170 results about "Autocorrelation matrix" patented technology

The invention discloses a remote sensing image target detection method based on an accelerated projection recurrent neural network. The method comprises the following steps: acquiring an original remote sensing image and calculating a remote sensing image autocorrelation matrix; constructing a linear constraint optimization mathematical model of filtering output based on the remote sensing image autocorrelation matrix, and obtaining a linear equation mathematical model through conversion; and solving the linear equation mathematical model by adopting an accelerated projection recurrent neural network algorithm, carrying out vector decomposition on a solving result to obtain a filtering coefficient vector, obtaining a remote sensing image after filtering output based on the filtering coefficient vector, and completing remote sensing image target detection. According to the remote sensing image target detection method based on the accelerated projection recurrent neural network, the convergence speed is higher, the calculation time is shorter, the image threshold classification capability is higher and the target detection precision is higher when the optimization problem is solved.

The invention provides an Internet of Vehicles broadcast frame radio frequency fingerprint identification method based on LMMSE channel estimation. The method comprises the following steps: obtaining subcarrier data in a resource grid based on an obtained signal of a physical side link broadcast channel; performing root-mean-square delay expansion and channel autocorrelation matrix construction based on the subcarrier data, and combining priori signal-to-noise ratio regularization and time domain windowing operation to obtain a channel estimation value; initial radio frequency fingerprint features are obtained through the channel equalization and the channel estimation value, and the improved neural network and the initial radio frequency fingerprint features are utilized to carry out classification identification on the Internet of Vehicles equipment. Influences of noise and channels on fingerprints are effectively considered, and the purpose of extracting fingerprints of different devices in a complex environment more accurately is achieved.

The invention provides a video anomaly detection method and system based on time interaction entropy, and the method comprises the steps: obtaining an original video stream, segmenting the original video stream into video clips, generating a time disturbance sample through the video clips, carrying out the feature extraction through a feature extraction module, and solving a mean value, so as to obtain a video feature sequence, processing the video feature sequence by using a time structure entropy calculation module to obtain a time structure entropy sequence, and processing the semantic feature by using an attention modulation module to obtain a structure entropy modulation attention output feature; and processing the global time volatility priori descriptors and the structure entropy modulation attention output features by using an exception scoring module to obtain exception scores. According to the method, lightweight design is adopted, a parameter-free time structure entropy calculation module is designed, and through immediate feature centralization and time autocorrelation matrix calculation in a sliding window, trainable parameters are not needed, so that the calculation overhead is remarkably reduced.

The invention discloses a fault identification method and device for a rotating mechanism, computer equipment and the rotating mechanism. The method comprises the following steps: acquiring rotating speed information and a time domain vibration signal of the rotating mechanism; converting the time domain vibration signal into an angular domain vibration signal based on the rotating speed information; performing feature extraction on the angular domain vibration signal based on the autocorrelation matrix of the angular domain vibration signal and the filter coefficient to update the filter coefficient and obtain a target filter coefficient; and determining fault characteristics of the rotating mechanism based on a target vibration signal obtained by processing the angular domain vibration signal based on the target filtering coefficient. According to the method provided by the invention, the fault features of the unsteady vibration signals of the rotating mechanism can be effectively extracted, so that reliable fault diagnosis of the rotating mechanism is realized.

The invention discloses a power distribution network power prediction method, system and device based on multi-source heterogeneous data fusion, and a storage medium, and relates to the technical field of power distribution network data processing and prediction, and the method comprises the steps: constructing a data autocorrelation matrix based on the comprehensive operation and maintenance data of a power distribution network; performing eigenvalue decomposition on the data autocorrelation matrix, and constructing a principal component matrix; performing dimension reduction processing on zero-mean data obtained in the process of constructing the data autocorrelation matrix by using the principal component matrix to obtain operation and maintenance feature fusion associated data of the power distribution network; an echo state network model is established and trained, the operation and maintenance feature fusion associated data of the power distribution network is used as the input of the echo state network model, and the output power of the power distribution network is predicted; according to the method, the data dimension is reduced, the calculation complexity is reduced, and the output power of the power distribution network can be predicted more accurately; the operation efficiency and the management level of the power distribution network can be improved, and powerful support is provided for stable supply of a power system.

