9results about How to "Improve final performance" patented technology

This invention discloses a method for removing micro-water from polar lubricating oil based on the synergy of electric field polarization and crystal field. The method includes: first, filtering the polar lubricating oil containing trace amounts of water to remove mechanical impurities, then heating it to 25℃~35℃; then introducing it into an electric field polarization unit, where it is polarized using an electric field until the dielectric constant of the lubricating oil increases by 3%~8% relative to its initial value; next, introducing it into a crystal field unit, where it is stirred and mixed to fully contact an inorganic crystalline particle adsorbent with a superhydrophilic surface, using the adsorbent to remove water from the oil; finally, passing it through a fine filtration unit to remove the water-swollen inorganic crystalline particle adsorbent, thus completing the micro-water removal of the polar lubricating oil. This invention achieves efficient and deep dehydration of polar lubricating oil at room temperature, with advantages such as low energy consumption, low cost, no additive loss, and no impact on oil performance.

The invention provides a forming process of a complex special-shaped section launching barrel. The forming process comprises the following steps of production preparation, forming of a functional layer and an inner pipe body, filling of a guide sliding groove and forming of an outer pipe body, overall curing and pre-pipe-hoop forming and aftertreatment. According to the invention, a sharp and complex geometrical shape is converted into a curved surface in smooth transition through the pretreatment procedure of filling wires into the guide chute, so that a perfect base is provided for the subsequent laying of a structural layer, and the hidden danger of overhead is fundamentally eliminated; a vacuum bag technology is combined with hand lay-up forming, and a silicone rubber block is used in a key area for local sealing and flow guiding. According to the combined process, uniform compaction pressure can be applied within the range of the whole structure, interlayer bubbles and redundant resin are efficiently removed, a product with high compactness and low porosity is obtained, and the reliability of consistency of mechanical properties of the product is greatly improved.

The application discloses a kind of unsupervised exposure correction methods, method includes: constructing semantic prompt guide's exposure perception module, the exposure level of image is perceived at visual semantic level;Exposure parameter modulation is constructed pseudo-supervision generation module, according to image exposure level adaptive exposure modulation is carried out, and synthesis is used to guide model training pseudo-supervision image;Construct color degradation compensation exposure correction module, by joint execution luminance correction and color compensation, the luminance distortion in abnormal exposure image and color degradation are recovered;Unsupervised exposure correction network is trained using the network optimization mechanism of multiple loss functions;Based on the unsupervised exposure correction network after training, input image is handled, and image luminance level and color fidelity are effectively improved.The application generates pseudo-supervision image according to image exposure level adaptive to guide model training, and compensates the color information of image while luminance correction, to realize the collaborative recovery of luminance and color.

The invention provides a language model training method and device, and the method comprises the steps: obtaining original dialogue type text data, and converting dialogue type data in the original dialogue type text data, so as to obtain a plurality of pieces of text type data; wherein the original dialogue type text data comprises at least two types of text data; performing text data mixing on the multiple pieces of text type data in a preset mixing mode to obtain target training data; wherein the preset mixing mode comprises a title-free mixing mode and / or a multi-title mixing mode; and inputting the target training data into a to-be-trained language model, and training the to-be-trained language model by using the target training data to obtain a target language model. Through the method and the device, the calculation overhead of the model is controlled while the text expression diversity is ensured, so that the training efficiency and the final performance of the language model are remarkably improved.

The invention relates to the technical field of intelligent decision making, can be applied to business system platforms such as financial science and technology, medical health and the like, and discloses a question answer generation method, device, equipment and medium, and the method comprises the steps: collecting a plurality of question answers generated by questions, collecting the question answers generated by each question into a question and answer sample set, calculating a reward model score and a reference model benchmark score of each question and answer sample in the question and answer sample set one by one, fitting into a confidence coefficient curve, and dividing an answer retention interval; screening answers to the questions according to the answer retention interval, and calculating and updating an answer retention proportion in the question and answer sample set; optimizing the question and answer model according to the answer retention proportion, and generating a plurality of candidate answers of the user question by using the optimized question and answer model; and calculating candidate reward scores, retaining candidate answers corresponding to the candidate reward scores in the answer retaining interval, and performing semantic fusion on the answers to generate a final question answer of the user question. The accuracy of question answer generation can be improved.

