17 results about "Regularized least squares" patented technology

This invention relates to the field of navigation and timing technology, specifically to a GNSS timing data prediction method and related equipment. The method involves acquiring Roland timing data when GNSS timing data is invalid and preprocessing it. The preprocessed data is then input into a trained single-hidden-layer extreme learning machine model. This model calculates the hidden-layer output matrix by randomly initializing the input layer weight matrix and bias vector, and solves for the output weights using regularized least squares. The model outputs a normalized GNSS timing data prediction value, which is then denormalized to obtain the final GNSS timing prediction result. This method combines the high stability of the Roland system with the efficient computational power of the extreme learning machine to achieve accurate GNSS timing data prediction. It is suitable for enhancing the resilience of positioning, navigation, and timing systems in critical infrastructure scenarios, ensuring the continuous operation of the system even when GNSS signals are interfered with or fail.

The application provides an underwater structure noise control method based on a parametric secondary sound source, which comprises the following steps: firstly, using the sound pressure and normal velocity information of a structure surface, a far-field noise prediction method is adopted to obtain the radiation noise at a virtual reference point in the far field, thereby providing an input basis for parametric secondary sound source control; secondly, a multi-channel parametric sound field model is established to verify the controllability of the parametric secondary sound field, and a regularization least square algorithm is adopted to obtain multi-channel signal emission parameters by taking the sum of the sound pressures at multiple virtual reference points in the far field as the minimum target; finally, when the monitoring information of the underwater structure is missing, a local control strategy is adopted to obtain the underwater structure noise control effect. The application is directed to directional control in the accurate area of far-field noise prediction, solves the problem of poor traditional active noise reduction effect caused by missing monitoring information, improves the reliability and stability of the system, and lays a theoretical foundation for the application of underwater structure active noise control technology.

The application discloses a kind of pure angle target trajectory estimation methods, this method will typical target motion model be converted into the autoregressive moving average model of position component to realize state decoupling, by adjusting the relationship between pseudo-linear least square cost function and process noise energy to obtain regularized pseudo-linear least square model.To reduce the estimation deviation generated by the correlation between pseudo-linear measurement matrix and noise, the regularized instrumental variable least square position estimation is obtained by using instrumental variable estimation method.The method of the application obtains the analytical expression of batch estimation of azimuth position component in the framework of regularized least square estimation.The method has linear time calculation complexity, and its estimation performance is irrelevant to initial state.

The invention discloses a virtual inertia identification method suitable for a doubly-fed wind power plant. The method comprises the following steps: constructing a second-order transfer function and a virtual rotational inertia expression of grid-connected power and power grid frequency of the doubly-fed wind power plant; collecting data, selecting a stable active power sequence and a power grid angular frequency sequence as identification data and denoising, and then selecting quasi-steady-state data to calculate a droop control coefficient; and discretizing the transfer function, identifying a coefficient of the discretized transfer function based on a regularized least square method, restoring a second-order transfer function based on the identified parameter, and calculating the virtual rotational inertia by using the coefficient of the second-order transfer function and a virtual rotational inertia expression. According to the method, no phase delay exists in the process of identifying the virtual inertia of the doubly-fed wind power plant, and the peak characteristic of inertia response is reserved; the inertia identification error caused by parameter coupling in the traditional method is solved; and by introducing a regularization factor, a severe oscillation phenomenon of fitting parameters is prevented, and the identification precision of the virtual inertia is remarkably improved.

