80 results about "Mean square" patented technology

The invention discloses a pilot optimization method for large-scale MIMO channel estimation based on structural compressed sensing. The method comprises the steps of establishing a channel estimation model for a large-scale MIMO-OFDM (Multiple-Input-Multiple-Output-Orthogonal Frequency Division Multiplexing) system when pilots are placed in an overlapping mode; simplifying the channel estimation model for the large-scale MIMO-OFDM system, thereby enabling the channel estimation model to correspond to a structural compressed sensing model; and obtaining an optimum pilot matrix through utilization of a pilot optimization algorithm. Through adoption of the optimum pilot matrix, according to the channel estimation of the large-scale MIMO system based on structural compressed sensing, the mean square errors MSEs of the channel estimation are clearly reduced, and the channel estimation performance is improved.

The invention discloses a multivariable time sequence prediction method and system based on an implicit neural network. The method comprises the following steps: 1) collecting data and preprocessing the data; 2) performing window division on the standardized or normalized multivariable time sequence and determining the length of a to-be-predicted window; 3) performing variable correlation coding on the input window to obtain a variable-level feature vector; 4) the implicit neural network based on time attention predicts target parameters by using the variable features in the step 3), and implicit neural representation of the target sequence is modeled through the parameters; 5) taking the output of the implicit nerve representation and the original input window as the input of the multi-head attention predictor, and obtaining a prediction result through cross-sequence cross attention calculation performed in the implicit space and multi-layer perceptron conversion output dimension; and 6) training and optimizing model parameters, calculating a mean square error of a prediction result and a real result, taking the mean square error as a loss function, carrying out back propagation to optimize trainable parameters of the variable correlation coding module, the implicit neural network module multi-head attention predictor and the multi-layer perceptron, and then repeating the steps 3) to 6) to obtain the multi-head attention predictor. Until the preset number of iterations is reached or the error of the model on the verification set meets the requirement of early stop; and 7) performing prediction by using a model of training convergence, and performing reverse normalization on a prediction result to obtain a final prediction result. The method has good generalization, and meanwhile, the interpretability of the attention mechanism is remarkably improved by generating the hidden space characteristics of the trend component and the season component.

The invention relates to a KV cache compression method based on attention alignment, and belongs to the technical field of large language models. Comprising the following steps: a target model self-generates dialogue data reply, and constructs training data; adding a soft token into a word list of the target model, and carrying out soft token random initialization; finely adjusting specified parameters, splicing soft tokens at the tail of an original input sequence, and transmitting the soft tokens to the model; and respectively calculating the soft token and the attention distribution of the self-generated reply, calculating the mean square error of the soft token and the attention distribution of the self-generated reply as a loss function, and completing training. According to the method, a limited number of soft tokens playing an auxiliary role are introduced into original input, and importance discrimination of KV cache elements is realized and expelling is completed by aligning attention distribution of the soft tokens and real generated content; according to the method, better compromise is achieved between sequence length compression and performance loss, and the loss of model performance is better controlled while it is guaranteed that KV cache video memory space occupation is reduced.

The invention belongs to the field of water-wind-light multi-energy complementary optimal scheduling, and particularly discloses a water-wind-light optimal scheduling method and system based on a two-stage dual-population evolutionary algorithm, and the method comprises the steps: taking the water level of a reservoir as a decision variable, taking the maximum power generation amount and the minimum residual load mean square deviation as objective functions, and constructing a multi-objective scheduling model; the two populations are initialized, and the water-wind-light multi-target scheduling model is solved; the two populations are initially in a fixed division stage, and operators A and B are respectively adopted for iterative updating; when a switching condition is met, switching to a self-adaptive cooperation stage for iterative updating, determining the selection probability of the operators A and B according to population performance at the moment, and selecting the operators based on the selection probability for iterative updating; the switching condition is that after the population is iteratively updated each time, if the population diversity is smaller than a threshold value or the evaluation frequency reaches the threshold value, stage switching is carried out. According to the method, the contradiction between convergence and diversity in water-wind-light multi-objective optimization can be solved, and accurate optimization is realized.