Coding including dynamic range handling of high dimensional inverse autocorrelation in optical flow refinement includes obtaining a refinement model from available warped refinement models, wherein the available warped refinement models include a four-parameter scaling refinement model, a three-parameter scaling refinement model, and a four-parameter rotational refinement model, obtaining refined motion vectors using the warped refinement model and previously obtained reference frame data in the absence of data expressly indicating the refined motion vectors in the encoded bitstream, wherein obtaining the refined motion vectors includes using a dynamic range adjusted autocorrelation matrix, generating refined prediction block data using the refined motion vectors, generating reconstructed block data using the refined prediction block data, including the reconstructed block data in reconstructed frame data for the current frame, and outputting the reconstructed frame data.

The invention provides a shape function KL expansion random field discretization method considering a slope soil body irregular domain, and belongs to the technical field of soil body parameter random field analysis. Comprising the following steps: establishing a slope numerical model and dividing a discrete grid and a form function grid; establishing a probability distribution model of a soil parameter random field; calculating an element stiffness submatrix ke and an element mass matrix Ee; calculating an autocorrelation matrix rho and a global stiffness matrix B; assembling a global matrix; calculating a characteristic value lambda k, a characteristic function phi k and an expansion term number M; and random field discretization of the slope numerical model is realized. The random field discrete technology provided by the invention is used for efficiently and accurately analyzing the random distribution characteristics of the soil parameters in the rock-soil structure. The discrete random field model can be efficiently constructed in a complex geological structure, the calculation cost is remarkably reduced, the method can be conveniently combined with numerical simulation tools such as a finite element method, and the requirement of actual engineering for efficiency is met while the precision is improved.

The invention discloses a multi-view heterogeneous cascade non-stationary time sequence prediction method based on Mama improvement, and belongs to the technical field of time sequence analysis. The prediction method comprises the following steps: collecting and preprocessing time sequence data of a target domain; the time sequence data are stabilized and decomposed; the decomposed seasonal part is embedded from a univariate view angle and a multivariate view angle respectively; an embedding result is correspondingly input into a Mama encoder and a multi-granularity cascade Mama encoder decoder heterogeneous module for feature learning; performing stationarity correction on the features based on an autocorrelation matrix; and predicting the feature representation after stability correction and the decomposed trend part, adding prediction results, and carrying out inverse normalization to obtain a final prediction result. The method provided by the invention solves the technical problem that the trend and periodicity of dynamic evolution in data are difficult to capture when an existing method faces a non-stationary time sequence, and also solves the problems that an existing model is high in complexity and low in prediction accuracy.

A rolling bearing fault feature extraction method based on variable scale multipoint kurtosis deconvolution comprises the following steps: a, sampling a rolling bearing vibration acceleration signal to obtain a rolling bearing fault vibration signal; b, constructing a toeplitz autocorrelation matrix; c, constructing a variable-scale multipoint kurtosis deconvolution filter; d, filtering the rolling bearing fault vibration signal by using an optimal deconvolution filter to obtain a fault impact signal; e, performing envelope demodulation processing on the fault impact signal, extracting an envelope of the fault impact signal, and obtaining an envelope spectrum through spectral analysis; and f, judging the fault type of the rolling bearing according to the envelope spectrum. According to the method, the optimal target vector of the deconvolution is searched by constructing the variable-scale multipoint kurtosis index, the optimal filter is constructed, the fault impact signal is extracted through the deconvolution, the fault impact signal is subjected to envelope analysis, the fault feature frequency of the rolling bearing is extracted, and the fault feature of the rolling bearing can be accurately extracted.

The invention discloses a fault diagnosis method and system based on strong-adaptability image information enhancement, and the method comprises the following steps: segmenting a one-dimensional signal into data segments with equal length, obtaining a feature matrix of each data segment, a row autocorrelation matrix and a column autocorrelation matrix, coding the feature matrixes into images, generating a plurality of images through a plurality of data segments, and storing the images in a database; forming an image information data set after image enhancement; and dividing the image information data set into a training set and a test set, inputting the training set into the double-pooling multi-scale convolutional neural network model for training, then inputting the test set into the trained double-pooling multi-scale convolutional neural network, and outputting a fault diagnosis result. By adopting the technical scheme, effective diagnosis of the mechanical equipment is realized based on image information enhancement and feature fusion.

