65 results about "Time series classification" patented technology

A method and system are provided. The method includes extracting shapelets from each of a plurality of time series dimensions of multi-dimensional time series data. The method further includes building a plurality of decision-tree classifiers, one for each time series dimension, responsive to the shapelets extracted therefrom. The method also includes generating a pairwise similarity matrix between respective different ones of the plurality of time series dimensions using the shapelets as intermediaries for determining similarity. The method additionally includes applying a feature selection technique to the matrix to determine respective feature weights for each of shapelet features of the shapelets and respective classifier weights for each of the decision-tree classifiers that uses the shapelet features. The method further includes combining decisions issued from the decision-tree classifiers to generate a final verdict of classification for a time series dimension responsive to the respective feature weights and the respective classifier weights.

The present invention provides a method, system, and computer program product for detecting memory management anti-patterns. The method comprises: periodically generating census data for at least one object type in a heap; collecting the periodically generated census data for each object type into a time series; analyzing each time series; and classifying each time series into one of a plurality of different categories based on the analysis.

The invention discloses a multi-dimensional time sequence classification method based on mahalanobis distance DTW, and relates to the multi-dimensional time sequence classification method. In order to solve the problems that aiming at satellite telemetry data, a fixed point segmentation effect is non-ideal, due to the facts that relativity exists between multi-dimensional time sequences and small deviation exists between the time sequences, a measuring result is not accurate, therefore a classification result is not accurate, and the multi-dimensional time sequence classification method based on the mahalanobis distance DTW is provided. The method comprises the steps that 1 a multi-dimensional time sequence X={x <1>, x <2>, ..., x<j>, ..., x<n>} used for training and a classification label L={l<1>, l<2>, ..., l<n>}are obtained; 2 a to-be-classified multi-dimensional time sequence X'={x' <1>, x' <2>, ..., x'<j>, ..., x'<n>} is extracted; 3 a DTW distance sequence between the X'={x' <1>, x' <2>, ..., x'<j>, ..., x'<n>} and the X={x <1>, x <2>, ..., x<j>, ..., x<n>} is calculated; 4 classification is conducted on the to-be-classified multi-dimensional time sequence X'={x' <1>, x' <2>, ..., x'<j>, ..., x'<m>} according to neighboring numbers of K which is set by using a KNN classification method based on the mahalanobis DTW distance, and the classification of the to-be-classified multi-dimensional time sequence is determined. The method is applied to the field of multi-dimensional time sequence classification.

The invention provides a classification method based on a kernel feature extraction early prediction multivariate time series category according to early prediction multivariate time series classification. To extract the essential features of variable time series, first the variable time series undergo feature extraction respectively, and a clustering method is adopted to reduce redundancy features, remove noise and improve classification stability; then, to improve classification efficiency, precision and early degree, a method for comprehensively evaluating feature performances is provided on the basis of accuracy rate, recall rate and the early degree and the like, and the optimal feature in each cluster is selected to serve as a kernel feature of a variable; and finally, two simple effective classifier construction methods are provided on the basis of a kernel feature set of each variable. Correctness and effectiveness of the method and an algorithm are proven through experiments, and experiment results prove that a classifier can have high accuracy rate and good early degree.

The invention discloses a Chinese model training, a Chinese image recognition method, a device, equipment and a medium. The Chinese model training method comprises the following steps: obtaining a training handwritten Chinese image; dividing the training handwritten Chinese image into a training set and a test set according to a preset proportion; the training handwritten Chinese images in the training set being sequentially annotated, and the annotated handwritten Chinese images being inputted into the convolution neural network; the time series classification algorithm being used to classifyconvolution neural network into two types, wherein one is long time memory neural network, and the other is short time memory neural network; the network parameters of the long and short time memoryneural network being updated to obtain the original handwritten character recognition model; the original handwritten Chinese character recognition model being tested by training handwritten Chinese images in the test set; when the test accuracy is greater than the preset accuracy, the target handwritten character recognition model being obtained. The Chinese model training method has the advantages of high training efficiency and high recognition accuracy.

The invention discloses a multi-variable time sequence classifying method based on semantic selection. Time sequence classification as a machine learning hot research branch has wide application prospects in computer vision, financial analysis, biostatistics and the like. According to a sparse adaptive-semantic learning network (SA-SLN), by means of a sparse convolution kernel, the space semanticconcept and time short-term dependence of multi-variable time sequences are extracted at the same time, an attention diversion method is put forward to be used for selecting the semantic concept, andfinally modeling is conducted on the long-term dependence of the sequences through LSTMs. By means of the SA-SLN, based on the situation that inter-frame properties have correlation, automatic space feature extraction and time-sequence dependence multi-step modeling are realized, the defects of an existing method are overcome, and the currently optimum result is obtained on three public data sets.