The invention discloses a motor multi-objective optimization design method and system based on a Kriging agent model. Comprising the following steps: constructing an initial data set through an experimental design method and finite element simulation calculation, so as to train an initial Kriging agent model which takes a motor design variable combination as an input variable and takes motor performance as an output response; the proxy model serves as a target function, a multi-target optimization problem is solved, the proxy model is updated, and an updating strategy is as follows: after each round of optimization is finished, a high-error solution in a current Pareto solution set is screened based on a Kriging model prediction mean square error, a finite element response of the high-error solution is obtained and supplemented to a data set, and the proxy model is updated and trained; and repeating the process until the optimization result converges. According to the method, high-uncertainty region samples are selectively supplemented, the calculation cost is remarkably reduced while the local prediction precision of the proxy model at the Pareto leading edge is improved, and the method has good practicability and economical efficiency.

The invention relates to a multi-channel multi-mode space-time 3D convolution and Transform fused satellite-borne GNSS-R wave height inversion model, and aims to solve the problems that a traditional empirical model is poor in adaptability to complex oceans and existing deep learning cannot fully excavate space-time correlation. According to the method, three modules are innovatively fused for cooperative processing: firstly, a 3D CNN-ConvLSTM is utilized to extract local spatial-temporal characteristics of multichannel GNSS-R data and capture a dynamic time sequence; secondly, performing deep coding on the sea surface environment parameters through a multi-head attention mechanism of Transform, and establishing global dependence between features; and finally, realizing cross-modal fusion by adopting a weighted summation and feature splicing strategy. And through training of an Adam optimizer and control of an early stop strategy, optimization is carried out by taking a mean square error as a loss function. Experiments show that compared with a traditional model and a machine learning model, the method has the advantages that the significant wave height estimation error is reduced by 40-53%, the correlation coefficient reaches 0.84-0.91, the precision and generalization ability under the complex sea condition are remarkably improved, and reliable support is provided for ocean remote sensing monitoring and disaster early warning.

The invention discloses a current transformer calibration method and device based on minimum mean square M estimation and a medium, and belongs to the technical field of current transformer calibration. The method comprises the steps that response signals are extracted from a to-be-calibrated transformer and a standard transformer, and response signal vectors are established; parameters of an N-order RVSSLMSM adaptive filter are initialized, the initial parameters and a to-be-calibrated mutual inductor response signal vector are substituted into the adaptive filter to obtain filter output, an error vector of an adaptive filter output signal vector and a standard mutual inductor response signal vector is calculated, and therefore a variance value is updated; and according to the variance variation, updating the parameters of the adaptive filter, substituting the response signal vector of the to-be-calibrated mutual inductor into the updated adaptive filter, and repeating the steps until the residual sum is converged or reaches a certain number of iterations, thereby obtaining the optimal coefficient of the filter and completing the calibration. And the calibration precision and the system robustness are effectively improved.

The invention discloses an interpretable data-knowledge-free distillation method based on a class activation graph, and the method comprises the following steps: inputting random noise to a generator, and generating a composite image similar to the distribution of a target data set; respectively inputting the synthesized image into a teacher model and a student model which are subjected to structure adjustment; respectively generating corresponding class activation diagrams through the teacher model and the student model; inputting the teacher class activation graph and the student class activation graph into a pre-trained scale mapping network, and performing cross-scale scaling on the class activation graphs; performing regularization processing on the scaled class activation diagram to obtain a standardized teacher class activation diagram and a student class activation diagram; calculating mean square error loss based on the teacher type activation graph and the student type activation graph to realize knowledge transmission; and finally, optimizing the generator in combination with information entropy loss, one-time heat loss and activation loss. According to the method, effective knowledge distillation can be realized under the condition of no real data, and the interpretability and stability of the distillation process are improved.

The invention provides an uplink MIMO-NOMA (Multiple Input Multiple Output-Non-Orthogonal Multiple Access) system optimization method with a maximized sum rate, which comprises the following steps: firstly, in order to suppress interference between a transmission cluster and a reflection cluster, providing an inter-cluster SIC method: clustering users according to transmission characteristics, adopting space division multiple access in each cluster, and performing inter-cluster SIC of NOMA transmission so as to reduce the interference between the clusters; secondly, a non-convex optimization problem aiming at maximizing the total rate of the system is formulated and solved by using an alternating optimization framework, in the framework, a receiving equalizer is optimized according to a minimum mean square error standard, and transmitting power control of a user is processed through successive convex approximation, so that sub-problems are easy to process; and finally, the beam forming of the active STAR-RIS adopts a sequential rotation design. According to the invention, the dual-fading effect is effectively overcome by adjusting the phase shift and amplitude of electromagnetic waves, an alternating optimization framework is also provided to solve the described non-convex problem, user power is controlled through continuous convex approximation, and active STAR-RIS beam forming is realized through sequential rotation.

