475 results about "Gaussian noise" patented technology

The invention provides a vibration signal space-time reconstruction method based on a multi-modal condition diffusion model, and relates to the technical field of vibration signal reconstruction, and the method comprises the steps: firstly collecting structure vibration response through multiple sensors, constructing a multi-dimensional vibration signal matrix, and automatically recognizing a space continuous missing region and a time random missing region; performing coarse reconstruction on the missing region by adopting self-adaptive multi-scale interpolation so as to recover the basic trend and frequency band characteristics of the signal; a pseudo-missing mask is further applied to complete data, a training sample is constructed through a self-supervision strategy, and the model is guided to learn spatio-temporal correlation features and missing modes; in a training stage, a diffusion model is used as a generation framework, Gaussian noise disturbance is applied to a missing region, four types of condition embedding of time, space, trend and frequency domain are introduced in a denoising inversion process, signal periodicity, multi-sensor space coupling, low-frequency change and a physical frequency spectrum structure are respectively described, and the noise is reduced; and high-fidelity signal reconstruction under multi-modal information joint constraint is realized.

The invention provides a self-adaptive bearing fault diagnosis method based on multi-base wavelet fusion. The objective of the invention is to solve the problems of noise reduction, insufficient feature extraction and low diagnosis precision under noise conditions. A Kaisixi University bearing public data set is used as original data, and Gaussian noise with different SNRs is superposed to simulate various noise intensities. And uniformly carrying out length alignment, down-sampling, equal-length segmentation, division and normalization preprocessing. Then, wavelet bases such as sym4, db4, coif5 and the like are adopted for parallel multi-scale decomposition and reconstruction; and adaptively determining the number of decomposition layers and a threshold strategy according to the noise level, and generating a de-noising branch. And performing weighted fusion on the denoising results of the branches, and performing iterative denoising on the residual error. Signals subjected to noise reduction processing are sent to a double-branch convolution-cycle-attention network, a convolution layer extracts features, an LSTM and a self-attention module capture time sequence changes, and accurate recognition of various bearing faults is achieved. The training adopts a segmented attenuation learning rate and an early stop strategy, and the robustness and generalization ability of different SNR working conditions are improved.

The invention belongs to the technical field of seismic exploration, and discloses a robust full-waveform inversion method, system and device based on generative modeling, and a medium, and the method comprises the steps: obtaining seismic observation data; the method comprises the following steps: starting from random Gaussian noise, establishing a model space comprising a plurality of initial velocity models through an unconditional score model, and determining a global optimal initial velocity model of the model space by adopting a strategy search algorithm; de-noising is carried out on the global optimal initial velocity model based on back diffusion, seismological observation data are introduced as observation constraints, the velocity model is updated by calculating the mismatch gradient of forward modeling data of the current velocity model and the seismological observation data, and an implicit condition sample is obtained; performing forward diffusion processing on the implicit condition sample to obtain a speed model for a subsequent annealing time step; and when the annealing process reaches a preset condition, outputting a final speed model. According to the method, the robustness and accuracy of full-waveform inversion are improved, and the bottleneck problem of traditional full-waveform inversion is solved.

The present invention provides a context-enhanced image generation method, and a model training method and system, comprising: acquiring a training image data set; using a preset noise-adding mechanism to perform multi-timestep noise-adding on a training image, wherein a variance of Gaussian noise added at each timestep depends on a current timestep and progressively increases until a clean training image is transformed into standard Gaussian noise so as to obtain a noisy image at each timestep; using a random masking mechanism to generate a mask, and according to the mask, discarding a masked region of the noisy image and discarding an unmasked region of the clean image; and on the basis of reconstructing a discarded region using a non-discarded noisy image and the clean image, training a context-enhanced image generation model so as to obtain a trained image generation model. The present invention effectively improves the context comprehension capabilities of image generation methods, improves image generation quality, and achieves high-resolution diversified image generation.

