118 results about "Orthogonal transformation" patented technology

The invention relates to the technical field of motor control, and discloses a low-noise flux linkage adaptive feedback control system of a high-speed motor. The system comprises an acquisition module used for acquiring current, voltage and vibration data of a high-speed motor at various operation rotating speeds to obtain operation parameters and noise characteristic data of the high-speed motor; the transformation processing module is used for performing frequency domain orthogonal transformation processing on the operation parameters and the noise characteristic data of the high-speed motor to obtain flux linkage real-time observation data; the self-adaptive analysis module is used for executing multi-mode working condition self-adaptive analysis according to the flux linkage real-time observation data and generating flux linkage control reference values of different rotating speed intervals; and the output module is used for inputting the flux linkage control reference values of different rotating speed intervals into a magnetic-machine coupling control algorithm, and calculating and outputting a motor low-noise driving control signal. According to the method, the problem of magnetic flux path distortion under high-speed operation is effectively solved, and downtime and efficiency loss caused by a traditional off-line calibration method can be avoided.

The invention relates to the technical field of digital watermarking, and discloses a self-adaptive image watermark embedding and recovering system, which comprises an image preprocessing module used for carrying out discrete wavelet transform on a carrier image to be embedded with a watermark, decomposing the image into a plurality of frequency sub-bands, and selecting a low-frequency sub-band as a watermark embedding region; the watermark preprocessing module is used for carrying out orthogonal transformation on the watermark image, so that orthogonality is kept among a plurality of watermarks, and mutual interference of the watermarks is reduced; and the watermark embedding module is used for embedding the watermark into the low-frequency sub-band of the carrier image based on the output of the watermark preprocessing module. According to the method, the concealment and robustness of the watermarks are improved through the technologies of self-adaptive optimization of the watermark strength, redundancy embedding and orthogonal transformation, and compared with the prior art, the interference between the watermarks is effectively reduced, and the resistance of the watermarks to attacks such as compression, noise, cutting and rotation is enhanced.

The invention provides a power transmission network equipment insulation aging state evaluation and life prediction method and system, and relates to the technical field of power equipment state monitoring, and the method comprises the steps: obtaining multi-dimensional operation monitoring data, extracting features through multi-scale time-frequency decomposition, and carrying out the orthogonal transformation to obtain a decoupling feature vector; calculating a mahalanobis distance to obtain an aging state deviation metric value; constructing a stress accumulation damage factor to perform nonlinear modulation on the degradation rate; an aging state trajectory is predicted based on the modified degradation rate and a remaining life is determined. The method can accurately evaluate the insulation aging state and predict the residual life of the equipment, and improves the evaluation precision and reliability.

The invention provides a bandwidth distortion self-adaptive multi-layer scalable feature compressor and a method and system thereof, and relates to the technical field of machine learning. According to the invention, through a code rate distortion joint controller, bit width distribution and fragment size of each layer of a layering number are dynamically decided according to a real-time network state training stage and a precision demand; the input features are mapped to a subspace with more concentrated energy through a layered learnable orthogonal transformation module, and feature channels are clustered and grouped; different bit widths are distributed to the coefficients based on the importance degree through a non-uniform quantization module; dynamically adjusting the fragmentation size and the fragmentation sending rhythm through an RTT adaptive scheduling module according to the round-trip delay and the jitter degree; recording an accumulated reconstruction error through an error tracking compensation module and carrying out residual encoding; and periodically transmitting the anchoring layer through the key frame synchronization module to eliminate accumulated errors. According to the method, the robustness, convergence stability and communication efficiency of distributed training in a complex network environment are remarkably improved.

In order to overcome the defects of an existing low-rank model in non-stationary trend, pseudo-periodic mode and transient anomaly processing, the invention provides a time series prediction method based on self-adaptive low-rank representation, original time series data is decomposed into a low-rank component (L) and a sparse component (S) through robust principal component tracking (PCP) and principal component analysis (PCA) algorithms, and the low-rank component (L) and the sparse component (S) are subjected to low-rank prediction. And adaptive separation of a trend-periodic component and a residual (abnormal / transient) component is realized. According to the method, an orthogonal transformation matrix (A) is constructed, a data-driven orthogonal basis (B) and a Fourier orthogonal basis (UF, VF) are fused in the matrix, and time sequence data are mapped to a low-rank potential space with higher characterization capacity. The method adopts a convolution kernel norm minimization (CNNM) frame for prediction, a learnable dynamic search window is established under the frame, the size (wx, n) of the window is adaptively adjusted according to a sequence local feature (| un |), and through fusion of fast Fourier transform (FFT) decomposition and introduced motion vector information, the motion vector information of the motion vector is obtained. And accurate acquisition of multi-scale time sequence characteristics (such as periodic modes of different frequencies) is realized. Under the low-rank constraint, the method provided by the invention can effectively process the complex dynamic characteristics of high-dimensional sensing data, and meanwhile, the prediction precision of a non-stationary sequence is remarkably improved through a parameter adaptive mechanism.

