7results about How to "Avoid convergence" patented technology

The present application belongs to the technical field of artificial intelligence and multi-modal large language model, and particularly relates to a hallucination reduction method and device for multi-modal large language model. The method comprises the following steps: obtaining multi-expert visual tokens of an original image and text tokens of a text prompt, and inputting the connection of the two into a multi-modal large language model; calculating a similarity matrix of the multi-expert visual tokens and an adaptive clustering threshold, performing hierarchical clustering on the multi-expert visual tokens, merging redundant tokens and retaining complementary visual evidence; constructing a negative sample pool and distributing it to an auxiliary expert with the largest feature difference through a CLIP anchored complementary rejection fine-tuning strategy, and performing specialized training on the corresponding projectors of each expert; and inputting the clustered and fused visual tokens and the text tokens into the multi-modal large language model to generate target text output. The present application solves the problems of information redundancy, large computational overhead and insufficient exertion of expert advantages in the existing multi-expert multi-modal large language model hallucination reduction method.

This invention provides a method for extracting image watermarks based on a single-step inversion diffusion model, comprising: acquiring a synthetic image, wherein the synthetic image is generated by a VAE decoder based on the latent variables output by multi-step denoising of initial noise according to the diffusion model, and the symbol of the initial noise is a symbol mask obtained by encrypting the watermark information used for tracing through an encryption algorithm; mapping the synthetic image to the latent space where the diffusion model denoising operation is located to obtain latent space coding features; performing a symbol classification task based on the latent space coding features without using continuous initial noise regression to predict the symbol mask prediction value of the initial noise; acquiring the decoding algorithm corresponding to the encryption algorithm; and extracting watermark information from the symbol mask prediction value according to the decoding algorithm.

According to the TBM full-circumference supporting shoe based on the cam hinge mechanism, a motor is used for driving a cam plate to rotate and driving a sliding pin and an ejector rod to move, unfolding and folding of the hinge mechanism are achieved, and when the hinge mechanism is unfolded, a hinge can make full contact with a full-circumference rock wall to serve as the supporting shoe to disperse supporting pressure and provide tunneling counter force. By the adoption of the technical scheme, rock supporting in the full-circumference range of a tunnel can be achieved, tunneling counter-force is dispersed to large-area surrounding rock, and the pressure acting on the surrounding rock is effectively reduced, so that the supporting device can be applied to soft and broken surrounding rock zones; and in addition, the occupied space in the machine is small, and the device can be carried in a small-diameter or ultra-small-diameter TBM (Tunnel Boring Machine).

The invention discloses a membrane unit correction constitutive method for explicit dynamic analysis. The method comprises the following steps: providing a tension state stiffness matrix for a data inspection stage; officially calculating, assuming that each integration point is in a tension state at first, and calculating the assumed stress and strain of the integration points by using a tension state stiffness matrix; calling a tool subprogram to calculate principal stress and principal strain of the assumed stress and strain; judging whether the integral point is actually in a tensioned state, a folded state or a relaxed state by adopting a stress-strain combination judgment criterion according to the magnitude relationship between the principal stress and a preset threshold value and the magnitude relationship between the principal strain and the preset threshold value; if the integration point is in a tension state, the actual stress is equal to the assumed stress; if the integral point is in a wrinkle state, calculating the actual stress by adopting a wrinkle state stiffness matrix; if the integral point is in a relaxed state, the actual stress is calculated by adopting a relaxed state stiffness matrix. The method considers the wrinkle and relaxation characteristics of the film, can be used for explicit analysis and calculation, and avoids the problem that the implicit analysis is not easy to converge.

This invention discloses an optimized cell layout structure for a trench MOSFET and its fabrication method. The optimized cell layout structure of the trench MOSFET includes an epitaxial layer, within which multiple trench structures extending in the thickness direction and of different depths are formed. The trench walls are provided with a gate oxide layer and filled with a polysilicon gate. A first well region and a second well region with different extension depths are formed between the trenches, and a corresponding conductivity-type doped region is formed on the surface of each well region. By constructing trench and well region structures of different depths, this invention enables the formation of channel units with different conduction characteristics inside the device, achieving current distribution and flow when the device is turned on, improving current density concentration, enhancing current distribution uniformity, and improving device reliability.

The invention discloses a unit layout optimization structure of a trench MOSFET and a preparation method thereof, the unit layout optimization structure of the trench MOSFET comprises an epitaxial layer, a plurality of trench structures which extend in the thickness direction and are different in depth are formed in the epitaxial layer, the trench wall of each trench structure is provided with a gate oxide layer and is filled with a polycrystalline silicon gate, and the gate oxide layer and the polycrystalline silicon gate are formed in the epitaxial layer. A first well region and a second well region which are different in extending depth are formed between the grooves, and corresponding conductive doped regions are formed on the surface layers of the well regions respectively. According to the invention, channel units with different conduction characteristics are formed in the device by constructing grooves and well region structures with different depths, so that current distribution flow is realized when the device is conducted, the current density concentration phenomenon is improved, and the current distribution uniformity and the device reliability are improved.

The invention relates to the technical field of underground garage safety management, in particular to an underground garage iron nail safety inspection method and device. The method comprises the following steps: S1, constructing a training data set containing three marked values of normality, iron nail throwing and iron nail throwing; s2, importing the training data set into a YOLOv8 target detection model for training to obtain an iron nail recognition model; a round hyper-parameter and a batch hyper-parameter in the YOLOv8 target detection model are determined through a speed delay particle swarm algorithm, and the rest hyper-parameters use default values; according to the speed delay particle swarm optimization, speed delay parameters are increased to serve as a third state option of the particles, so that the speed of the particles at the next iteration times is kept unchanged; s3, obtaining and preprocessing an original video image to obtain an input video image; s4, inputting the input video image into the iron nail recognition model, and outputting a recognition result; and S5, when the identification result is that the iron nail is thrown or the iron nail is thrown, the intelligent inspection vehicle is controlled to process the iron nail, and real-time early warning monitoring and processing can be carried out on the iron nail.