The invention is suitable for the technical field of data pre-processing, and provides a text classification method for different subject topics. The method comprises the following steps of: A, performing word selection on each subject by using a chi-square test to form a feature vocabulary of the subject; B, performing subject classification on the selected feature words by using a Naive Bayes model; and C, performing secondary classification on first two structures with maximum probabilities given in the Naive Bayes model classification by using a support vector machine to obtain a first result. Through two-time classification, the average accuracy of classification is improved. The text classification method for different subject topics is simple to realize and operate, convenient and accurate to use and more accurate in classification between the subjects, and effectively improves the accuracy of classification between the adjacent subjects.

The invention provides a heavy landing analysis method based on a multi-branch time convolutional network. The method comprises the following steps: acquiring original parameter data and a dynamic time point; performing convolution operation on the original parameter data by using the improved time convolution network to generate a feature map of each parameter; performing feature extraction on the feature map to generate overall feature representation; learning a preset category by using the overall feature representation to obtain a parameter level of the preset category and a weight occupied by a feature map of each parameter; and according to the parameter level and the weight occupied by the feature map of each parameter, carrying out linear combination on the feature maps in the overall feature representation to obtain a final class activation mapping map, and according to the class activation mapping map, carrying out analysis on airplane heavy landing. According to the method, a new thought is provided for safety accidents or overrun events in the aviation field, reference is provided for interpretability work of the time sequence classification problem, technical reference is provided for flight safety, and the method has good theoretical and application values.

The invention discloses a convolutional echo state network based time series classification method. An echo state network has a time series core and an echo state property, wherein the time series core refers to that the echo state network maps inputted signals into a high-dimensional space of a reserve pool, and the echo state property refers to that the network has a short-term historical information memory capacity. In the convolutional neural network, multi-scale characteristics in the echo state network can be extracted through a multi-scale convolutional layer, and multi-scale time series invariance can be kept through maximal pooling in time direction. By combination of the echo state network and the convolutional neural network, a convolutional echo state network model is provided;by the model for operations including multi-scale convolution, maximal pooling in the time direction and the like of state represent information outputted by the echo state network, advantage complementation of the echo state network and the convolutional neural network is realized, and high efficiency of an echo state network learning mode is kept while advantages of the convolutional neural network in characteristic extraction are achieved.

The present invention relates to methods for mining real-world databases that have mixed data types (e.g., scalar, binary, category, etc.) to extract an implicit time-sequence to the data and to utilize the extracted information to compute categories for the input data and to predict categorization of future input data vectors. Many real-world databases may not have explicit time data yet there may be inherent time data which may be extracted from the database itself. The present invention extracts such inherent time sequence data and utilizes it to classify the data vectors at each instant in time for purposes of categorizing the data at that time instant. The present invention has wide applicability and may find use in fields such as manufacturing, financial services, or government. In particular, the present invention may be used to identify potential threats, to predict the presence of a threat, and even to evaluate the degree of threat posed. For purposes of this discussion, the threats may be security threats or other adverse events occurring at a particular company, location, or systems, such as a manufacturing or information systems.

The invention discloses a new video semantic extraction method based on a deep learning model. The new video semantic extraction method comprises the following steps: obtaining semantic structured video data by combining and segmenting a video frame sequence on the basis of a video physical structure; using a sliding window to process the semantic structured video data into the input data of a three-dimensional convolutional neural network; creating a three-dimensional convolutional neural network model, and using the output data of the sliding window as training data; using the output resultbased on the three-dimensional convolutional neural network as the input of the continuous time series classification algorithm, and completing the training of three-dimensional convolutional neural network parameters by the backpropagation algorithm; and using the trained three-dimensional convolutional neural network-continuous time series classification algorithm as a sports video semantic extraction model to extract video semantics. The proposed video semantic structuring method is combines with the three-dimensional convolutional neural network and the continuous time series classification algorithm, which can capture the connection between actions and improve the accuracy of sports video semantic extraction.