The block to be encoded is encoded using a first quantization parameter corresponding to a coefficient of each color component in the block to be encoded when it is determined that orthogonal transform processing is to be performed on the coefficient of each color component in the block to be encoded, and the block to be encoded is encoded using a second quantization parameter obtained by correcting the first quantization parameter when it is determined that orthogonal transform processing is not to be performed on the coefficient of each color component in the block to be encoded. A predetermined determination based on the first quantization parameter and a predetermined value is performed, and the second quantization parameter is derived by correcting the first quantization parameter in accordance with a determination result of the predetermined determination.

An encoder includes: circuitry; and memory coupled to the circuitry. In the encoder, in operation, the circuitry: performs calculation that involves multiplication and from which a plurality of discrete integers are yieldable as a calculation result, to derive one of the plurality of discrete integers as a quantization parameter threshold; and when orthogonal transform is skipped for a current block of an image, determines a quantization parameter value for the current block, using the quantization parameter threshold. The calculation involves multiplying a first value by a first fixed value, the first value and the first fixed value being each an integer. The first value is a limit value to be included in a header of a bitstream.

An encoder includes circuitry and memory coupled to the circuitry. In residual coding of a current block to which an orthogonal transform is not applied, when coefficient information flags relating to a coefficient in the current block are encoded, the circuitry: performs a level mapping process to transform the coefficient to a second coefficient by using a predicted value, in which the predicted value is determined based on neighboring coefficients of the coefficient within the current block; encodes second coefficient information flags by Context-based Adaptive Binary Arithmetic Coding (CABAC), each of the second coefficient information flags relating to the second coefficient; and encodes a remainder value of the coefficient with Golomb-Rice code, and when the coefficient information flags are not encoded, in operation, the circuitry: skips the level mapping process; and encodes a value of the coefficient with the Golomb-Rice code.

The invention belongs to the technical field of optical fiber sensing, and relates to an optical time domain reflection monitoring method and device based on forward and reverse signal joint optimization, and the method comprises the steps: obtaining reverse and forward OTDR observation signals of a to-be-measured optical fiber; constructing a reverse OTDR linear signal reconstruction item based on a product of the orthogonal transformation matrix and a to-be-solved first sparse coefficient vector; constructing a forward OTDR linear signal reconstruction item based on the product of the orthogonal transformation matrix and a to-be-solved second sparse coefficient vector; constructing a data fidelity item based on the OTDR observation signal and the linear signal reconstruction item; constructing a joint sparse regular term based on the structural sparsity of the first sparse coefficient vector and the second sparse coefficient vector to be solved; and constructing a spatial constraint regular term based on the continuous smoothness of the reconstructed OTDR linear signal in the space, thereby constructing a joint optimization function and solving the joint optimization function to obtain a first sparse coefficient vector and a second sparse coefficient vector, and further reconstructing the OTDR linear signal to detect the optical fiber to be detected.

To reduce processing delay by simplifying processing.SOLUTION: An encoder (100) includes circuitry circuit and a memory which is connected to the circuitry. In a case where a restriction on the number of context adaptive coding processes allows the context adaptive coding to be collectively performed on a plurality of coefficient information flags, the circuitry encodes the plurality of coefficient information flags by the context adaptive coding and encodes a remainder value of a flag by Golomb-Rice coding. In a case where the restriction on the number of context adaptive coding processes does not allow the context adaptive coding to be collectively performed on the plurality of coefficient information flags, the circuitry skips the encoding of the plurality of coefficient information flags and encodes a value of a coefficient by the Golomb-Rice coding. In a case where orthogonal transform is skipped, the circuitry encodes a plurality of absolute value flags relating to magnitudes of absolute values of coefficients by the context adaptive coding. The plurality of coefficient information flags includes a flag indicating whether or not the coefficient is an odd number or an even number.SELECTED DRAWING: Figure 56