In examples, a filter used to denoise shadows for a pixel(s) may be adapted based at least on variance in temporally accumulated ray-traced samples. A range of filter values for a spatiotemporal filter may be defined based on the variance and used to exclude temporal ray-traced samples that are outside of the range. Data used to compute a first moment of a distribution used to compute variance may be used to compute a second moment of the distribution. For binary signals, such as visibility, the first moment (e.g., accumulated mean) may be equivalent to a second moment (e.g., the mean squared). In further respects, spatial filtering of a pixel(s) may be skipped based on comparing the mean of variance of the pixel(s) to one or more thresholds and based on the accumulated number of values for the pixel.

The invention provides a load prediction method and related equipment. The method comprises the following steps: carrying out preprocessing and characterization processing on parameters correspondingly obtained on a day to be predicted to obtain feature data; inputting the feature data into a trained mean square error prediction model to obtain a mean square error value corresponding to the day to be predicted; the mean square error value is a mean square error between the to-be-predicted day and each historical day; determining a candidate day based on the mean square error value; calculating the trend similarity between each candidate day and the day to be predicted, and determining at least one similar day based on the trend similarity; and determining the load value of the to-be-predicted day based on the load value of the similar day and the mean square error value. According to the embodiment of the invention, through characterization processing of the to-be-predicted day parameters, in combination with the trained mean square error prediction model and similar day analysis, the accuracy and reliability of load prediction are effectively improved, and the method is suitable for complex and changeable practical application scenes.

The application discloses a path planning method based on deep reinforcement learning-fast exploration random tree, comprising the following steps: S1, obtaining a starting point and an ending point; S2, calculating candidate path points and selecting an action in an action state according to a Q value; S3, calculating a reward value and a new action state after the action is executed; S4, storing the action state, the action, the reward value and the new action state to an experience pool, in response to the number of stored experience values in the experience pool being greater than a batch size, randomly selecting experience values of the batch size, and updating the Q value and a time difference error through a policy network; S5, updating policy network parameters through a mean square error loss, and calculating a target network update step according to the time difference error; S6, judging whether a searched path reaches the ending point or satisfies a set maximum path point search number, if yes, outputting a current path, and if not, returning to S2. The application improves the search efficiency of the algorithm without increasing the search time of the algorithm.

The application discloses a current transformer calibration method and device based on minimum mean square M estimation, and a medium, and belongs to the technical field of current transformer calibration. The method extracts response signals from a to-be-calibrated transformer and a standard transformer, and establishes a response signal vector. Parameters of an N-order RVSSLMSM adaptive filter are initialized, the initial parameters and the response signal vector of the to-be-calibrated transformer are brought into the adaptive filter to obtain filter output, an error vector of the adaptive filter output signal vector and the response signal vector of the standard transformer is calculated, and a variance value is updated. According to the size of the variance change amount, the adaptive filter parameters are updated, the response signal vector of the to-be-calibrated transformer is brought into the updated adaptive filter, and the cycle is repeated until the residual error converges or a certain number of iterations is reached, the optimal coefficient of the filter is obtained, the calibration is completed, and the calibration accuracy and the system robustness are effectively improved.

The invention discloses a method for predicting an impact displacement time history curve of a concrete filled steel tubular column under the influence of defect coupling, comprising the following steps: S1, acquiring a data set: acquiring test data and verified finite element numerical simulation data, the data set sample comprising structural characteristic parameters of a component and a corresponding displacement time history curve under transverse impact, s2, data preprocessing: extracting geometric, material, defect and impact working condition parameters of the component to form an input feature vector, standardizing input features, performing time alignment and resampling on a displacement time history, and constructing a time sequence input tensor of a unified dimension; s3, a CNN-LSTM-Attention model is established, and the CNN-LSTM- S4, training and optimizing the model by using sample data, and taking a mean square error as a loss function; s5, target component parameters are input, and an impact displacement time history prediction result is output. According to the method, the impact displacement time history of the concrete filled steel tubular column containing multiple defects can be accurately and efficiently predicted, and the impact resistance evaluation efficiency of the structure is improved.