The invention provides an accurate calculation method and system for the axial pretightening force of an ultrasonic bolt, and the method combines a high-order statistic theory with a nonlinear filtering technology, and solves the problem that the measurement precision of a conventional method is insufficient under the conditions of strong noise and nonlinearity. Particle filtering and Kalman filtering are combined, so that the problem that a traditional Kalman filtering method easily causes sensitivity of an initial value of filtering divergence is solved; a double-Gaussian attenuation model is adopted to fit an echo envelope, the problem that a single Gaussian or single index model cannot adapt to envelope distortion caused by stress change is solved, and a nonlinear state space model is established to adapt to nonlinear echo signal characteristics under dynamic stress; gaussian noise is effectively suppressed by using a third-order cumulant cross-correlation algorithm, the time difference resolution in a low signal-to-noise ratio environment is remarkably improved, and the Gaussian noise and asymmetric interference are effectively suppressed.

Methods for motion-controllable video diffusion include extracting optical flow fields from an input video and computing warped noise by iteratively warping noise between consecutive frames using the optical flow fields. The iteratively warping includes (i) re-Gaussianizing expanded pixel regions by sampling fresh Gaussian noise, and (ii) aggregating contracted pixel regions by merging noise particles and renormalizing variance to preserve spatial Gaussianity. An output video is generated by initializing a diffusion process with the warped noise and iteratively denoising to produce temporally coherent output frames. Various other methods, systems, and computer-readable media are also disclosed.

The invention relates to a training method for providing an underwater vehicle main body structure design model and an underwater vehicle main body structure generation method. The training method of the underwater vehicle main body structure design model comprises the following steps: firstly, generating a noise adding main body structure based on a randomly initialized ordinal number stamp k, a sample main body structure and randomly initialized Gaussian noise generation data; secondly, determining predicted Gaussian noise data based on the ordinal number stamp k, the performance parameters of the sample main body structure and the noise adding main body structure by adopting a main body structure design model; and finally, determining a training loss function based on the generated Gaussian noise data and the predicted Gaussian noise data, and training the main structure design model based on the training loss function. Under the condition that only few sample main body structures exist, enough training sample data used for main body structure design model training can be generated, and model training under a small amount of sample main body structure data is achieved.

The invention belongs to the technical field of battery management systems, and particularly relates to a battery state-of-charge estimation method for resisting a voltage sensor fault, which comprises the following steps: S1, data acquisition and system initialization; s2, state prediction; s3, measurement updating and online compensation are carried out; and S4, recursive circulation is carried out. According to the method, the direct-current bias quantity generated by a voltage sensor due to zero drift is widened into a state variable of a battery model, and the state variable, ohm internal resistance, polarization internal resistance, polarization voltage and charge state of a battery jointly form an expansion state vector; and performing real-time recursive estimation on the extended state vector by adopting an unscented Kalman filtering algorithm, and simultaneously obtaining an estimated value of the offset of the sensor and an estimated value of the SOC of the battery subjected to deviation compensation. Through combination of state extension and unscented Kalman filtering, fault bias of the voltage sensor can be identified and compensated online, non-Gaussian noise interference is effectively suppressed, and accuracy and robustness of battery SOC estimation in a complex industrial environment are remarkably improved.

The invention discloses a tomato leaf disease and insect pest identification method, which comprises the following steps of: acquiring tomato leaf images in different planting environments by using a high-definition camera, performing data enhancement in modes of random overturning, rotation, brightness contrast adjustment, Gaussian noise addition and the like after the tomato leaf images are labeled by agricultural experts, and normalizing the images; constructing a recognition model which is based on a lightweight convolutional neural network such as MobileNetV3 and is fused with an SE attention mechanism module and an FPN multi-scale feature fusion module; dividing the data set into a training set, a verification set and a test set according to 8: 1: 1, training a model by adopting an Adam optimizer, adjusting parameters according to the performance of the verification set, evaluating the model by using the test set, and optimizing according to a result; during actual application, images are collected in real time, preprocessed and input into the model for recognition, and pest and disease type and position information are output and displayed for early warning. According to the method, the recognition accuracy rate exceeds 95%, the hardware requirement is reduced, the robustness and generalization ability are enhanced, real-time monitoring and early warning are achieved, and agricultural economic losses can be reduced.