The invention discloses a mining area load spectrum anomaly detection method based on standard variable analysis, and belongs to the technical field of hydraulic pump anomaly detection.The mining area load spectrum anomaly detection method comprises the steps that S1, standardized data are obtained through data preprocessing, and a typical working condition data set of a hydraulic pump is constructed in combination with working conditions of an excavator; s2, constructing a historical vector and a future vector, and constructing a historical observation matrix and a future observation matrix according to the historical vector and the future vector; s3, constructing a Hankel matrix according to the autocorrelation matrix and the cross-correlation matrix, decomposing the Hankel matrix and determining a model order; s4, mapping the original data to a standard variable space and a residual space, and respectively evaluating the total variable quantity of the standard variable in the state space and the sum of squares of change errors in the residual space; and S5, determining an evaluation threshold value, and if the evaluation threshold value exceeds the control line, judging that the hydraulic pump operates abnormally. The method is based on standard variable analysis, adopts the pressure pulsation data of the hydraulic pump to perform anomaly detection, is sensitive to the internal running state of the pump, is not easily influenced by the external environment, and can perform early warning on faults of the hydraulic pump.

A data processing method, apparatus and device, and a computer readable storage medium relate to the technical field of target detection, and comprise the steps of obtaining millimeter-wave radar point voxel data corresponding to millimeter-wave radar data; encoding the millimeter wave radar point voxel data to generate target voxel sequence features; performing multi-layer linear calculation on the target voxel sequence features to generate query sequence features, key sequence features and value sequence features; calculating the query sequence features and the key sequence features based on an attention mechanism to obtain an autocorrelation matrix; and coding the autocorrelation matrix and the value sequence features through a multi-head self-attention mechanism, and performing residual connection on a coding result and the target voxel sequence features to generate millimeter wave radar autocorrelation enhancement sequence features. According to the method and the device, the sparsity problem of the Radar data can be effectively improved, so that the utilization efficiency of the Radar data is improved.

A sea clutter suppression method based on self-correlation features belongs to the field of image processing of measurement signals, and comprises the following steps: firstly, obtaining a self-correlation function of an RT distance-time image of a sea surface and constructing a self-correlation matrix, then constructing a covariance matrix of speckle components according to the self-correlation matrix, and performing eigenvalue decomposition on the covariance matrix of the speckle components to obtain an eigenvalue decomposition matrix; obtaining a feature vector matrix of the speckle components; then, according to the number of feature vectors of a covariance matrix of the sea surface RT distance-time image, obtaining a subspace of sea clutters; and finally, substituting the subspace of the sea clutter into an orthogonal projection suppression algorithm to obtain an orthogonal projection operator, and effectively suppressing the sea clutter in the sea surface RT distance-time image by using the orthogonal projection operator. The method can directly utilize the received data to construct the feature subspace of the sea clutter, does not need to depend on reference data of adjacent units, and avoids inaccuracy in covariance matrix estimation.

The embodiment of the invention provides a hyperspectral remote sensing image target detection method and device, a readable storage medium and electronic equipment. The method comprises the steps of obtaining three-dimensional hyperspectral remote sensing image data to be processed; processing the three-dimensional hyperspectral remote sensing image data by using a pre-constructed hyperspectral image filter to obtain target image data; the process of constructing the filter comprises the following steps: acquiring preprocessed three-dimensional hyperspectral remote sensing image data; determining a target spectral data set based on the three-dimensional hyperspectral remote sensing image data, and determining a sample autocorrelation matrix in a constraint energy minimization algorithm based on the target spectral data set; a hyperspectral image filter is determined based on a sample autocorrelation matrix and a target spectral data set, the sample autocorrelation matrix and the target spectral data set are combined to determine the hyperspectral image filter, and the hyperspectral image filter reduces the influence on the matrix when a target ground object participates in operation as a background ground object. And the influence of the target spectrum and the background spectrum on the autocorrelation matrix is effectively overcome.

The present invention discloses a GNSS deception interference detection method and device based on signal direction vector autocorrelation, which relates to the field of satellite navigation technology. The method includes: storing and updating the direction vectors corresponding to each satellite signal and arranging them into a matrix; multiplying the matrix by the conjugate transpose of the matrix to obtain an autocorrelation matrix; calculating the modulus and phase values ​​of all elements of the autocorrelation matrix, and calculating the 1-norm or infinity-norm of the autocorrelation matrix based on the modulus; and judging whether the current satellite signal contains a deception signal based on the 1-norm, infinity-norm or phase value of the autocorrelation matrix. The method provided by the present invention not only enhances and clarifies the role of collaborative detection between multiple satellite signals and improves the accuracy of detection, but also the detection results can be quantified, and the threshold value is easy to set.