The application relates to the technical field of high-dimensional data, in particular to a high-dimensional data fitting algorithm and system based on an improved gradient descent, the algorithm is deeply coupled with a long short-term memory (LSTM) network, a gated differential adaptive momentum variable gradient descent (AM-VGD) is designed for high-dimensional time series data, high-precision fitting of high-dimensional data is realized, the gated differential adaptive momentum variable gradient descent (AM-VGD) is proposed, different momentum factors and gradient decoupling items are designed according to the gradient characteristics of the LSTM forget gate, input gate and output gate, the gradient propagation efficiency is improved by 40% under high dimension, the model convergence iteration number is reduced to <=500 times, the convergence speed is improved by more than 2 times, a stacked auto-encoder (SAE) is fused to decouple and extract high-dimensional features, a high-dimensional decoupling regular term of the gradient descent is combined, the collinearity interference between features is eliminated, the fitting determination coefficient R2 of high-dimensional time series data is greater than or equal to 0.98, the mean square error (MSE) is reduced by more than 60%, and there is no overfitting / underfitting phenomenon.

The invention discloses a lightweight attention-based long-term time series prediction method and system for an MLP architecture. The method comprises the following steps of 1, preprocessing time series data; 2, data grouping reconstruction is carried out; 3, performing feature extraction and representation on the obtained time sequence feature vector to obtain a time sequence feature; 4, processing the time sequence feature vector to obtain a prediction time sequence generated based on the training data set; step 5, training to obtain a time sequence prediction model PCOTS; step 6, obtaining a prediction time sequence generated based on the test data set; 7, calculating MAE and MSE between the prediction sequential sequence obtained in the step 6 and the prediction sequence in the test data set; 8, repeating the steps 2 to 7 until the MAE obtained in the step 7 and the mean square error MSE are not reduced any more, and obtaining an optimal time sequence prediction model PCOTS; and 9, inputting an input sequence given by a prediction task into the optimal time sequence prediction model PCOTS obtained in the step 8 to obtain a generated prediction time sequence.

The invention provides a music time coding method for artificial neural network modeling. The method aims at the coding requirement of music beat time, molecules are decomposed according to a fixed denominator for multi-bit independent classification coding, and the method comprises the following steps: setting the fixed denominator as a composite number and decomposing the composite number into a prime factor combination; the molecules are mapped into vectors, the value range of each bit is determined by a cardinal number of a prime factor combination, and the vectors are converted into one-hot coding sub-vectors with redundant components removed. The technical scheme of the invention solves the problem that the traditional mean square error (MSE) ignores the internal correlation of scores and has high sparsity in the topology recognition task of an optical music recognition system (OMR), and compared with a traditional coding method adopting floating point numerical output and a mean square error loss function under the conditions of the same data set, model architecture and training hyper-parameters, the method provided by the invention has the advantages that the coding efficiency is greatly improved; according to the method, the comprehensive error index is reduced by about 21%, and the model precision and the training efficiency are improved.

The invention discloses a self-adaptive BFP quantization method and device for deep neural network acceleration, and the method comprises the steps: calculating the mean square error of all blocks when a current tensor carries out the BFP quantization through employing a current block size, and if the mean square error is smaller than or equal to a quantization error threshold value, carrying out the BFP quantization. If not, performing BFP quantization on the current tensor according to the current block size, ending and exiting; otherwise, the current tensor is divided into a subset meeting the quantization error threshold value and a subset not meeting the quantization error threshold value, the subset meeting the quantization error threshold value is directly subjected to BFP quantization according to the current block size, and the subset not meeting the quantization error threshold value serves as a new current tensor; and reducing the size of the block to obtain a new current block size, and continuing iteration. The method aims at better balancing the model precision and the hardware overhead, and fundamentally reducing the dependence on floating point operation, so that the hardware cost is reduced.