Approaches presented herein provide for the unconditional generation of novel three dimensional (3D) object shape representations, such as point clouds or meshes. In at least one embodiment, a first denoising diffusion model (DDM) can be trained to synthesize a 1D shape latent from Gaussian noise, and a second DDM can be trained to generate a set of latent points conditioned on this 1D shape latent. The shape latent and set of latent points can be provided to a decoder to generate a 3D point cloud representative of a random object from among the object classes on which the models were trained. A surface reconstruction process may be used to generate a surface mesh from this generated point cloud. Such an approach can scale to complex and / or multimodal distributions, and can be highly flexible as it can be adapted to various tasks such as multimodal voxel- or text-guided synthesis.

The invention discloses a digital fingerprint generation method for terminal equipment, and relates to the technical field of digital fingerprint generation, comprising the following steps: S1, collecting multi-dimensional hierarchical features; s2, privacy enhancement; s3, fingerprint anti-regeneration generation is carried out; and S4, fingerprint life cycle management. According to the terminal equipment digital fingerprint generation method, multi-dimensional fine-grained feature acquisition is realized from a non-sensitive layer, a semi-sensitive layer and a high-sensitive layer, a feature weight is optimized through a federated learning framework, fingerprint drift caused by hardware replacement and software updating is effectively resisted, and the user experience is improved. Dynamic Gaussian noise is injected into numerical semi-sensitive features to balance privacy protection and feature effectiveness, 2048-bit key Paillier homomorphic encryption is adopted for high-sensitive features, meanwhile, cross-scene tracking risks are blocked through scenarized sub-fingerprint generation and an original feature immediate destruction mechanism, and the privacy protection and feature effectiveness are improved. And an anti-duplication architecture fusing a Merkle tree and a bloom filter is introduced to greatly reduce the fingerprint coincidence collision rate.

The invention discloses a flow field key structure detection and feature point extraction method based on an optimized YOLO model, and the method comprises the steps: S1, obtaining flow field time sequence schlieren images of an air-breathing aircraft under different working conditions through a high-speed schlieren collection system in a wind tunnel test; s2, preprocessing the flow field time sequence schlieren image to obtain a data set corresponding to the strong shock wave of the isolation section of the air-breathing aircraft, and dividing the data set into a training set, a verification set and a test set in proportion; s3, constructing an optimized YOLO target detection model, and obtaining an optimal weight model after iterative training and verification; and S4, adopting the optimal weight model to complete target detection and feature point extraction of the test set or the schlieren image to be detected. According to the method, the target area is focused through area cutting, meanwhile, a targeted data enhancement strategy of Gaussian noise, brightness adjustment, contrast ratio adjustment and saturation is designed, a high-quality annotation data set is constructed, and the model detection precision is effectively improved.

The invention discloses a structural health monitoring abnormal data reconstruction method based on a submerged space diffusion model, and belongs to the technical field of structural health monitoring. The method comprises the following steps: firstly, converting one-dimensional structure monitoring time sequence data into a two-dimensional image, mapping the two-dimensional image to a low-dimensional potential space by using a pre-trained variational auto-encoder, and obtaining a potential variable containing semantic information of an original signal; then, defining a forward diffusion process in the submerged space, and establishing an evolution path from the structured features to Gaussian noise; constructing a conditional diffusion U-Net network containing time step embedding and frequency domain conditional coding, and training the network through a mask region selectivity mechanism to learn noise inversion distribution; in a reverse generation stage, a known area forced alignment strategy and a bidirectional resampling mechanism are introduced, and a soft mask smoothing technology is combined. The invention aims to obtain a high-quality time sequence signal which can meet the requirements of subsequent modal parameter identification and long-term performance evaluation by using a generative probability inversion mechanism.

