16results about How to "Reduce operation" patented technology

The invention discloses a fine tuning method and system for precise parameters of a pre-training language model. The method comprises the following steps: constructing a robust adapter; a plurality of constructed robust adapters are connected in series to form a robust adapter chain, and the maximum chain node is determined; preparing a weight of a pre-training language model and a data set of a corresponding task label; firstly, each robust adapter in the robust adapter chain is subjected to preliminary training, optimization and approximate fine tuning of parameters of the robust adapters; after the parameters of the robust adapter are approximately fine-tuned, the parameters are combined with the weight of the preorder frozen pre-training language model to complete calculation of a current robust adapter chain node, residual information used for next iteration is generated, and the parameters of the current robust adapter are updated at the same time; and carrying out iterative training, updating residual information until the number of iterations is equal to the maximum node of the robust adapter chain, and obtaining the optimal weight of the pre-training language model and the optimal parameters of the robust adapters, so that more stable convergence and simpler hyper-parameter adjustment can be provided.

The invention provides a large model self-attention optimization method and device, equipment and a medium, and relates to the technical field of artificial intelligence. The method comprises the following steps: obtaining a KQ large matrix obtained after an input sequence passes through a large model self-attention structure, and dividing the KQ large matrix into a plurality of KQ small matrixes; carrying out local attention calculation on each KQ small matrix, and carrying out similarity descending sorting on the KQ small matrixes and other KQ small matrixes to obtain a sorting list; and for each element in the input sequence, according to the calculated sorting list, selecting the local attention corresponding to the KQ small matrix with the similarity ranking in the first r to carry out weighted calculation, and taking the local attention as the attention representation of the current element until all the elements of the input sequence are completed, so as to obtain an attention result after the input sequence is optimized. According to the method, the calculation complexity of the model is reduced, remote context information can be captured, and the accuracy of the model is improved.

The application relates to the technical field of automatic control of unmanned aerial vehicles, in particular to a control system of an unmanned aerial vehicle, which comprises a bus, a self-pilot, a plurality of rudder controllers and a plurality of rudder surface driving mechanisms, the self-pilot and each rudder controller are connected to the bus, each rudder controller is electrically connected with the corresponding rudder surface driving mechanism; each rudder controller is provided with a control distribution table mapped with a flight configuration; during the flight of the unmanned aerial vehicle, each rudder controller determines the corresponding control distribution table according to the flight configuration broadcast by the self-pilot; according to the control distribution table, the weight of each control signal participating in control is determined, and the total control signal is obtained by weighted summation of each control signal participating in control; each rudder controller outputs the total control signal to the corresponding rudder surface driving mechanism to realize the control of the rudder surface driving mechanism. That is, the application scheme can relieve the operation and storage pressure of the self-pilot, simplify the configuration of the self-pilot and improve the universality of the self-pilot.

This invention specifically relates to an information management platform based on Internet of Things (IoT) technology, specifically in the fields of IoT data security and information management technology. It includes: a source-side data entanglement generation module; an edge-side local ledger maintenance module; a business event-triggered anchoring module; and a cloud-based auditing and dispute arbitration module. In this invention, standardized data collection and lightweight chained hash entanglement calculation lock the time-series data from the physical source of data generation. A lightweight algorithm with hardware-level optimization is used, occupying only 32 bytes of volatile memory to cache the latest hash value, adapting to the low power consumption, low computational cost, and small storage characteristics of IoT terminals. The edge side is equipped with a real-time packet-by-packet chain break detection and self-healing mechanism, which can quickly detect and attempt to repair hash chain breaks, preventing tamper-proofing failure.

