and collecting factors influencing power load data by using longitudinal data to obtain a multivariable time sequence data set D1, and preprocessing the data set D1 to obtain a data set D2; s2, correlation analysis is carried out, and dominant features and auxiliary features are divided based on the correlation degree of each variable and the load in the MIC quantitative data set D2; respectively inputting the dominant features and the auxiliary features into heterogeneous LSTM branch modeling, and splicing output results to generate a fusion feature map; and S3, constructing a BO-LSTM neural network, inputting the fusion feature map into a double-flow time sequence learning module, extracting dominant features and auxiliary features, performing deep representation, performing splicing, introducing a channel attention mechanism to perform weighting processing on fusion feature vectors, and outputting a power load prediction value through a residual error correction module. According to the method, the prediction precision and robustness are remarkably improved, and the real-time scheduling of the port shore power system is supported.

The invention relates to the technical field of sea flux monitoring, in particular to a pollutant sea flux monitoring and predicting method based on deep learning. The method comprises the following steps: collecting historical flow and pollutant concentration data of a target area according to different time intervals, and locking an area to be analyzed according to the historical flow and pollutant concentration data; then obtaining a regional water sample, extracting concentration and flow data of nitrite, ammonia nitrogen and nitrate, and constructing a training set; after the training set and the real-time flow test set are standardized, deep time sequence learning is carried out through a TCN model, and the future flow and the concentration of each pollutant are predicted; and finally, calculating the into-sea flux based on a prediction result, and realizing high-precision pollutant into-sea flux monitoring and forecasting. According to the method, deep causal convolution and residual feature extraction are performed on long-time sequence data by using the time sequence convolution network, so that the prediction precision and efficiency of the pollutant into-sea flux under complex terrains, extreme climates and human activity dense areas are improved.

A hydraulic loading system fault diagnosis method based on sequence learning comprises the steps that firstly, sample data and corresponding labels are imported, and the imported sample data are flattened and then subjected to normalization processing; then, a three-layer one-dimensional CNN convolution structure is adopted, each layer comprises convolution, batch normalization, ReLU activation function and maximum pooling operation, and a feature matrix is obtained; then, inputting the feature matrix into a single-layer one-way GRU network to capture a dynamic time sequence dependency relationship, and taking a final hidden state of the dynamic time sequence dependency relationship as a global feature representation; and finally, the global features are mapped to a fault category space by a full connection layer, and a Softmax activation function is used to complete fault classification. Cross entropy is adopted as a loss function, the normalized data sample is used for training, and when the accuracy of the verification set reaches a design value or reaches the maximum training round number, the training is ended; according to the method, the automation degree and reliability of fault diagnosis of the hydraulic loading device are remarkably improved by fusing the advantages of the CNN in feature extraction and the GRU in time sequence modeling.

Key phrase generation using indefinite sequence learning is described. In accordance with the described techniques, a sequence generation model generates a sequence of key phrases based on an input document. During the generation task, the sequence generation model omits use of a self-generated sequence termination token. Key phrases in the sequence are then output as recommended key phrases for the input document.

The application discloses a kind of communication power supply operating state intelligent monitoring methods based on machine learning, comprising the following steps: S1, obtains multiple-source monitoring data and pre-processes;S2, sequence segmentation is carried out using sliding time window and statistical feature vector is extracted, forms state variable set;S3, each state variable is encoded to construct state node and establish weighted directed connection, and construct operating state evolution diagram;S4, state transition matrix is constructed, main transition feature is extracted and low rank is approximately reconstructed, sequence learning is carried out to construct state update function;S5, to state node, multiple-step recursion is carried out, and the Euclidean distance between recursive state and risk boundary is calculated, to determine state monitoring result;S6, error is calculated and state update function parameter is iteratively updated.The application can realize the multiple-step recursion prediction and risk trend identification of communication power supply operating state, improve the accuracy and stability of communication power supply operating state monitoring.

The invention discloses a field exercise and training medical service equipment management system based on a block chain, and relates to the technical field of field emergency medical equipment management and power supply safeguard.The viability of field medical service equipment in an extreme environment is remarkably improved through deep coupling of a block chain architecture and an intelligent algorithm; the dynamic analysis unit adopts an online sequence learning model to compress and analyze the window size according to the real-time altitude and power supply quality index, so that the capture speed of a plateau voltage sag event is increased to millisecond level, and the downtime risk caused by instantaneous power failure of key equipment such as a cardio-pulmonary resuscitation machine is thoroughly avoided; on the premise of protecting data privacy, the federal aggregation mechanism fuses fluctuation characteristics of a plurality of generators, the generated prediction model can adapt to different altitude intervals, and the problem of cross-scene failure of a single model is solved.