The invention discloses a three-dimensional molecule generation method fusing conditional flow matching and Van der Waals radius constraint, which comprises the following steps: processing a molecule training data set, and extracting a total number of atoms and a training element component histogram; based on the optimal transmission path interpolation, combining the sampling time step and the standard Gaussian noise to construct a noise coordinate and a target condition velocity field; the noise coordinates are input into a continuous flow matching prediction model, node features are extracted through affine transformation modulation, a prediction velocity field is obtained, soft atom type distribution is generated, and the expected Van der Waals radius of each atom type is calculated; calculating flow matching loss through a prediction velocity field and a target condition velocity field, calculating a geometric collision penalty term in combination with an expected Van der Waals radius and a noise coordinate, and constructing a total loss function training model parameter; and defining an ordinary differential equation by using the trained parameters for solving, and outputting a three-dimensional molecular structure file. According to the method, atom space overlapping is inhibited, and the physical rationality and chemical effectiveness of generated molecules are improved.

The invention provides a privacy protection federated learning method for high-time-space-flux medical data, and belongs to the technical field of medical information intelligent processing. According to the technical scheme, the method comprises the following steps that S10, a central server broadcasts a current global model to each participating client, and each client carries out model training locally; s20, updating and adding Gaussian noise to model parameters by the client; s30, guiding the client to align the local feature extracted by the client with the global feature prototype in each round of training; and S40, performing scaling processing on the parameters with low importance. The method has the beneficial effects that the problem of model performance reduction caused by differential privacy noise of a medical institution is effectively relieved, and the clinical prediction accuracy of federated learning under a privacy protection condition is improved.

The invention discloses a multi-objective reinforcement learning man-machine cooperative assembly task allocation method based on neighborhood parameter migration, and the method comprises the steps: building a mathematical model which aims at minimizing the physiological fatigue accumulated value of a human operator and minimizing the maximum completion time for a multi-objective optimization problem of task allocation in a man-machine cooperative assembly system; a multi-target problem is decomposed into N standard sub-problems by adopting a weighting and decomposition strategy, and training is accelerated through a neighborhood parameter migration strategy; each sub-problem is solved based on a near-end strategy optimization algorithm of an Actor-Critic framework, Gaussian noise is added to an Actor network to simulate environment uncertainty, an action mask mechanism is introduced to process priority constraints of assembly tasks, and it is ensured that a generated task allocation scheme is always feasible. According to the method, the convergence speed and diversity of the Pareto solution set can be remarkably improved, and the assembly efficiency and operator fatigue are effectively balanced.

The invention discloses a maximum correlation entropy adaptive dynamic estimation method containing a distributed photovoltaic power distribution network, and belongs to the technical field of power systems. The invention discloses a maximum correlation entropy adaptive dynamic estimation method for a distributed photovoltaic power distribution network, and the method comprises the steps: building a two-stage three-phase distributed photovoltaic physical model containing a feedback control link, and forming a photovoltaic system state equation containing a differential link; establishing a power distribution network state equation based on a Holt's two-parameter exponential smoothing method; the kernel width parameter of the maximum correlation entropy is adjusted in real time according to the measurement error statistical characteristics; a maximum correlation entropy criterion is adopted as a target function, and a robust Kalman gain updating formula is obtained in combination with a Gaussian kernel function and an adaptive weight matrix; and integrating the robust Kalman gain updating formula into the integral framework of the volume Kalman filtering to complete the dynamic state estimation of the power distribution network. By adopting the maximum correlation entropy adaptive dynamic estimation method containing the distributed photovoltaic power distribution network, the problem that an existing dynamic state estimation method is insufficient in estimation precision in the distributed photovoltaic and non-Gaussian noise power distribution network can be solved.