The invention is suitable for the technical field of earthquake message pushing, and provides an earthquake early warning message broadcasting method, device and equipment and a storage medium, and the method comprises the steps: obtaining the epicenter coordinate, the epicenter depth and the epicenter magnitude of a target earthquake; based on the epicenter coordinates, the epicenter depth and the epicenter magnitude, obtaining historical earthquake region magnitude distribution information of each historical earthquake closest to the target earthquake; determining theoretical radiuses of different earthquake types according to the earthquake magnitude distribution information of each historical earthquake region; based on the theoretical radius corresponding to each earthquake type, determining earthquake magnitude areas corresponding to different earthquake types with the epicenter coordinate as the center; and sequentially sending earthquake early warning messages to users in the corresponding earthquake magnitude areas according to the severity of the earthquake types from large to small. According to the invention, the earthquake early warning information is pushed to the user quickly and accurately.

This invention relates to a marine exhaust gas treatment system, specifically a shipboard biphase absorbent carbon capture system based on a centrifugal phase-coalescing-heat exchanger integrated unit. This invention proposes a highly efficient and low-energy-consumption shipboard carbon capture system based on a centrifugal phase-coalescing-heat exchanger. This technology combines a centrifugal phase separator, a heat exchanger, and a coalescer. The lean and rich liquids flowing from the bottom of the absorber tower undergo heat exchange and centrifugal phase separation in the integrated centrifugal phase-coalescing-heat exchanger. This technology not only reduces the need for reheating, temperature rise, and centrifugal separation after the phase separator, effectively solving the problem of low phase separation efficiency caused by ship rolling, but also effectively reduces equipment size and floor space by combining the phase separator and centrifugal heat exchanger.

This invention relates to the field of battery life prediction technology, specifically providing a method, apparatus, and medium for predicting the remaining life of a battery, aiming to solve the problem of inaccurate prediction of the remaining battery life. To this end, the method includes acquiring the battery's full life-cycle operating parameters and dividing these parameters into training and test sets; selecting highly correlated features from the training set to establish a training set health model; and calculating the actual health HI of the test set based on its features. Test And based on the health model of the training set, predict the predicted health HI of the test set. Pred The two models are compared to select the K closest battery health models; the final lifetime TL of these K closest battery health models is obtained, and the predicted lifetime TL of the test set batteries is calculated. Pred Predicted remaining lifetime RUL = TL Pred -TL Curr , among which, TL Curr This method aims to accurately predict the remaining lifespan of a test cell.

The invention relates to the technical field of twisted yarn monitoring, in particular to a twisted yarn monitoring device of a glass fiber yarn twisting machine, which comprises a yarn twisting machine body, a cross rod is fixedly mounted at the upper end of the yarn twisting machine body, a monitoring device base is fixedly mounted on the outer side of the cross rod, and a monitoring device support is rotatably mounted on the outer surface of the monitoring device base; the monitoring device has the advantages that the monitoring device base is used for supporting and installing the whole monitoring device support, YARN monitoring in the twisting process is achieved through the additionally-arranged monitoring device support, whether the rotating speed of a motor at each spindle position is abnormal or not is monitored at any time through the multiple sets of flash frequency measuring instruments and the yarn breaking automatic stop sensors, and the monitoring effect is good. Abnormal conditions in the curling process can be found in time, the twisting action can be interrupted in time, loss of finished yarn is controlled, and continuous operation of defective products is reduced; by means of the upturning mechanism, it is ensured that the whole monitoring device support can be upturned when yarn twisting is finished and doffing operation is needed.

The invention discloses a student attendance management device and management method, and the student attendance management device not only comprises a registration identification device, a resource manager, an information input device and a management system, but also comprises a cache comparison pool used for information data comparison operation. Aiming at the problem that an existing attendance system needs one-to-one corresponding complex statistical operation in data comparison and statistics, a short byte splicing mode is adopted, student information mac frames, course information mac frames and regional information mac frames are spliced, so that effective data of each information branch is extracted, classified attendance information is obtained through comparison and classification operation, and the efficiency of attendance checking is improved. And only a single area attendance information output mac frame is implanted, that is, the attendance information of a certain course, activity or conference is output through the single mac data frame through the operation of the cache comparison pool, so that the workload of operation, comparison and statistics of attendance data is greatly reduced, and a novel management system integrating time management, area management and personnel management is realized.