The invention provides an unsupervised joint optimization method for realizing hyperspectral imaging and super-resolution reconstruction under the condition of extremely low sampling rate. The method comprises the following steps: firstly, based on a single pixel imaging (SPI) physical model and high-resolution RGB image gray level prior, adopting a pre-reconstruction algorithm (LS-RGP) fused by regularization least square and gray level prior to quickly recover a preliminary low-resolution hyperspectral image from a measurement vector; secondly, an Untrained RGB-Guidded Hypersection Recovery Network (UHRNet) is introduced, and a preliminary result is gradually refined and recovered through multiple physical and semantic constraints such as measurement consistency, Fourier domain sparse regularization, gradient sharpness loss, spatial smoothing total variation and VGG-based perception loss under the condition of not depending on large-scale annotation data; and finally, designing a super-resolution network (USRNet) based on a multi-head self-attention Transform, and by combining a cross-modal attention mechanism and high-resolution RGB guidance, realizing unsupervised reconstruction of a 512 * 512 high-resolution hyperspectral chart by simulating a super-sampling measurement operator and jointly optimizing measurement consistency, frequency domain regularization, three-dimensional total variation and structural similarity loss. The method does not need to depend on massive pre-training samples, at lt; the method can efficiently recover an image with high spatial resolution and high spectral precision under the conditions of high spectral resolution, 1%-5% sampling rate and low signal-to-noise ratio, and is especially suitable for hyperspectral imaging and monitoring application in a resource-constrained environment.

The invention discloses a bamboo strip character denoising method based on a fractional order diffusion equation inverse problem POD algorithm, and belongs to the field of image processing and inverse problem calculation, and the method comprises the steps: carrying out the preprocessing of a bamboo strip image, and obtaining a terminal observation image; based on the set diffusion time and the fractional order, a finite difference method is adopted to solve a positive problem of a time fractional order diffusion equation, and a snapshot data set is generated; based on the snapshot data set, a singular value decomposition method is adopted to extract a main mode, and a reduced-order model is established; based on the terminal observation image and the reduced-order model, constructing a regularization least square optimization problem and adopting a gradient iteration algorithm for solving, and performing inversion to obtain a clear image at an initial moment; and processing the clear image at the initial moment by adopting a threshold segmentation method, and outputting a finally recovered high-definition bamboo strip image. The invention provides an efficient and stable treatment means with clear physical significance for digital protection of the fragile cultural relics such as the bamboo strips.

The invention discloses a transient impact enhancement extraction method and device under strong background noise and a medium, and belongs to the technical field of fault diagnosis. The method comprises the following steps: collecting noise-containing vibration signals of the rotating machinery; constructing a discrete cosine transform over-complete atom dictionary; initializing a residual error, a sparse coefficient and a support set; determining a threshold to evaluate the support set; selecting an optimal atomic index on the basis of target function maximization; updating the support set and calculating a sparse coefficient through regularization least square; and iterating the reconstructed signal and updating the residual until the maximum number of iterations is met, and outputting an enhanced transient impact signal. According to the method, the effect of adaptively extracting the complete transient impact characteristics under strong background noise is achieved.

The application discloses a kind of lightweight remote sensing image geometric positioning method, system and equipment, the method includes the following steps: S1. Obtain remote sensing image and create coordinate dataset;S2. The parameters of remote sensing image geometric positioning model are initialized, and input weight matrix is generated using orthogonal random projection;S3. Remote sensing image geometric positioning model is trained according to coordinate dataset, hidden layer output is generated by nonlinear feature mapping, and output weight matrix is constructed according to regularized least squares method combined with multi-level solution strategy;S4. According to cross-validation division, the performance of the model is evaluated, and the model is retrained after optimizing the regularization parameter;S5. Based on the model parameters of training completion, test data is input into the trained remote sensing image geometric positioning model to perform regression calculation, and the calculated coordinate data are compared with the coordinates calculated by satellite sensor model, and the running time is recorded, to verify the efficiency and accuracy of the model. The application is applied to the technical field of photogrammetry.