The application belongs to the field of medical prediction, and discloses an information mining method for heterogeneous time series data, comprising: acquiring electronic medical record data and constructing a hypergraph, analyzing and calculating the hypergraph to obtain embedding representation data, weighting the embedding representation data based on an attention mechanism to obtain embedding sequence data, constructing a sequence learning model and performing hidden state access to obtain hidden representation data and weight data thereof, weighting the embedding sequence data to obtain embedding sequence hidden data; training the sequence learning model through time training parameter data, weighting the embedding sequence hidden data through the trained sequence learning model to obtain time dimension hidden data, constructing a full connection network to analyze the time dimension hidden data to obtain medical event prediction data. The technical scheme disclosed by the application can learn complex information in the time dimension by using time step information, and can obtain accurate medical event prediction results.

The invention discloses a two-dimensional-to-three-dimensional attitude skeleton reconstruction system and method based on sequential prediction and occlusion semantic compensation, and the method comprises the steps: obtaining skeleton joint points from two-dimensional image data as a plurality of two-dimensional feature points, and building an initial three-dimensional attitude skeleton according to a skeleton model and a projection inversion technology; thirdly, inputting historical frame data into the sequence learning model to output a predicted attitude, calculating a deviation value between the predicted attitude and an initial three-dimensional attitude skeleton, automatically correcting the stability of the three-dimensional attitude when the detected deviation exceeds a preset threshold value, and automatically correcting the stability of the three-dimensional attitude when two-dimensional feature point missing is found. And finding out the matched similar attitude to estimate and complement the missing feature points of the initial three-dimensional attitude skeleton, and immediately generating a complete and accurate final three-dimensional attitude skeleton, so as to achieve the technical effect of improving the stability and naturalness of the reconstruction of the three-dimensional attitude skeleton.

The invention discloses an artificial neural network non-sample incremental learning method based on a sleep mechanism, and relates to the technical field of artificial intelligence. The method comprises the following steps: training an artificial neural network through a data set of a target category task to obtain an artificial neural network in a waking learning stage; replacing an activation function of the artificial neural network after the waking learning stage with a Heaviside step function, periodically adding Poisson noise to a random square of an average image of the data set, and inputting the random square into the artificial neural network; on the basis of the replaced Heaviside step function, neurons in an activated state are determined; according to the neurons in the activated state, the connection weight between the neurons is updated based on the Hertz rule; and recovering the original activation function to obtain the artificial neural network in the sleep playback stage, namely the artificial neural network after incremental learning of the target class tasks is completed. According to the method, the problem of disastrous forgetting generally existing when the neural network faces multi-task sequential learning is relieved.

The invention discloses an Internet marketing data processing method and system based on artificial intelligence. Comprising the steps that S1, a marketing data set is acquired, the marketing data set comprises the data distribution amount of a plurality of different blocks, and the blocks are different types of internet platform areas of user interaction types; s2, acquiring a user data set, wherein the user data set is feedback data of a user to the data distribution amount in different blocks; s3, establishing mapping between the marketing data set and the user data set, and encrypting the mapping; and S4, establishing a sequence learning model based on the encrypted mapping, and generating interaction prediction data. The invention provides a mapping relationship between marketing data and user data, and can also provide an interactive prediction mode for predicting the performance of the model.

The invention discloses a digital refocusing method applied to a microscope, which belongs to a computer vision technology in the field of information, and is technically characterized in that aiming at an out-of-focus image which is not in a focusing range, registration of a focusing out-of-focus image pair can be utilized. Further, inputting into a u-net network based on deep learning for virtual refocusing, taking a focused image as a label image for supervision, performing down-sampling feature extraction and up-sampling feature recovery and learning on the input out-of-focus image through a network layer, and performing loss function calculation and back propagation to optimize the network step by step, so as to obtain an out-of-focus image; therefore, the network can learn the mapping relation between the out-of-focus image and the in-focus image. The method has the advantages that under the condition that an existing system of the microscope is not changed, digital refocusing is carried out on an image with serious positive and negative defocusing, a focused image can be obtained only by using the neural network, and the refocused image and a real focused image have high numerical values in the aspects of SSIM and PSNR. In addition, the method has no special requirements on hardware environment conditions, and only part of data sets need to be acquired and processed in advance.