The present disclosure provides a data processing method, apparatus and device, a vehicle, a medium, and a program product. The method comprises: acquiring millimeter-wave radar point voxel data corresponding to millimeter-wave radar data; encoding the millimeter-wave radar point voxel data to generate a target voxel sequence feature; performing multi-layer linear computation on the target voxel sequence feature to generate a query sequence feature, a key sequence feature, and a value sequence feature; computing the query sequence feature and the key sequence feature on the basis of an attention mechanism to obtain an autocorrelation matrix; and encoding the autocorrelation matrix and the value sequence feature, and performing residual connection on an encoding result and the target voxel sequence feature to generate a millimeter-wave radar autocorrelation-enhanced sequence feature.

The invention relates to the technical field of radar signal processing, in particular to a deep learning-based target detection method in a sea clutter environment, which comprises the following steps of: acquiring echo signals in the sea clutter environment, arranging the echo signals according to a high-speed time domain and a low-speed time domain, and sampling to obtain a complex data matrix; analyzing the complex data matrix to construct an autocorrelation matrix, and obtaining input data through the autocorrelation matrix; the method comprises the following steps: constructing a self-supervised learning model based on a Transform encoder by adopting input data, and carrying out training; and obtaining data to be detected, inputting the data to be detected into the trained self-supervised learning model based on the Transform encoder to output average similarity, and obtaining a detection result of the sea clutter environment through the average similarity. Conversion analysis is carried out based on the thunder wave signal, the expression ability is improved through the model, the target and the sea clutter are distinguished, the detection ability of the weak target under the strong clutter background is effectively improved, and the problem that the target is submerged by the sea clutter and is difficult to detect under the condition of a low signal-to-noise ratio is solved.

This invention provides a sound source localization method, apparatus, and product based on incremental learning and imbalance correction. The method includes: acquiring a dataset for a new task, a parameter-frozen feature extractor, and regularization parameters; enhancing the tail category of the new task dataset; initializing an autocorrelation matrix, a cross-correlation matrix, and a category count based on the number of categories in the new task dataset; iterating through each sample in the new task dataset, updating the autocorrelation matrix, cross-correlation matrix, and category count; calculating the category weight for each category and calculating the Gini coefficient describing the category distribution; adaptively adjusting the regularization parameters based on the Gini coefficient; and solving for and outputting the optimal weight matrix based on the adjusted regularization parameters and aggregating the global autocorrelation matrix and global cross-correlation matrix of different categories. The optimal weight matrix is ​​used to update the sound source localization model. This invention addresses the problem of intra-task and inter-task imbalance in sound source localization, improving the sound source localization effect.

The invention relates to the technical field of satellite communication, in particular to a low-orbit satellite beam self-adaptive acquisition method based on fused double architectures, which comprises the following steps: realizing local clock calibration by locking a high-orbit satellite broadcast channel and acquiring a target low-orbit satellite almanac; a terminal pose and an array surface temperature gradient are acquired in real time by using an inertial measurement unit and a temperature sensor network, and a multi-dimensional physical characteristic tensor is constructed. Then, inputting the feature tensor into a nonlinear compensation model integrating multi-scale cavity convolution and a spatial self-attention mechanism, and reasoning and outputting a phase pre-distortion matrix; and by executing channel-level Hadamard product operation of the pre-distortion matrix and the theoretical pointing matrix, generating a wave control driving matrix and synthesizing a physical receiving wave beam. After a downlink signal is captured, a real Doppler frequency shift residual error is extracted, an error autocorrelation matrix is constructed, and an almanac compensation weight is reversely corrected. According to the invention, through multi-dimensional feature perception and algorithm logic closed loop, robust beam capture is realized.