The invention discloses a multi-parameter collaborative optimization method for an MOSFET (Metal-Oxide-Semiconductor Field Effect Transistor) driving circuit based on deep reinforcement learning. The method comprises the following steps: constructing a parameterized MOSFET driving circuit simulation environment taking SPICE as a core; a decision framework is established based on a simulation environment, and an MOSFET driving circuit parameter optimization problem is converted into a strategy learning process of an intelligent agent: the process is guided by a discrete action space representing MOSFET driving circuit parameter adjustment, a state space containing driving circuit working conditions and PCB layout parameters, and a composite reward function; the intelligent agent obtains a converged strategy model through strategy learning in a parameterized environment; the strategy model can be called as a decision basis for parameter optimization of the MOSFET driving circuit during deployment. According to the MOSFET driving circuit multi-parameter collaborative optimization method based on deep reinforcement learning, automatic global optimization of high-dimensional parameters and intelligent balance among multiple targets are achieved, and the design efficiency, performance and reliability of the MOSFET driving circuit are remarkably improved.

This invention discloses a hybrid energy laser-powered UAV communication method and system based on reinforcement learning, belonging to the field of UAV energy management technology in 6G air-to-ground communication. Addressing the problem of rotorcraft UAVs struggling to maintain long-term service in hotspot scenarios due to limited onboard battery capacity, a photovoltaic-grid hybrid power supply laser transmitter is constructed to continuously wirelessly charge the UAV. The UAV uses non-orthogonal multiple access (NOMA) to serve ground users. With the goal of maximizing the total system throughput, under constraints such as UAV battery state of charge and user service quality, the UAV trajectory, laser emission power, and fractional-order power allocation exponent are jointly optimized. This non-convex temporally coupled problem is modeled as a Markov decision process, and an improved dual-delay deep deterministic policy gradient algorithm (I-TD3) integrating priority experience replay and n-step reward is proposed for solving it.

This invention provides a low-carbon, non-fired brick made entirely from solid waste and its preparation method, belonging to the technical field of solid waste resource utilization and low-carbon building materials. The non-fired brick is made from specific weight parts of industrial solid waste raw materials, including 30-45 parts steel slag, 30-45 parts blast furnace slag, 10-40 parts calcium carbide slag, and 20 parts water. The preparation method includes: mixing and stirring the above raw materials into a slurry, pressing it into brick blanks, then placing the brick blanks in a carbonization curing chamber, introducing CO2 industrial waste gas at a pressure of 0.1 MPa for 4-24 hours of carbonization curing, and finally transferring them to a constant temperature and humidity environment for 26-28 days of conventional curing. This method utilizes the calcium carbonate and silica gel generated during the carbonization process to construct an early strength framework, and further optimizes the microstructure through the synergistic hydration reaction of the solid waste components during subsequent curing. The resulting non-fired brick has excellent mechanical properties, good durability, and significant carbon fixation effect, while achieving low-cost production.

The application discloses a robot pose estimation method based on deep learning, which solves the contradiction between labor-intensive pose annotation in real-world images and the difference between virtual training domain and real domain. The application proposes an unsupervised learning paradigm, which includes four parts. First, based on rendered robot pose estimation, the rendered image estimated in the current state is updated together with the real image to predict the pose. Then, the robot key point detection is input into the encoder-decoder network structure to predict the confidence map. Then, the projected key points of the pose estimation and the detected key points are converted into a distributed situation to construct a divergence-based objective function. Finally, the divergence objective function is approximated by using the integral based on sampling. The comprehensive design enables the unsupervised learning to stably enhance the network.