The invention discloses a DDPM-based CT image metal artifact elimination method, and aims to solve the artifact problem caused by a metal object in an existing CT image and improve image quality and diagnosis reliability. A diffusion model of unconditional training is adopted, step-by-step back diffusion repair of an artifact area is carried out in a sinogram domain, an unrepaired area is dynamically adjusted by combining with a metal mask, accurate repair of the artifact area is achieved, and original data of the area which is not affected by artifacts are kept. In the training stage of the system, artifact-free data are gradually converted into standard Gaussian noise through forward diffusion; in the inference stage, data are gradually recovered by utilizing back diffusion, block repair is carried out on an artifact region by combining with a metal mask, and a complete sinogram is generated through region merging. And finally, reconstructing a CT image by using a filtered back projection algorithm, and optimizing boundary transition through a smoothing algorithm to ensure seamless connection between the metal object and surrounding tissues.

The invention discloses a bone injury detection method based on ultrasonic guided waves and a multi-branch convolutional neural network, and the method comprises the steps: firstly collecting UGW signals of bone tissues, and carrying out the standardization preprocessing of the UGW signals, so as to eliminate the amplitude difference between different samples; then, data enhancement strategies such as Gaussian noise disturbance and amplitude scaling are adopted, a finite sample set is effectively expanded, and the robustness and generalization ability of the model are improved; a multi-branch convolutional neural network (TB-CNN) structure is constructed, the network designs independent branches for time domain, frequency domain and time-frequency domain input, deep features are automatically extracted from the respective branches, and feature splicing and multi-level information integration are realized in a fusion module; finally, regression prediction of the bone injury depth is realized through a full connection layer, so that continuous value output is obtained, and the severity of the bone injury can be reflected more finely.

The invention discloses a GNN-MAPPO task unloading method based on dominant value noise addition, which is characterized in that an MLP is changed into a GNN on the basis of the existing MAPPO framework, a multi-agent system can be directly modeled into a graph structure, an interaction relationship among multiple agents can be better established, Gaussian noise is added on the dominant value, the exploration capability of a model is enhanced, and overfitting is reduced. According to the method, the powerful graph structure learning ability of GNN is combined with an innovative dominant value noise adding mechanism, and the mixed reward function is elaborately designed, so that the MAPPO algorithm can more effectively learn a cooperation strategy between agents and optimize time delay and energy consumption in the aspect of edge computing task unloading, and the efficiency of the MAPPO algorithm is improved. And the exploration capability of the strategy and the avoidance capability of the communication risk can be obviously enhanced, so that a more robust and efficient intelligent task unloading scheme adapting to a dynamic environment can be obtained.

The invention discloses a multi-mode emotion fusion robot voice style conversion method and equipment. The method comprises the following steps: extracting emotion feature vectors of text, audio and image input through a multi-mode emotion feature fusion extraction network; constructing a probability path taking Gaussian noise as a starting point and a target Mel spectrum as a terminal, taking the emotional characteristics as a condition guide vector, and predicting a vector field from the noise to target distribution by using a time condition U-Net; wasserstein-2 regularization constraint based on an optimal transmission theory is introduced to learn a smooth and efficient probability flow path; and finally, reconstructing a high-fidelity voice waveform with the target emotion through an acoustic decoder. According to the method, the problems of insufficient emotion expressive force and low generation efficiency in a traditional speech synthesis technology are solved, end-to-end optimization of emotion feature cross-modal mapping and speech generation is realized, and the emotion naturalness and the generation stability of synthesized speech are remarkably improved.