The invention discloses a method for calculating secondary ripple current of a single-phase PWM (Pulse Width Modulation) rectifier, which comprises the following steps of: A, obtaining input voltage us and input current is of the rectifier through sampling, and respectively obtaining orthogonal components us alpha and us beta of the voltage and orthogonal components is alpha and is beta of the current through orthogonal transformation or delay of a quarter period; respectively calculating a voltage amplitude Us and a current amplitude Is; the included angle between the input voltage and the current is phi, namely a power factor angle; b, the active power P and the reactive power Q of the input power of the rectifier are calculated, and then the power factor angle phi of the grid-side current and the voltage is calculated; obtaining the angular velocity omega of the input voltage by using a phase-locked loop; c, according to the inductance value L, the voltage amplitude Us, the current amplitude Is, the power factor angle phi and the angular velocity omega of the PWM rectifier, the alternating current side secondary pulsation power Pshc of the rectifier is calculated; and D, calculating the secondary ripple current of the DC side of the PWM rectifier in real time by using the PWM DC side given voltage Vdcref and the AC side secondary pulse power Pshc. The method is different from a method for obtaining the reference value of the secondary ripple current on the DC side of the PWM rectifier by adopting current sampling and filtering and the like, and the reference value of the secondary ripple current is directly calculated by directly utilizing the existing parameters of the PWM rectifier; and the precision of the secondary ripple current reference value is further improved on the basis of eliminating a current sensor and a circuit thereof.

The invention belongs to the technical field of natural language processing, and particularly discloses a method for constructing a geographic knowledge graph alignment model, which comprises the following steps of: constructing a geographic knowledge graph structure containing multi-dimensional spatial information; creating a corresponding reverse link for the relation link between each geographic entity by using a reverse link generation strategy; distributing corresponding weights to different neighbor entities of each geographic entity by using a spatial perception attention mechanism, and calculating an output feature of each geographic entity by combining an orthogonal transformation matrix of the relationship and the spatial distance; and generating a negative sample through the adversarial network, training a pre-constructed deep learning network by using the negative sample, performing error calculation by using a combination of vector loss and geographic loss as total loss in the training process, and updating a corresponding weight parameter according to an error result until network convergence reaches an expected target, thereby completing training. According to the method and the device, the accuracy of geographic entity alignment in the geographic knowledge graph can be improved.

The present application relates to the technical field of feature vector merging in large model inference acceleration, in particular to a feature vector similarity measurement method and device based on transformation domain, by constructing a new type of calculation path of "transformation-measurement", using an orthogonal transformation matrix to project the feature vector to the transformation domain and then calculating the absolute value error sum, effectively extracting the internal structured features, solving the defect of insufficient accuracy of traditional lightweight measurement indicators, designing a multiplierless hardware acceleration architecture supporting the algorithm, by strictly constraining the element values of the orthogonal transformation matrix within the integer set of {-1, 0, 1}, completely equivalent to the matrix multiplication into simple addition and subtraction operation, completely avoiding multiplication, division and square root operation, greatly reducing the chip area and power consumption overhead, in addition, the pareto optimality of model accuracy and hardware efficiency is also realized, and the configuration capability of searching and customizing the transformation matrix for specific models is also provided.

The invention provides an encoding device, a decoding device, and a non-transitory storage medium. The encoding device encodes a flag by context-adaptive encoding if the number of processing times of context-adaptive encoding is limited to allow context-adaptive encoding of a plurality of coefficient information flags together in both cases of applying a different syntax and skipping an orthogonal transform in residual encoding of a current block, and then performs encoding of the flag by the context-adaptive encoding if the number of processing times of the context-adaptive encoding is limited to allow context-adaptive encoding of a plurality of coefficient information flags. And encoding the residual value of the coefficient by Golomb encoding, if not allowed, skipping the encoding of the flag, encoding the value of the coefficient by Golomb encoding, and if the orthogonal transformation is skipped, encoding the coefficient information flag other than the plurality of absolute value flags after encoding and before encoding the residual value of the coefficient. A plurality of absolute value flags related to whether or not the absolute value of the coefficient is greater than a prescribed value, which is an integer greater than 1, are encoded by context adaptive coding, and if the orthogonal transformation is skipped, the flags further include a flag indicating whether the coefficient is a positive value or a negative value.