The invention relates to the technical field of pulse power and electromagnetic manufacturing, and discloses a pulse magnetizer which comprises a charging module, a capacitance matrix module, an acquisition module and a control module. The acquisition module adopts a Kelvin four-wire system structure to synchronously acquire a voltage and current discrete sequence at a discharge moment; the control module is based on a heterogeneous architecture of a digital signal processor and a field programmable gate array, equivalent resistance and inductance parameters of a load are identified in real time by using a regularization least square algorithm, and a closed-loop control logic based on single waveform reverse analysis is adopted by a system core: a nonlinear equation is solved by using a Newton iteration method to reconstruct a capacitance matrix; the pulse width is accurately locked; and meanwhile, an iterative learning algorithm is combined to correct the charging voltage, compensate the capacitance discrete quantization deviation and lock the peak current. According to the invention, the self-adaptive compensation of the magnetizing waveform to the load parameter drift is realized, the current repetition precision is obviously improved, and the service life of the system is obviously prolonged.

The invention provides an arbitrary array MIMO radar parameter estimation method and device, and belongs to the field of radar signal processing. The method comprises the following steps: in an offline stage, constructing an array manifold matrix of a physical arbitrary array and a virtual uniform linear array at a plurality of calibration angles, solving a regularization least square problem, and obtaining a transmitting end interpolation matrix and a receiving end interpolation matrix; in the on-line stage, digital pre-weighting is carried out on a baseband transmitting signal by using a transmitting end interpolation matrix, a radar echo signal of a target is received, a three-dimensional mixed data tensor is constructed, and then demodulation is carried out; and decomposing the demodulated tensor by adopting a parallel factor analysis algorithm to obtain a virtual transmitting factor matrix and a physical receiving factor matrix, and estimating the direction of arrival and the direction of arrival of the target in combination with a receiving end interpolation matrix. Through space virtualization, an advanced tensor algorithm is expanded to any actual array platform from an ideal uniform array, and the problem of algorithm applicability under hardware constraint is solved.

The application discloses a kind of ocean gravity field modeling methods under seabed topography constraint fusing frequency domain and space domain, it is related to ocean gravity field technical field, including the following steps: S1 pretreatment multi-source data, with EGM2008 as reference, 3σ criterion eliminates ship survey anomaly point, GMT's x2sys module adjustment weakens systematic error;S2 based on Parker fast Fourier transform forward topographic gravity anomaly, determine cut-off wavelength after power spectrum analysis, construct initial gravity background field;S3 recursive quadtree adaptive partitioning;S4 constructs topographic constraint polyhedral function model fitting residual;S5 regularized least squares solution parameter, compensate and refine residual to synthesize high-precision model.The application introduces seabed topography constraint in dual domain, combines adaptive partitioning, realizes high-low frequency information complementation, breaks through calculation bottleneck, fits physical law, absorbs systematic error, suppresses boundary effect, greatly improves the precision and robustness of ocean gravity field model.

The invention provides a charging facility metering error dynamic estimation method and device based on least square and ADMM optimization, and the method comprises the steps: obtaining and preprocessing charging pile group data, and constructing a multi-source spatio-temporal data set; performing noise filtering by an attention-based time convolutional network auto-encoder; analyzing the electrical association between the charging piles through a dynamic graph comparison learning algorithm, and generating an electrical association matrix reflecting the topological structure between the charging devices; constructing a physical information embedded time-varying regularization least square optimization model by using the electrical incidence matrix, the equipment physical parameters and the total meter and branch meter electric quantity difference value of the charging pile, and determining the coupling optimization problem of the electric meter error and the line loss; and carrying out inversion solution on the ammeter error and the line loss by adopting a time-varying transaction direction multiplier optimization algorithm, and generating a real-time ammeter error coefficient and a line loss estimated value. According to the method, accurate separation of the metering error and the line loss is realized, and the accuracy and the reliability of error detection of the metering system are remarkably improved.