The invention discloses a non-Gaussian noise pollution spiral shaft rod type part measurement signal processing method, which can realize reliable purification of measurement signals, stable separation of bending deviation and spiral raceway morphology items and fitting reconstruction when pollution such as heavy tails, multiple peaks, burrs and sheet abnormal points exists. The method comprises the following steps: acquiring and synchronously registering measurement data of rotation and axial movement, establishing a joint fitting model for decomposing readings into deformation items and spiral raceway morphology items, and expressing the morphology items by multi-order harmonics, and a three-stage process is adopted to complete baseline establishment and coarse elimination, phase locking and fine filtering in sequence, and the harmonic order of the optimal spiral raceway morphology item is determined in a self-adaptive manner and finally reconstructed. The invention provides an effective solution for the problems that the continuous rotary scanning measurement signal of the spiral shaft rod type part is unstable in recovery and is easy to distort in reconstruction under the non-Gaussian noise pollution.

The invention discloses a rolling bearing intelligent fault diagnosis method based on data quality leading. The method comprises the following steps: (1) collecting a vibration signal of a rolling bearing; (2) carrying out preprocessing and data enhancement on the collected signals, specifically, (2.1) optimizing a quantitative mapping relation between a sampling length and a fault period number based on a bearing fault characteristic frequency and a periodic impact theory, optimizing a signal length to be 2048 sampling points, and setting a sliding window overlapping rate to be 30%; (2.2) constructing a five-dimensional aggressive data enhancement strategy covering Gaussian noise injection, amplitude scaling, time translation, random flipping and impulse noise disturbance; (3) carrying out fault diagnosis on the rolling bearing; and (4) according to an experiment result, selecting an optimal data quality optimization strategy and a model architecture to carry out rolling bearing fault diagnosis, and outputting a fault diagnosis result. By optimizing the data quality, the accuracy and generalization ability of rolling bearing fault diagnosis are improved, and the problem caused by insufficient attention to the data quality in the prior art is solved.

The invention discloses a night enhanced imaging and dynamic denoising method for an automobile data recorder, relates to the technical field of vehicle-mounted image processing and computational photography, and is used for solving the problems of insufficient brightness and too strong noise of night videos of the automobile data recorder under a low-illumination condition. Estimating the zero-mean Gaussian noise intensity of a read link and the Poisson noise intensity increasing along with the brightness in the same frame, converting the zero-mean Gaussian noise intensity and the Poisson noise intensity into brightness equivalent parameters, and introducing a unified control function G as the scale reference of all enhancement and noise reduction links; defining a sharing priori H and carrying out unified constraint in dense optical flow, depth time sequence denoising and local tone mapping; and in combination with the motion mask and the background model, performing differential space-time denoising and proportional write-back, and in cooperation with motion blur detection, spatial variation motion kernel and constrained deconvolution and multi-frame detail recharge, obtaining target brightness and outputting an RGB frame after proportional write-back.

The invention discloses an SAM medical image segmentation method based on QR-KAN and MSMDA feature enhancement, and relates to the technical field of medical image segmentation. The method comprises the following steps: carrying out data preprocessing operation of normalization and data enhancement on an image; designing a QR-KAN feature enhancement module and an MSMDA attention mechanism module to transform an SAM image encoder to obtain an encoder, and performing feature extraction and enhancement on the preprocessed image by the encoder; and the decoder decodes the features output by the encoder by using the multi-head self-attention module and the deformable attention module, and the decoded feature sequence is input to the prediction head to output a prediction result. According to the invention, the QR-KAN feature enhancement module, the MSMDA multi-scale and multi-dimensional attention mechanism and the Gaussian noise injection module are designed, so that the boundary precision and detail recovery capability of medical image segmentation can be effectively improved, and the method has strong advantages especially when low-contrast and blurred images are processed.

The invention discloses an automatic driving safety state estimation method and system based on statistical similarity, and the method comprises the steps: firstly constructing a sensor network for sensing the two-dimensional position information and speed information of a moving target: building a system state equation and an observation equation based on the dynamic characteristics of a target vehicle; constructing a local optimization objective function on each distributed node, deducing a local posteriori mean value and covariance update formula, and realizing high-precision state estimation of a two-dimensional target; an alternating direction multiplier method is adopted as a distributed information processing strategy, consistency constraint conditions are introduced, estimation result fusion and global state consistency estimation between nodes are achieved, and vehicle safety state estimation is completed. According to the method, high-precision state estimation and fusion of the two-dimensional target can be realized in a complex environment of non-Gaussian noise and hostile attack, the state estimation precision and the system robustness of the automatic driving vehicle are improved, and the method has a relatively high engineering application value.