A data processing method relates to the field of artificial intelligence, comprising: obtaining a text to be processed and a pretrained language model, the pretrained language model comprising a feature extraction network and a prediction network; extracting features from the text to be processed using the feature extraction network of the pretrained language model to obtain a feature representation of the data to be processed, wherein the feature representation is a complex number; performing an orthogonal transformation on the length-unitized feature representation using the prediction network of the pretrained language model to obtain an orthogonal transformed result, and determining a text prediction result based on the orthogonal transformed result. This application can perform operations on the pretrained language model on a quantum circuit, and by constructing a pretrained language model represented by a complex value, the representation capability of the model is improved, thereby enhancing the performance of the network.

A decoding unit decodes data corresponding to an N×M array of quantized coefficients from a bit stream. An inverse quantization unit derives orthogonal transform coefficients from the N×M array of quantized coefficients by using at least a quantization matrix. An inverse orthogonal transform unit performs inverse orthogonal transform on the orthogonal transform coefficients generated by the inverse quantization unit to generate prediction residuals corresponding to a block of a P×Q array of pixels. An inverse quantization unit derives the orthogonal transform coefficients by using at least a quantization matrix of an N×M array of elements, and the inverse orthogonal transform unit generates prediction residuals for the P×Q array of pixels having a size larger than the N×M array.

The invention discloses an intelligent analysis method based on high-throughput phenotypic data, which comprises the following steps: preprocessing high-throughput genotypic data, calculating the deletion rate and heterozygosity of a sample and an SNP (Single Nucleotide Polymorphism) site, and filtering; principal component analysis is carried out based on the filtered data, and the data dimension is reduced through orthogonal transformation; constructing an evolutionary tree, and analyzing an evolutionary relationship among samples in the group; an analysis result is output and visually displayed; the method supports genotype data files in HMP or VCF formats, provides various visual displays of MAF distribution diagrams, SNP density diagrams, PCA scatter diagrams, evolutionary tree diagrams and the like, and can output analysis result files in various formats; the method improves the efficiency and accuracy of genome data analysis, and has important value in population genetics research and molecular breeding application.

This invention relates to the field of natural language processing technology and discloses a multilingual word vector weighted alignment method based on orthogonal Protodyakonov analysis. The method includes the following steps: S1: Randomly initialize an orthogonal matrix R, and randomly rotate the Chinese word vectors, i.e., X < -XR; substitute the weight matrix A, calculate the weighted orthogonal transformation matrix W according to the provided weighted Protodyakonov analysis, and obtain the weighted aligned Chinese word vector XW2. The Chinese word vectors are downloaded through a given word vector link to obtain Chinese and English word vectors X and Y. A given alignment dictionary L is used to weight sentiment words, with the goal of weighted alignment from X to Y. The method provided by this invention can perform weighted alignment according to the needs of downstream tasks, further improving the performance of downstream tasks, realizing a shared word vector space, and providing a mathematical proof of the weighted orthogonal Protodyakonov analysis.

The present disclosure relates to an image processing device and method capable of reducing an increase in processing delay.A data path of a one-dimensional input signal array of residual data is switched in a forward direction or a reverse direction, and matrix operation of one-dimensional orthogonal transform is performed on the one-dimensional input signal array input via the switched data path. The present disclosure can be applied to, for example, an image processing device, an image coding device, an image decoding device, a transmission device, a reception device, a transmission / reception device, an information processing device, an imaging device, a regeneration device, an electronic apparatus, an image processing method, an information processing method, or the like.

An orthogonal transform unit orthogonally transforms prediction residuals in a block of a P×Q array of pixels, thereby generating an N×M (N is an integer satisfying N<P, and M is an integer satisfying M<Q) array of orthogonal transform coefficients. A quantization unit quantizes the N×M array of the orthogonal transform coefficients using at least a quantization matrix of an N×M array of elements, and thereby generates an N×M array of quantized coefficients.