The present application relates to a kind of lightning intensity prediction method and system based on deep time sequence learning, belong to lightning early warning technical field, solve the problem of insufficient fine prediction ability of lightning intensity in prior art, peak time positioning is not accurate.The method comprises: obtaining current atmospheric electric field intensity data, using peak detection algorithm, determine the current peak position;According to the current atmospheric electric field intensity data, determine the current intensity time sequence data;Current peak position, and current intensity time sequence data, input lightning intensity prediction model, determine the first predicted peak position and the first predicted intensity peak;Wherein, lightning intensity prediction model includes time feature fusion layer, encoder layer and double-head output layer.The accurate timing and quantitative prediction of sparse, sudden lightning intensity peak are realized.

The invention provides an intelligent triage priority evaluation method and system based on deep time sequence learning, and relates to the technical field of medical triage, and the method comprises the steps: constructing a feature vector through processing patient registration, symptom text and physiological data, constructing a patient relation graph based on a symptom and department relation, and generating a node enhancement feature through a graph attention network. In combination with the historical treatment sequence of the patient, a time sequence enhancement priority feature is generated through a variational encoder, a priority score is obtained through multiple times of random inactivation deduction, a waiting queue is dynamically reordered according to the priority score, a consulting room distribution instruction is generated, and accurate and efficient evaluation of the triage priority is achieved.

The invention relates to the technical field of network communication and information processing, discloses a multi-dimensional information network intelligent identification and accurate data packet capture method and system, and aims to solve the problems of single identification dimension, redundant packet capture, difficulty in adapting to a dynamic network and the like in the prior art. Comprising the steps of collecting flow multi-dimensional feature extraction, and generating a time sequence feature matrix; training a sequence learning model to construct a transaction fingerprint database; comparing the dynamic fingerprint with a fingerprint database in real time, and identifying a specific network transaction; and after a specific transaction is identified, generating a dynamic packet capture rule and executing accurate data packet capture. According to the scheme, deep intelligent identification of network transactions is realized, 'transaction fingerprints' are formed, a dynamic network is effectively handled, the problems of large packet capture data volume and low signal-to-noise ratio are solved, and the efficiency is improved.

The invention discloses a plateau railway icing early warning method and system based on sequence learning, and solves the problem of data labeling when no direct icing record exists by collecting and preprocessing daily scale meteorological parameter data and applying an innovative method based on meteorological experience criteria to generate icing labels according to multiple meteorological parameter thresholds and logic judgment. A layered self-attention mechanism is utilized, meteorological data features are extracted by a meteorological feature layer, cross-day meteorological dependence features are extracted by a time layer through sliding window self-attention, observation point geographic association feature mining is realized by fusing longitude and latitude position coding in a space layer, daily scale data are efficiently adapted, and extra terrain information is not needed; and constructing a time-space fusion icing early warning model to output an icing probability, and setting a dynamic threshold value based on observation point clustering to carry out graded early warning decision making. According to the method, the precision and efficiency of icing early warning of the plateau railway can be remarkably improved, and a powerful guarantee is provided for safe operation of the railway.

The present disclosure relates to a learning system that sequentially learns a prediction model for a manufacturing device which has different product requirements. This learning system comprises: at least one processor; and a storage device that retains, as learning data, a parameter which uses a probability distribution to express a learning value calculated on the basis of data acquired during manufacturing and the timing at which the parameter was updated last. The at least one processor uses a prediction model corrected on the basis of the learning data to calculate a set value for controlling a manufacturing device, acquires a manufacturing performance value corresponding to a prediction value of the prediction model, computes a parameter expressing a performance value probability distribution on the basis of at least one acquired performance value, updates the learning data, and stores the learning data in a storage device.

The present application relates to the technical field of sea flux monitoring, and particularly relates to a pollution sea flux monitoring and prediction method based on deep learning. The method comprises the following steps: collecting historical flow and pollutant concentration data of a target area at different time intervals, and locking the area to be analyzed accordingly; then obtaining water samples in the area, extracting nitrite, ammonia nitrogen, nitrate concentration and flow data, and constructing a training set; after standardizing the training set and real-time flow test set, performing deep time sequence learning with a TCN model to predict future flow and pollutant concentration; and finally calculating the sea flux based on the prediction result to realize high-precision sea flux monitoring and prediction of pollutants. The present application improves the prediction accuracy and efficiency of sea flux of pollutants in complex terrain, extreme climate and densely populated areas by using time sequence convolution network to perform deep causal convolution and residual feature extraction on long time sequence data.