The invention discloses an image semantic segmentation method and system based on cross-modal hierarchical knowledge distillation, and the method comprises the steps: inputting an image to a multi-modal teacher model and a multi-modal student model, obtaining teacher features and student features of each stage, and outputting a prediction map; the teacher features of all the stages are fused, common modal features are enhanced, fusion features corresponding to all the stages are obtained, and the fusion features and the student features are divided into high-level features and low-level features according to different stages; the low-level teacher fusion features and the low-level student features are converted into frequency domain graphs respectively, and high-frequency parts are taken out to transmit detail knowledge; respectively calculating a cross-correlation matrix and an autocorrelation matrix by using the high-level teacher fusion features and the high-level student features, and transmitting structural knowledge; and aligning the teacher output prediction map and the student output prediction map, and transmitting response knowledge. According to the invention, the multi-modal features are fused in a mode of modal enhancement, the modal gap between the teacher fusion features and the student features is reduced, and the single-modal student model segmentation precision is improved.

The present application relates to a channel estimation method, device, electronic device and storage medium, which are applied to the field of communication technology, wherein the method includes: obtaining the frequency domain channel estimation result of the pilot position channel; determining the noise power corresponding to the frequency domain channel estimation result; obtaining the frequency domain autocorrelation matrix and the time domain autocorrelation matrix corresponding to the pilot position channel; performing dimensionality reduction processing on the frequency domain autocorrelation matrix to obtain the reduced dimension matrix; determining the filter coefficient of the minimum mean square error criterion corresponding to the time domain autocorrelation matrix, the reduced dimension matrix and the noise power; and calculating the time and frequency domain channel estimation result of the pilot position channel according to the filter coefficient of the minimum mean square error criterion and the frequency domain channel estimation result. In order to solve the problem in the prior art that the channel estimation based on the minimum mean square error criterion is performed by using methods such as SVD (singular value decomposition), and after the filter coefficient is calculated, the channel at the RS position is filtered, the process has a large amount of calculation due to the high dimension.

The invention discloses a time sequence prediction method and device, a program product, a storage medium and application thereof, relates to the field of time sequence analysis, and aims to solve the problems of high calculation complexity and poor reliability of a prediction result in a time sequence prediction task. The method comprises the following steps: training an inference model by using a time sequence, wherein the inference model extracts a dependency relationship among variables based on an autocorrelation matrix of an input original time sequence; integrating the dependency relationship into an original time sequence to obtain a key time sequence feature; learning a long-term dependency relationship between variables from the key time sequence features; and mapping a time sequence of future time steps according to the long-term dependency relationship. According to the method, the accurate prediction of the time sequence is realized with low calculation complexity, and the prediction credibility is high.

The present application relates to the field of radio signal identification, and particularly relates to a cross-domain few-shot specific emitter identification method based on self-supervised learning and transfer learning, comprising: adopting a self-supervised joint embedding method, pre-training a first model and a second model by using an upstream task pre-training dataset, regularizing a first autocorrelation matrix output by the first model by using von Neumann entropy, calculating the cosine similarity between the feature projections output by the first model and the second model, and updating the parameters of the first model based on the regularized autocorrelation matrix and the cosine similarity; migrating the pre-trained first encoder to a feature extractor, inputting a downstream task fine-tuning dataset into the connected feature extractor and classification head to fine-tune the feature extractor and the classification head; inputting a wireless signal into the fine-tuned feature extractor and classification head to obtain the predicted transmitter category of the signal. The present application achieves good identification accuracy under the conditions of few samples and low cost.

The invention discloses a sparse array DOA estimation method based on a residual network, and the method comprises the following steps: reconstructing an ideal Here-Toeplitz covariance matrix of a virtual uniform linear array through deep learning, and achieving the end-to-end training in combination with the differentiability of a Root-MUSIC algorithm. By introducing the multi-delay autocorrelation matrix tensor and residual learning, the method can effectively process challenging scenes such as coherent signals, a small number of snapshots and broadband signals. The effectiveness of the method is verified in the embodiment, high-precision estimation can be achieved under various conditions, the target that the number of estimated information sources is larger than the number of physical array elements is successfully achieved, and a new solution is provided for sparse array DOA estimation.

A MIMO, equalizer circuit (300) comprises a controller (339); a calculation circuit (302) that generates and outputs an autocorrelation matrix W (308) in response to a received covariance matrix RI (123) and channel estimate matrix H (121). A matrix manipulation circuit (316) receives and combines a received inverted autocorrelation matrix (315) and channel estimate matrix H (121) and outputs an equalization gain, P, matrix (324); receives and combines the inverted autocorrelation matrix (315) and a received signal vector y (301) and the channel estimate matrix H (121) and outputs a 'f' vector (323). A successive interference cancellation, SIC, detection circuit (325) converts the received P matrix (324) into a first orthogonal unitary matrix Q (328) and a first upper triangular matrix R (327), and converts the matrix Q (328) and the matrix R (327) and f vector (323) into a SIC estimated, equalized symbol, Z2, vector (330) that is output.