The application discloses a channel estimation method and related device, wherein the method comprises: performing minimum mean square error (MMSE) channel estimation based on the product of the first amplitude factor and the first noise power of the first dimension, the first channel autocorrelation matrix of the first dimension, and the first diagonal matrix to obtain the channel estimation value matrix of the first dimension; wherein the channel estimation value matrix of the first dimension is obtained by performing MMSE channel estimation in the non-last step of the multi-dimensional step-by-step MMSE channel estimation; the first amplitude factor is a positive number less than or equal to 1; and performing MMSE channel estimation based on the channel estimation value matrix of the first dimension, the second channel autocorrelation matrix of the last dimension, the second noise power and the second diagonal matrix to obtain the multi-dimensional channel estimation value matrix. The method can improve the performance of the multi-dimensional step-by-step MMSE channel estimation.

Unlike conventional methods that assess the soundness and safety of structures by directly applying external forces to the object, this method provides a structural diagnostic and evaluation method based on the ambient micro-vibrations of a structure that can diagnose and evaluate the seismic resistance, soundness, and effectiveness of countermeasures and reinforcement work in an inexpensive and rapid manner. [Solution] In a method for evaluating the performance of a structure by observing ambient microtremors, ambient microtremor observation history is observed simultaneously at multiple observation points within the structure, and an estimated value of an index used for seismic design of the structure is calculated using the mean square (RMS) of these time histories, and the seismic performance of the structure based on the observations is evaluated using the ratio of this value to the value of the index used at the time of design.

The present disclosure relates to a model training method and device and a readable storage medium, wherein the method performs knowledge distillation on each first intermediate subnetwork of a teacher model, thereby avoiding the selection and setting problems of the distillation position and the number of distillation layers, and achieving better distillation effect in the intermediate subnetwork distillation task. In addition, the knowledge distillation is performed by using the boundary mean square error loss, which not only enables the student model to learn the output of the teacher model, but also enables the student model to learn whether the teacher model activates the output, so that the student model learns more knowledge and is beneficial to improve the performance of the target student model.

The application discloses a movable array element anti-interference position beam joint method facing unknown direction, introduces movable array elements, and makes array geometry structure be treated as adjustable optimization variable in deployment stage. By changing equivalent array element spacing and pattern structure of the array, the separable degree of expected signal and interference signal incident direction on the spatial response is improved, and the destructive influence of strong interference on the array output when falling in the high side lobe / grating lobe direction of the array is reduced. In each candidate position, an adaptive beam forming process with a pilot sequence as a reference is adopted to directly generate a beam forming weight vector corresponding to the position, so that the mean square error of the receiving output in the pilot interval is minimized. The "movable array element position-performance index" is modeled as a target function to be optimized, and a global adaptive search method based on a probability model is adopted to select the next evaluation position in an iterative manner, instead of exhausting evaluation on all candidate positions.

This disclosure relates to the field of bridge engineering, specifically to a design method, system, and medium for HPC-RC combined eccentrically compressed columns. The method includes: defining a design space; selecting optimal design data combinations based on engineering specification constraints, bearing capacity constraints, and cost-effectiveness values ​​to construct a design data sample set; constructing a two-branch heterogeneous neural network model, the model including an input layer, a shared feature extraction layer, a diameter prediction classification branch, and a reinforcement area prediction regression branch; constructing a loss function composed of focal loss and mean square error loss; training the two-branch heterogeneous neural network model using the design data sample set; and using the trained two-branch heterogeneous neural network model to predict the diameter and total reinforcement area of ​​the eccentrically compressed column. This disclosure improves the stability and accuracy of the design, reduces the computational burden, increases computational speed, and reduces computational complexity.

Methods, systems and apparatus for estimating the fidelity of quantum logic gates. In one aspect, a method includes defining multiple sets of random quantum circuits; for each set of random quantum circuits: selecting an observable for each element in the set of random quantum circuits, wherein each selected observable corresponds to a respective element of the set of random quantum circuits and is dependent on the element to which it corresponds; estimating a value of a polarization parameter for the set of random quantum circuits, comprising performing a least mean squares minimization based on multiple expectation values, wherein each expectation value comprises an expectation value of a respective selected observable with respect to an output of an experimental implementation of a random quantum circuit corresponding to the respective selected observable; and processing the estimated polarization parameter values to obtain an estimate of the fidelity of the n-qubit quantum logic gate.