This invention relates to the field of tactile sensing technology, specifically to a six-dimensional force tactile sensor based on FBG, a demodulation method, and a fabrication method. The demodulation method includes: acquiring the first center wavelength drift of the sensing fiber optic component and the temperature center wavelength drift of the reference fiber optic component; calculating the temperature change from the temperature center wavelength drift; compensating the first center wavelength shift with the temperature change to obtain the wavelength drift vector of the sensing group; calibrating the sensing group to obtain a sensitivity matrix; combining the wavelength drift vector with the sensitivity matrix and establishing a linear model; and solving the linear model using regularized least squares to obtain an estimated value. The embodiment achieves synchronous acquisition of multi-source optical signals by acquiring the center wavelength drifts of the sensing group and the reference fiber optic component; calculating the temperature change based on the temperature center wavelength drift and compensating for the wavelength drift of the sensing group accordingly effectively eliminates temperature interference and obtains the wavelength drift vector of the pure mechanical response.

The invention relates to the technical field of navigation and timing, in particular to a GNSS timing data prediction method and related equipment, and the method comprises the steps: obtaining and preprocessing Roland timing data when the GNSS timing data is invalid; and inputting the preprocessed data into a trained single hidden layer extreme learning machine model, calculating a hidden layer output matrix by the model through randomly initializing an input layer weight matrix and an offset vector, and solving an output weight by adopting a regularization least square method. And the model outputs a normalized GNSS timing data prediction value, and a final GNSS timing prediction result is obtained after reverse normalization processing. According to the method, high stability of the Rowland system and efficient computing power of the extreme learning machine are combined, accurate prediction of GNSS timing data is achieved, the method is suitable for positioning, navigation and time service toughness enhancement scenes of key infrastructures, and the continuous operation capacity of the system when GNSS signals are interfered or fail is guaranteed.

The present application relates to the technical field of pulse power and electromagnetic manufacturing, and discloses a pulse magnetizing machine, which comprises a charging module, a capacitor matrix module, an acquisition module and a control module; the acquisition module synchronously acquires voltage and current discrete sequences at the discharge moment in a Kelvin four-wire structure; the control module is based on a heterogeneous architecture of a digital signal processor and a field programmable gate array, uses a regularized least square algorithm to identify equivalent resistance and inductance parameters of a load in real time, and adopts a closed-loop control logic based on single waveform reverse analysis as a system core; a Newton iteration method is used to solve nonlinear equations to reconstruct the capacitor matrix and accurately lock the pulse width; at the same time, the charging voltage is corrected by combining an iterative learning algorithm, the capacitor discrete quantization deviation is compensated, and the peak current is locked. The present application realizes adaptive compensation of the magnetizing waveform to load parameter drift, and significantly improves the current repetition accuracy and the operation life of the system.

The invention discloses an industrial robot pose error prediction method, system, medium and equipment based on geometric and non-geometric fusion driving, and the method comprises the steps: collecting nominal joint angles and actual tail end poses corresponding to a plurality of expected pose points of an industrial robot in a task space, calculating a pose error formed by the nominal joint motion angle and the actual pose, and calculating the pose error of the industrial robot; constructing a pose error data set; constructing a geometric kinematics model of the robot based on the geometric parameters of the industrial robot, and obtaining estimated values of the geometric parameters of the robot by using a regularization least square method; constructing a fusion-driven neural network model; performing staged training on the fusion-driven neural network model through minimizing a loss function; and inputting a nominal joint angle corresponding to a to-be-predicted pose into the trained fusion-driven neural network model, and outputting a predicted pose error.

This invention discloses a chaotic time series prediction method based on multi-scale decomposition. First, the original one-dimensional chaotic time series data is subjected to learnable multi-scale decomposition. Then, the decomposed multi-scale time series features are input into a parallel two-branch long short-term memory network. One branch learns the global integrated state of the signal embedding over the time span, while the other branch characterizes the independent time series state representation of each decomposed component. The hidden states output by the two branches at each time step are correlated through a regularized least-squares fitting process to generate a set of prediction alignment weights. Based on these weights, the component features are weighted and converged to form a fused hidden state vector for prediction. Finally, the time series prediction result is output through a lightweight linear layer. This invention constructs an inherent structured mechanism while maintaining high prediction accuracy and provides an analyzable intermediate representation basis for the prediction results.