The invention discloses a displacement measurement method and device based on four-way interference signals, a medium and equipment, and relates to the technical field of semiconductor detection, and the method comprises the steps: carrying out the differential processing of four-way interference signal data with a preset phase relation, and obtaining two-way orthogonal signal data; respectively carrying out extreme value detection on the two paths of orthogonal signal data through a self-adaptive updating strategy to obtain extreme value detection results; performing first normalization processing, orthogonal transformation processing, second normalization processing and phase resolving processing on the two paths of orthogonal signal data in sequence based on an extreme value detection result to obtain phase data; calculating the phase difference between any two adjacent elements in the phase data to obtain phase difference data; and calculating a displacement increment corresponding to each phase difference in the phase difference data through an intelligent direction identification algorithm, and accumulating the displacement increments to obtain a displacement measurement result corresponding to the four paths of interference signal data. According to the invention, a displacement measurement result with higher precision and higher stability can be provided.

This invention relates to the field of biometric recognition technology and discloses a visual recognition-based aquaculture classification system and method, including: a feature preprocessing module, an index mapping module, a feature projection module, and a classification determination module. The preprocessing module acquires the visual feature array of the individual to be identified. The mapping module generates a mask matrix representing the distribution of dirt based on the gradient distribution law. The projection module extracts effective feature components by multiplying element by element and performs dimensionality truncation on the orthogonal transformation basis according to the failure bit index in the mask matrix to construct the projection submatrix. The determination module performs isomorphic dimensionality truncation on the reference feature vector according to the failure bit index, calculates the distance metric value in the dimensionality-reduced feature subspace, and determines the target category of the individual. This invention eliminates the dimensionality collapse phenomenon caused by local dirt by transforming the body surface occlusion state into topological constraint parameters, ensuring the scale consistency of feature comparison under different recognition conditions.

The present application relates to a kind of tree species identification method based on spectral orthogonal transformation and multivariate filtering, comprising the following steps: S1, the high spatial resolution remote sensing image of forest area is collected, and pretreatment is carried out;S2, preliminary interpretation is carried out to the image after pretreatment, and tree species sample is obtained in combination with field investigation, tree species sample is marked on the image after pretreatment, and is divided into training sample and verification sample;S3, the spectral band of remote sensing image is orthogonally transformed, and a set of two two orthogonal feature layers are generated;S4, multivariate filtering processing is carried out to the layer after orthogonal transformation, and filter feature is extracted, and hybrid feature set is constructed;S5, based on hybrid feature set, tree species is identified using machine learning classification method, and classification result and precision statistics are output.The present application reduces the correlation between layers by orthogonal transformation of multispectral data, then highlights the difference of tree species by four filters of texture, low-pass, median and enhanced Lee, to promote the identification of tree species.

An encoder includes circuitry and memory. In both of a first type of residual coding where an orthogonal transform is applied and a second type of residual coding where the orthogonal transform is skipped, wherein when a restriction on a number of CABAC processes allows CABAC coding of a set of coefficient information flags, the circuitry: encodes the coefficient information flags by CABAC; and otherwise, the circuitry: skips the CABAC encoding of the coefficient information flags; and the circuitry encodes a remainder value of the coefficient with Golomb-Rice code when the coefficient information flags are encoded; and otherwise the circuitry encodes a value of the coefficient with the Golomb-Rice code, wherein in the second type of residual coding, the circuitry encodes absolute value flags each relating to an absolute value of the coefficient after encoding the coefficient information flags and before encoding the remainder value of the coefficient.

The application discloses a multi-ground ranging instrument positioning method based on carrier and pseudo-range, and utilizes the carrier and pseudo-range to combine the positioning of the multi-ground ranging instrument, and utilizes high-precision carrier phase observation to improve the DME positioning precision. m The technical scheme of the application comprises the following steps: positioning is performed by using the first to m ground ranging instruments DME1 to DME m ; observation equations of single-difference carrier phase and pseudo-range observation of DME1 and DME2 are constructed; the single-difference observation equations of all visible DMEs at the current time are solved to obtain the user position, wherein the single-difference integer ambiguity is obtained through orthogonal transformation. When the integer ambiguity of the carrier is solved, multi-epoch recursion is adopted to further improve the solving precision of the single-difference integer ambiguity, and meanwhile, the influence of the excessive calculation amount on the real-time performance of the algorithm is avoided.