The invention discloses a service data anomaly detection method based on time sequence classification, which comprises the following steps: s1, extracting offline service data, classifying the offlineservice data according to a time sequence, and generating a sample library comprising different types of time sequences; s2, associating different types of time sequences in the sample library with different time sequence anomaly detection algorithms; s3, acquiring online service data, and classifying the online service data according to the time sequence based on the classification of the time sequence in the sample library; and S4, carrying out anomaly detection on the classified online time series according to an association relationship between time series classification and a time seriesanomaly detection algorithm. The method automatically classifies and identifies different types of time sequences, automatically selects parameters or algorithms to perform time sequence anomaly detection, automatically identifies the types of the time sequences when processing large-scale time sequence anomaly detection, reduces false alarms and missing alarms of alarms, and effectively saves the labor cost.

The invention discloses a time sequence classification method based on time sequence two-dimension and width learning, and the method comprises the following steps: 1), carrying out the normalization processing of input data, and obtaining X; 2) performing feature extraction on X by using four sub-width learning systems and width learning to obtain an input U of the deep cascade width learning system; and 3) classifying the input data U by the deep cascade width learning system. According to the method, width learning and time sequence two-dimensionalization are combined, a feature extraction mode is provided, the features of the time sequence can be fully extracted, the dimensionality reduction effect can be achieved, and data more suitable for a deep cascade width learning system can be extracted. After feature extraction is completed, a deep cascade width learning system is used for classification, so that the time cost and the calculation cost can be remarkably reduced. The invention further provides a time sequence classification system based on time sequence two-dimension and width learning.

The invention relates to a time sequence data classification method based on data feature fragments. The method comprises the following steps of firstly, carrying out subclass division on all sequences in different classes in a specific time sequence training set; meanwhile, selecting a corresponding central sequence; for the center sequence of each subclass, reusing the data trend points to select the data feature fragments of larger weights preferentially according to the weight sorting of the data features to thereby form a data feature fragment set Shapets, and finally finishing the finaltime sequence classification by using a corresponding Shapets conversion algorithm by taking the Shapets set as a core. According to the method, the efficiency of Shapetes generation is greatly improved, and meanwhile, the overall processing efficiency of the algorithm for classifying the time sequence based on the Shapetes is also improved, and the efficiency improvement amplitude is more than three orders of magnitudes.

The invention relates to a multivariate time sequence classification method and system based on full convolution attention, and the method comprises the steps: capturing the local variable features of a multivariate time sequence through employing a 2D convolution filter through employing a design idea of full convolution in the field of images, so as to learn the linkage relation between adjacent variables; capturing local time features of the multivariate time sequence by using a 2D convolution filter to learn trend information between adjacent time, and weakening the influence of mutation information on a result; adopting a convolution and self-attention model, obtaining multiple local features through multi-kernel convolution, enabling the self-attention model to calculate the weights of the multiple local features and non-local features, and providing different visual angles for reviewing multivariate time series data; adopting an attention model to respectively fuse variables and time characteristics of corresponding visual angles, and learning a global dependency relationship of the variables and a global dependency relationship of time; adopting a weight matrix method to fuse multi-view features, and learning more comprehensive and more accurate time variable interaction features.

The invention discloses an MH-LSTM anomaly detection method based on session feature similarity. The method comprises the steps: setting a sliding window to collect Web access data of a user, and carrying out the processing of the Web access data through Min-Hash, and extracting sequence features; training a detection model by using a time sequence classification algorithm of LSTM; and finally, detecting and positioning abnormal users in the captured Web session stream data by using the trained detection model. The method not only can effectively adapt to challenges in a streaming data environment, but also can maintain a high detection rate and a high recall rate.

The invention provides a slope deformation area division method based on dynamic time warping and k-means clustering, and the method comprises the steps: selecting a reference point displacement time sequence through comparing the change conditions of the accumulated displacement of each monitoring point in a monitoring interval; then, moving average smoothing preprocessing is adopted, the change trend of each monitoring point in the monitoring area is extracted, a monitoring point coordinate position set for follow-up classification is screened by setting a threshold value, and therefore the displacement time sequence similarity in the set is calculated through a dynamic time warping algorithm; and finally, establishing a monitoring point displacement time sequence classification model based on a k-means unsupervised clustering algorithm by taking the accumulated displacement in the monitoring interval and the DTW similarity matrix as input characteristics, thereby obtaining a preliminary division result of the slope deformation area under different cluster numbers, and comprehensively evaluating the classification result under different cluster numbers by calculating multiple indexes, and therefore, a simple and efficient classification method is provided for slope deformation area division.