The present disclosure relates to a method, device, storage medium and electronic device for estimating a pilot position channel, the method comprising: obtaining a frequency domain channel estimation result of the pilot position channel; determining the noise power corresponding to the frequency domain channel estimation result; obtaining a time-frequency domain autocorrelation matrix corresponding to the pilot position channel, and calculating a filter coefficient according to the time-frequency domain autocorrelation matrix and the noise power; and calculating the time-frequency domain channel estimation result of the pilot position channel according to the filter coefficient and the frequency domain channel estimation result. The time-frequency domain channel estimation result of the pilot position channel is calculated based on the frequency domain channel estimation result and the noise power of the pilot position channel, so that the noise power only needs to be estimated once. Compared with performing two noise estimations in the related art, the number of calculations is reduced, the calculation error is further reduced, and the performance can be improved.

The invention discloses a short-term power load prediction method and system based on improved DSSFA-SAC-ConvLSTM, and the method comprises the steps: firstly collecting and preprocessing regional historical power load data, then employing DSSFA to analyze and extract features, constructing a feature matrix, generating a distance adjacency matrix and a dynamic adjacency matrix based on a geographic distance and DSSFA features, fusing the distance adjacency matrix and the dynamic adjacency matrix into an autocorrelation matrix, and carrying out the prediction of a short-term power load based on the autocorrelation matrix. Then splicing the feature matrix and the autocorrelation matrix, inputting the spliced feature matrix and the autocorrelation matrix into a ConvLSTM model for training, adjusting a learning rate by adopting Warmup and improved OneCycleLR, and finally mapping an output feature into a power load prediction value; according to the method, key features are extracted through DSSFA, the data dimension is effectively reduced, important information is reserved, a dynamic adjacency matrix is generated in combination with a geographic distance and a self-attention mechanism, spatial information is fully considered, the accuracy of multi-region prediction is improved, the ConvLSTM learning rate is optimized by adopting a Warmup and an improved OneCycleLR strategy, the model training effect is enhanced, and the prediction efficiency is improved. And the accuracy of power load prediction is further improved, so that the power load prediction method has more excellent performance in a complex scene.

Systems and methods for quantum linear prediction include autocorrelations formed with QFTs, and a modified quantum HHL circuit that includes appropriate normalization and encoding steps for solving a linear system of equations, including normalization of the quantum autocorrelation sequence using a norm factor; measuring a probabilistic distribution associated with values of a quantum state solution vector representing a set of quantum autoregressive parameters that correlate with a linear relationship between the quantum autocorrelation matrix and the quantum autocorrelation sequence; and generating a set of quantum linear prediction coefficients by re-normalization of the quantum state solution vector using the norm factor associated with the quantum autocorrelation sequence of the preprocessed input.

The MIMO equalizer circuit (300) includes: a controller (339); a calculation circuit (302) generates and outputs an autocorrelation matrix W (308) in response to receiving the covariance matrix RI (123) and the channel estimation matrix H (121). A matrix calculation circuit (316) that receives and combines the inverse autocorrelation matrix (315) and the channel estimation matrix H (121), and outputs an equalization gain P matrix (324); an inverse autocorrelation matrix (315), a received signal vector y (301), and a channel estimation matrix H (121) are received and combined, and a vector'f '(323) is output. A successive interference cancellation (SIC) detection circuit (325) converts the received P matrix (324) into a first orthogonal unitary matrix Q (328) and a first upper triangular matrix R (327), and converts the matrix Q (328), the matrix R (327), and the f vector (323) into an SIC estimated equalization symbol Z2 vector (330), which is output.

This invention discloses a time series forecasting method, apparatus, program product, storage medium, and its applications, relating to the field of time series analysis, to address the problems of high computational complexity and poor reliability of prediction results in time series forecasting tasks. This invention trains an inference model using time series data. This inference model extracts dependencies between variables based on the autocorrelation matrix of the input original time series; integrates these dependencies into the original time series to obtain key time-series features; learns long-term dependencies between variables from these key time-series features; and maps the time series for future time steps from these long-term dependencies. This invention achieves accurate time series forecasting with high reliability at a computationally low level.