A method, apparatus, device, and medium for fine-tuning parameters of a pre-trained model are disclosed. The parameter fine-tuning method fine-tunes the parameters of a pre-trained model for natural language processing based on a text dataset. The pre-trained model includes a linear layer with an initial weight matrix. The parameter fine-tuning method includes: applying an orthogonal transformation based on at least one rotation matrix to the linear layer and obtaining an angle parameter matrix based on the orthogonal transformation; constructing an orthogonal fine-tuning layer based on the initial weight matrix and the angle parameter matrix; replacing the linear layer in the pre-trained model with the orthogonal fine-tuning layer to obtain the model to be fine-tuned; updating only the angle parameter matrix using a gradient backpropagation algorithm based on the text dataset to train the model to be fine-tuned; and outputting the trained model to be fine-tuned as the result of parameter fine-tuning.

The present application provides an image decoding apparatus and method, an image encoding apparatus and method, and a storage medium. A decoding unit decodes data corresponding to a first array of quantized coefficients including an N x M array of quantized coefficients corresponding to a first block and data corresponding to a second array of quantized coefficients including an N x M array of quantized coefficients corresponding to a second block. An inverse quantization unit derives a first array of orthogonal transform coefficients from the first array of quantized coefficients using at least a first quantization matrix of elements of the N x M array and derives a second array of orthogonal transform coefficients from the second array of quantized coefficients using at least a second quantization matrix of elements of the N x M array. An inverse orthogonal transform unit inverse orthogonally transforms the first array of orthogonal transform coefficients to generate a first prediction residual of pixels of a P x Q array and inverse orthogonally transforms the second array of orthogonal transform coefficients to generate a second prediction residual of pixels of the N x M array.

The invention belongs to the technical field of chemical engineering, artificial intelligence and model compression, and discloses a uniform quantification method for a large model in the chemical engineering field. On the premise that the expression capability of a model function is not changed, orthogonal equivalent transformation is introduced, combined smoothing processing is carried out on the weight and activation in a linear layer, and activation distribution approximately obeys uniform distribution in a quantization interval by constructing an orthogonal transformation matrix and systematically remodeling the activation distribution, so that the low-bit uniform quantization error is remarkably reduced. And carrying out joint quantization on the smoothed weight and activation by adopting a uniform low-bit uniform quantizer, further introducing a small number of learnable global orthogonal parameters, and carrying out lightweight optimization on module output errors before and after quantization. According to the method, quantitative perception training or large-scale fine tuning is not needed, calculation and storage overhead is low, engineering implementation is simple, inference performance close to full precision can be kept under the 4-bit or 6-bit quantization condition, and video memory occupation and energy consumption in the model deployment process are remarkably reduced.

An image encoding device 1 of the present disclosure includes a neighboring pixel non-reference predictor 11 configured to generate a predicted image by a predetermined neighboring pixel non-reference prediction for each pixel signal of an original image in a block unit, a filter processor 12 configured to perform a low-pass filter process on a prediction signal located at a boundary of a block of the predicted image by using a decoded neighboring signal neighboring to the predicted image under a predetermined control, a prediction residual signal generator 55 configured to generate a prediction residual signal of the block unit by using the predicted image, an orthogonal transformer 14 configured to perform an orthogonal transformation process on the prediction residual signal of the block unit under the predetermined control, and an orthogonal transformation selection controller 25 configured to control the filter processor 12 and the orthogonal transformer 14 and generate a predetermined transformation type identification signal. An image decoding device 5 of the present disclosure performs a decoding process based on a transformation type identification signal at the time of the predetermined neighboring pixel non-reference prediction.

This disclosure provides a method, apparatus, storage medium, and electronic device for quantizing feature data in a model. The method includes: acquiring feature data to be quantized from a target model deployed on a target device, wherein the feature data includes at least two data blocks; performing an inter-block transformation on the feature data using a pre-constructed inter-block orthogonal transformation matrix to obtain inter-block transformed data; performing an intra-block transformation on each data block included in the inter-block transformed data using a pre-constructed intra-block orthogonal transformation matrix to obtain intra-block transformed data; and quantizing the intra-block transformed data to obtain quantized feature data. This disclosure can evenly distribute the energy of the feature data across all blocks, eliminating the influence of high-variance blocks on the quantization sharing index, and redistributing the intra-block data so that the intra-block data is evenly mapped to each codeword interval of the codebook, solving the codebook collapse problem and thus significantly improving quantization accuracy.