The invention discloses a self-distillation method for making up the teacher-student gap through residual learning, and relates to the technical field of computers. The method at least comprises the following steps: S1, obtaining an original data set, and dividing the original data set into a training set and a test set; s2, preprocessing the original data set; s3, dividing the network model into a plurality of outlets, introducing a cross-layer feature fusion module, and designing a residual learning strategy for multi-outlet self-distillation; s4, constructing a complete knowledge distillation loss function, improving a feature loss function based on a mean square error, and proposing multi-scale loss; and S5, inputting training set image data into the modified network model, storing the weight of the trained model, and then loading the model for reasoning to obtain a prediction result. Through the design of the method, the problems that students are difficult to learn and the knowledge of teachers is single when the teacher-student gap is large in the self-distillation process can be solved, and the classification precision of the model is effectively improved on the premise of not increasing extra cost during reasoning.

This invention discloses an orbital jump processing method, electronic device, and computer program product. The orbital jump processing method includes: determining a first mean square error in fitting the current orbit of the satellite to a first orbit, and a second mean square error in fitting the current orbit to a second orbit; the first orbit is the orbit preceding the current orbit of the satellite, and the second orbit is a predicted orbit obtained by batch processing based on the current orbit; determining the value of the orbital smoothing weight matrix in a preset first observation equation based on the first mean square error and the second mean square error; the orbital smoothing weight matrix characterizes the degree of dependence on the current orbit or the first orbit during orbit fitting; and determining the orbital value used for orbit updating based on the first observation equation.

This invention provides a reference-free background calibration method for ultra-high-speed TIADCs based on minimum mean square (RMS) calibration. It employs a decimation-calibration-interpolation calibration mode. The original quantized data is decimated using a specific decimation rate to obtain multiple quantization code sets. A step-by-step grouping calibration algorithm is used based on the number of interleaved channels, selecting the first quantization code set for step-by-step calibration. Each step involves grouping the data and using a dual-channel synchronous detection method based on a binary search to search for the clock jitter value of each channel in the grouped quantized data. Other quantization code sets are then compensated using a first-order Taylor approximation in the digital domain based on the clock jitter value of the first quantization code set. Finally, the compensated multiple quantization code sets are re-interpolated to obtain the calibrated result of the original data. This invention eliminates the need for a reference channel, minimizing the RMS of the detected clock jitter value and thus improving compensation accuracy.

The application relates to the technical field of communication, and specifically provides a reconfigurable intelligent surface (RIS) assisted ocean communication method and system based on LR-BCD, which comprises the following steps: constructing an RIS assisted ocean communication system model; constructing an ocean double-path channel model based on the ocean double-path propagation theory and in combination with a sea surface reflection coefficient; extracting low-rank characteristics of the channel by using singular value decomposition (SVD), projecting high-dimensional beamforming and phase shift optimization problems into a low-dimensional subspace; constructing a low-dimensional and rate maximization problem, and converting the problem into a convex subproblem by using a weighted minimum mean square error (WMMSE) criterion; alternately updating a low-dimensional precoding matrix and a phase shift vector by using a block coordinate descent (BCD) algorithm; and finally reconstructing the optimized low-dimensional variables back to a high-dimensional space and performing discretization and quantization. The application utilizes the sparsity of the ocean channel, significantly reduces the calculation complexity of large-scale RIS optimization through dimension reduction processing, and meanwhile, maintains the system and rate performance equivalent to that of a full-rank algorithm.

The invention discloses a path planning method based on deep reinforcement learning-fast exploration random tree. The method comprises the following steps: S1, acquiring a starting point and an ending point; s2, calculating candidate path points, and selecting an action in the action state according to the Q value; s3, calculating a reward value and a new action state after the action is executed; s4, the action state, the action, the reward value and the new action state are stored in an experience pool, an experience value of the batch size is randomly selected in response to the fact that the storage number in the experience pool is larger than the batch size, and the Q value and the time sequence difference error are updated through a strategy network; s5, strategy network parameters are updated through mean square error loss, and the target network updating step length is calculated according to the time sequence difference error; s6, judging whether the searched path reaches the end point or meets the set maximum path point search number, and if so, outputting the current path; and if not, returning to S2. According to the method, the algorithm searching efficiency is improved, and the algorithm searching time is not increased.