The invention discloses a time series classification method and device. The method comprises the following steps of: aiming at a to-be-classified test time series, determining a first dynamic time regular distance between the test time series and each training time series in a pre-obtained training time series set; aiming at each determined first dynamic time regular distance, determining a first Gaussian function on the basis of the first dynamic time regular distance, and obtaining a Gaussian kernel transformation matrix of the test time series; obtaining a sparse representation coefficient matrix of the test time series according to the Gaussian kernel transformation matrix and a dictionary matrix of the pre-generated training time series set; aiming at each type, determining a residual error between the test time series and the type according to an atom, corresponding to a position of the type, in the dictionary matrix and the sparse representation coefficient matrix; and determining the type corresponding to the minimum residual error as a type of the test time series. By applying the time series classification method and device provided by embodiments of the invention, the time series classification accuracy can be improved.

The invention relates to a financial time series classification method for abnormal financial account detection, and application thereof. According to the method, a financial time sequence data set ofa financial account can be constructed and expanded from transaction flow data of an abnormal financial account and a normal financial account; a neural network model with a plurality of stacked Blocks (wherein each Block contains Local BiLSTM, Self-Attention, residual connection, Layer Normalization, Position-wise Feed-Forward Networks) is used to simultaneously extract local and global mode characteristics of a sequence from a financial time sequence; and finally, a softmax classification layer is used to classify the financial time sequence to achieve the detection function of abnormal financial accounts.

A time series classifying memory includes an enumerated group of synapses. Each synapse of the group has a junction including pre-synaptic emitters communicating with post synaptic receptors. A single pathway is input to each synapse of the group, and when innervated, stimulates the junctions. Each successive synapse has successively more post synaptic receptors according to a fixed ratio. At each junction, quanta of neurotransmitter emitted from the emitters bind with available receptors. When all the receptors at a junction have been bound, the junction goes refractory. The synapse adjacent to the last junction to go refractory is marked. During a training mode the mark is a Long Term Potentiation (LTP) mark. During a live mode the mark is a Short Term Potentiation (STP) mark. The input is classified upon successful correlation of the LTP mark with the STP mark. A clock operable with synaptic matrix models this process.

The invention relates to a time sequence classification method, system, medium and device based on multi-representation learning. The method comprises the following steps: (1) carrying out the multi-feature coding of a given time sequence based on different time sequence representation strategies; (2) realizing representation fusion and enhancement by using a residual network and a bidirectional long-short-term memory network; and (3) completing classification by using a multi-layer perceptron network, and realizing classification interpretability by using an attention mechanism. According to the method, a multi-channel time sequence representation learning model is constructed, so that the time sequence characteristics can be comprehensively understood based on various representation strategies. According to the representation fusion model based on the residual network and the bidirectional long and short term memory network, multi-view representation can be effectively fused and representation enhancement can be realized, so that the classification precision is effectively improved. According to the method, the important time sequence characteristics of the time sequence can be effectively recognized on the basis of the attention mechanism, namely, the interpretability basis of the classification result can be provided, and the classification interpretability is realized.

The invention discloses a time sequence classification and retrieval method based on deep multi-task representation learning. The method comprises the steps: (1) performing normalization processing ona given time sequence data set; (2) inputting the normalized time sequence into a multi-stage wavelet decomposition network, and obtaining a multi-scale subsequence set through n-stage decomposition;(3) inputting the multi-scale sub-sequence set into a residual network, and outputting shared representation after feature extraction and fusion are performed on each sub-sequence; (4) as for a timesequence classification task, inputting the shared representation into a classification network, outputting a classification representation after feature extraction, then inputting the classificationrepresentation into a classifier, and outputting a time sequence classification result after classification; and (5) as for the time sequence retrieval task, inputting the shared representation into aretrieval network, outputting the retrieval representation after feature extraction, and as for the retrieval representation, realizing the retrieval task by constructing a tree-based or hash-based index.

The invention discloses a financial time series prediction method, a server and a device. The financial time series prediction method comprises the following steps: establishing a financial time series prediction model by using a feature subset of the financial time series, a time window of the financial time series, and a financial time series classifier as three populations of a roulette cooperative co-evolutionary algorithm and finding the optimal value of the three populations; inputting the financial time series into the financial time series prediction model and outputting a prediction result. The financial time series prediction method improves the accuracy of financial time series predictions.