The invention relates to the technical field of signal processing, and provides an electron capture detector baseline drift correction method, device, equipment, medium and product, and the method comprises the steps: collecting a baseline signal, carrying out the wavelet feature extraction of the baseline signal, and constructing a drift time-frequency feature vector; the drift time-frequency feature vector is input into a drift prediction model, a baseline drift prediction sequence in a future correction period is obtained, and the model structure is established based on a gating circulation unit; and collecting the mixed signal in a future correction period, and performing baseline drift correction processing on the mixed signal based on the baseline drift prediction sequence to obtain a corrected signal. According to the baseline drift correction method for the electron capture detector, wavelet feature extraction and gating circulation unit prediction are fused, the advantages of time-frequency feature extraction and long time sequence learning can be combined, baseline complex drift can be processed in a targeted mode, and the adaptability to complex drift, long time sequence stability and weak signal protection are effectively improved.

The invention discloses a continuous learning image classification method based on prompt memory dynamic alignment, and belongs to the technical field of artificial intelligence and computer vision. According to the method, on the basis of a pre-trained visual Transform backbone network, continuous prompt memory, momentum updating, optimal transmission alignment and an attribution perception fusion mechanism are introduced, and stable storage, dynamic alignment and progressive fusion of cross-task prompt knowledge are realized in combination with class-level value memory and vMF geometric constraints. In a class increment scene of multi-task sequential learning and no sample playback, on the basis of Split CIFAR-100 data set verification, on the premise of controlling calculation and storage overhead, catastrophic forgetting can be effectively relieved, the old task keeping ability and the new task adaptation ability are improved, and therefore long-term image classification performance and stability are remarkably improved, and the method has the advantages of being high in practicability and easy to popularize. And the method has relatively low parameter overhead and good engineering expandability, and is suitable for various practical application scenes.

This invention discloses a multi-condition intelligent monitoring method and system for industrial processes based on continuous learning of element-aware dictionaries. It mainly comprises two parts: continuous learning of new modes and online monitoring. Continuous learning of new modes involves first sequentially learning the optimal dictionary and importance matrix corresponding to the first N modes; then, constructing an objective function based on the reconstruction error of the (N+1)th mode monitoring data and the loss of historical modes, and learning the optimal dictionary for the (N+1)th mode; finally, using this optimal dictionary for the latest mode, a monitoring model based on control limits is established. Online monitoring utilizes control limits to monitor new online data and determine whether the industrial process is currently malfunctioning. This invention can continuously learn new modes while maintaining a memory of historical modes, overcoming the "catastrophic forgetting" problem of traditional process monitoring methods in multi-modal situations, and achieving a high level of monitoring accuracy.

The application provides an energy storage type hydraulic wind turbine super-short-term output power prediction method, relates to the field of intelligent prediction of machine learning, and comprises the following steps: S1, collecting wind speed historical data; S2, obtaining a wind speed prediction value by adopting a CEEMDAN-LSTM combined model; S3, constructing an optimized WPC model; S4, obtaining a wind turbine output power prediction value; S5, constructing a hydraulic system efficiency physical model to obtain a hydraulic system transmission efficiency theoretical value; S6, constructing a hydraulic system transmission efficiency model to obtain a hydraulic system transmission efficiency prediction value; S7, obtaining a mechanical power prediction value of a hydraulic motor output shaft; S8, constructing a hydraulic energy storage prediction model; and S9, obtaining a final output power prediction value according to a PINN-LSTM. The application realizes high-precision, strong generalization and interpretable super-short-term output power prediction by introducing physical consistency constraints and a state feature sequence learning mechanism of machine learning.

This invention provides a method, device, electronic device, and storage medium for predicting the lifespan of a crane transmission system, belonging to the field of crane maintenance technology. The method includes: acquiring a time series of health indicators reflecting the overall degradation trend of the crane transmission system, the time series being constructed based on vibration signals collected from the transmission system under variable load and variable speed conditions; inputting the health indicator time series into a pre-trained residual nested long short-term memory attention prediction network; the residual nested long short-term memory attention prediction network being composed of stacked residual blocks; and outputting predicted remaining service life values ​​of key components of the transmission system through sequence learning and attention weighting processing of the residual nested long short-term memory attention prediction network. This invention improves the accuracy of crane transmission system lifespan prediction.

Methods and systems for detecting water quality may involve: classifying water quality using a water meter, incorporating data indicating changes in ultrasonic time-of-flight (ToF) behavior due to impurities or combinations of impurities mixed in the water; and classifying impurities in the water using a sequence learning unit. Data indicating changes in ultrasonic time-of-flight behavior can be obtained from one or more ultrasonic sensors associated with the water meter. The sequence learning unit can classify impurities in the water into water quality parameters, including one or more of the following: for example, TDS (total dissolved solids), pH value, residual chlorine data, turbidity information, and total organic carbon value. The data can be transmitted to the user via radio frequency frames.