The invention provides a point cloud key point generation method based on a key point detection network and a descriptor network, and the method comprises the steps: carrying out the random sampling of the key point detection network, obtaining candidate key points, carrying out the multi-level clustering with each candidate point as a center, expanding a sensing domain, and aggregating the neighborhood information, aggregated information is sent to a double-head attention mechanism to calculate a neighborhood point weight, and more representative key points are selected; the descriptor network receives the attention feature map of the detection network, carries out fusion learning on global information and local information of each key point, and calculates a descriptor and a direction vector for each key point. Secondary screening of the key points is carried out while the descriptors are optimized by using a joint matching loss function, and the key points with strong characteristic characterization force are obtained. Compared with the prior art, the time consumption is shortened by 90%, the key point repetition rate is 70%, the matching precision is 90% or above, and good robustness is achieved under different Gaussian noises.

The embodiment of the invention provides a calibration data acquisition method, a model quantification method, equipment, a storage medium and a computer program product, and is applied to the technical field of model quantification. Specifically, firstly, a Gaussian noise image conforming to Gaussian distribution is generated; then, segmenting the Gaussian noise image into a plurality of image blocks, inputting the plurality of image blocks into a visual Transform model, and obtaining an attention map output by an attention module included in an encoder in the visual Transform model; determining an attention alignment loss value based on the attention map and a predetermined attention priori map; and updating the Gaussian noise image according to the attention alignment loss value and a back propagation algorithm. According to the method, Gaussian noise image updating operation is executed through iteration for multiple times, the Gaussian noise image updated through last iteration serves as a forged image, the forged image serves as calibration data and is added into a calibration data set, and therefore possibility is provided for the calibration data set needed by visual Transform model quantification.

The invention discloses a control method and device of a computing power demand prediction system and a storage medium, and relates to the technical field of system resource management, and the method comprises the steps: splicing input condition features and random Gaussian noise, and carrying out the fusion of the condition features and random Gaussian noise, and obtaining a high-dimensional hidden space representation; performing sequence decoding on the high-dimensional hidden space representation, arranging according to a decoding sequence, and determining a resource demand prediction sequence in a future time period; based on the resource demand prediction sequence and a real load sequence, calculating weighted loss according to a preset quantile, and outputting a composite prediction error value; extracting long-range features of the resource demand prediction sequence and the real load sequence, and obtaining a scalar score through feature scaling processing; and performing weighted fusion on the composite prediction error value and the scalar score to construct a joint optimization target so as to alternately update a generator and a discriminator and iteratively generate a prediction generator model. The problem of low resource utilization rate is solved, and the server utilization rate is improved.

The invention relates to a zero-sample speech synthesis method and device, computer equipment, a storage medium and a program product, and the method comprises the steps: obtaining a target coding feature according to a reference speech and a target text; inputting the target coding features into a stream matching model to obtain a conditional velocity field and an unconditional velocity field; inputting the target coding feature into a prior model to obtain a prior speech feature; obtaining a prior generation flow field according to the prior voice features and the standard Gaussian noise; calculating a KL divergence value between the priori generated flow field and a preset real generated flow field, and taking a moment when the KL divergence value is smaller than or equal to a preset KL divergence threshold as an initial moment of the priori generated flow field; fusing the conditional velocity field and the unconditional velocity field to obtain a fused velocity field; and inputting the fusion velocity field from the initial moment to the target moment and the priori generated flow field into an ordinary differential equation solver to obtain the target speech features, thereby improving the speech quality of the synthesized speech.