The invention discloses a time sequence classification method based on deep reinforcement learning, and the method comprises the following steps: collecting a plurality of time sequences, obtaining sample data, and carrying out the preprocessing of the sample data; constructing a deep residual network, and updating the deep residual network through a deep reinforcement learning method according tothe preprocessed sample data; and inputting a to-be-tested time sequence into the updated deep residual network to obtain a classification result of the time sequence. According to the method, the samples are input into the deep reinforcement learning network in a disordered sequence, so that the deep reinforcement learning network is more robust, the optimal strategy of time sequence classification is searched in a reward and punishment setting mode, and the classification accuracy is high.

The invention discloses a time series classification method based on an improved spiking neural network. The method comprises the following steps of: intercepting a one-dimensional time sequence signal from original time sequence data, converting the one-dimensional time sequence signal into a two-dimensional texture image by using an improved recurrence plot, constructing an overall structure framework of an impulsive neural network, and reasonably zooming the two-dimensional texture image by using a down-sampling layer according to the size of the two-dimensional texture image; inputting thescaled two-dimensional texture image into an input layer of a network, and converting the two-dimensional texture image into a Poisson pulse sequence; determining the initial learning rate of the network and the maximum value and the minimum value of the boundary of the cyclic learning rate by using a cyclic learning rate method; continuously updating connection weights among neurons of the network by using a pre-synaptic and post-synaptic trace learning rule; and after network training is completed, classifying the time series by counting pulse triggering conditions of excitatory neurons.

The present invention provides a method, system, and computer program product for detecting memory management anti-patterns. The method comprises: periodically generating census data for at least one object type in a heap; collecting the periodically generated census data for each object type into a time series; analyzing each time series; and classifying each time series into one of a plurality of different categories based on the analysis.

The invention relates to a method, a device and a system for automatically adding new labeled time sequence samples of remote sensing images. According to the method, firstly, every unclassified and labeled sample in an unclassified and labeled sample set is sorted into a candidate set of a corresponding category according to the dynamic time warping distance between the corresponding unclassified and labeled time sequence sample and category centers of the classified and labeled time sequence sample set; secondarily, samples which are most distinguished from the category centers of different categories in the classified and labeled sample set are selected from the candidate sets in different categories and labeled. In the iterative new sample adding process, the diversity of the sample sets can be sufficiently guaranteed, high-quality training data can be provided for remote sensing big-data monitoring and scene recognition and other application scenes, and the implementation process is completely automatic, efficient and stable.

The invention provides a time sequence classification method based on LSTM and multi-scale FCN, belonging to the field of time sequence classification. The method comprises the following steps: setting a general structure of a multi-modal network; extracting time-dependent characteristics by using a long-short-term memory network; fully mining the geometric space characteristics of multiple granularities of a time sequence curve by using a full convolution module; integrating spatio-temporal characteristics and discriminating a sample according to the integrated spatio-temporal characteristics; and fully training a model by using a back propagation algorithm. According to the method, spatial features can be comprehensively explored through a large-scale and multi-scale receptive field, long-term and short-term dependency can be adaptively learned, and learned beneficial information is more comprehensive compared with an existing model. More accurate judgment can be given by more comprehensively mastering the difference characteristics of time sequences.

The invention relates to the technical field of data processing and provides a time series classification method and device based on characteristic sampling. The method comprises the following steps:converting time series data sets for training into training data sets with an equal-length characteristic with a characteristic sampling method, and converting the time series data sets for testing into test data sets with an equal-length characteristic; performing model training by using the training data sets with the equal-length characteristic with an integrated learning classification method;performing time series classification on the test data sets with the equal-length characteristic by the trained model. The time series data sets with different lengths are first converted into the data sets with the equal-length characteristic with the characteristic sampling method, and then classification is performed with the integrated learning classification method, so that accuracy of timeseries classification is improved, and large-scale time series data can be accurately classified.

The invention discloses a time sequence classification method which comprises the following steps of optimizing a GDTW kernel function in advance, and obtaining an improved GDTW kernel function; by means of the improved GDTW kernel function, performing kernel transformation on a preset time sequence training sample set and a time sequence testing sample; and performing classification on kernel transformation data of the testing sample by means of a preset classification algorithm and a time sequence kind label in the kernel transformation data of the training sample, thereby obtaining kind of the time sequence testing sample. According to the time sequence classification method, in calculating an Euclidean distance between two time sequence elements by the improved GDTW kernel function, the Euclidean distance between two time sequence elements which satisfy an optimal offset path is calculated so that the improved GDTW kernel function keeps the offset path information between the time sequences, thereby further improving a subsequent classification effect. Additionally, the invention correspondingly discloses a time sequence classification system.