The application discloses a kind of privacy calculation arrangement methods, systems, equipment and medium based on dynamic trust, it is related to data security technical field.The method comprises: receiving data processing request, and collecting context data including user trust, environmental risk and data sensitivity;Dynamic trust score is calculated based on context data, and target privacy calculation technology is dynamically selected according to score;Task is dispatched based on selected technology and guard agent is deployed in computing container;Periodically collect running state data as heartbeat to verify by guard agent, if verification fails, key interruption and memory erasure are triggered to fuse task;After completion of calculation, result and access control policy are encapsulated as security capsule;In response to capsule opening request, verify that terminal environment meets strategy, then decrypt rendering result in memory.The application realizes dynamic control from computing request to result use whole life cycle, solves the problem of authorized control.

The invention discloses an image processing method, a display driving circuit and a readable storage medium, and aims to optimize the power consumption performance of a display driving chip. The display driving integrated circuit comprises an image recognition module, a built-in RAM and a main control chip. And the image identification module performs consistency detection on each line of received original image data, and generates a flag bit for each line of data according to a detection result. And the row data with the flag bit is compressed and then stored in a built-in RAM of the display driving integrated circuit. When the original image data is decompressed, the flag bit is detected, and if the flag bit indicates that the data in the row is consistent, only the first compressed data in the row is decompressed and multiplexed into the recovery data in the whole row. According to the invention, by means of flag bit detection and data compression and multiplexing, the calculation complexity in the process of effect processing and decompression of the image received from the master control by the display driving integrated circuit is remarkably reduced, and the method is particularly suitable for a mode in which the display effect requirement is not high and the algorithm processing frequency and complexity can be correspondingly reduced.

The present application relates to a three-dimensional color image sensing device applied to an object to be measured in ambient light. A light emitter is used to emit excitation light with a polarized direction. A first unit pixel is used to sense the excitation light and the ambient light reflected by the object to be measured to generate a first sensing signal and a second sensing signal. A second unit pixel, a third unit pixel and a fourth unit pixel are used to receive the ambient light reflected by the object to be measured to generate an image signal. A processing circuit calculates depth information according to the first sensing signal and the second sensing signal, and generates two-dimensional color information through the image signal to generate three-dimensional color information according to the depth information and the two-dimensional color information. Thus, the present application provides a three-dimensional color image sensing device capable of modeling three-dimensional color images while reducing system operation and data storage requirements.

The invention discloses a motor controller and a space vector modulation method thereof, and a computer storage medium, and the method comprises the following steps: obtaining a voltage instruction under a dq coordinate system, and defining the voltage instruction under the dq coordinate system as a first voltage instruction; converting the first voltage instruction into a voltage instruction under an abc coordinate system, and defining the voltage instruction under the abc coordinate system as a second voltage instruction; zero-sequence voltage offset is added to the second voltage instruction, so that the second voltage instruction is translated to a non-negative range to obtain a third voltage instruction; performing normalization processing on the third voltage instruction to obtain a normalized voltage vector; and mapping the normalized voltage vector into a duty cycle of PWM modulation. The calculation efficiency of the control algorithm can be improved, and the method is more suitable for real-time control of the motor.

Embodiments of the present disclosure provide a decryption method, a variable storage method, and related devices, equipment, media and products. The decryption method monitors the current signal of the chip through the current probe, monitors the electromagnetic signal of the SRAM through the electromagnetic probe, determines the key bit value corresponding to the coordinate value stored in the SRAM based on the signal change mode of the current signal and the electromagnetic signal, and then obtains the decryption result. Through the current-electromagnetic dual-channel acquisition and analysis scheme, the ability to obtain effective information in the decryption process is improved, the redundant information processing process is reduced, and the password cracking efficiency is improved. The variable storage method stores the coordinate values of at least two coordinate dimensions of at least one password intermediate variable in at least two SRAMs, or stores the coordinate values of at least one coordinate dimension of at least two password intermediate variables in the same SRAM, which can avoid local electromagnetic radiation analysis and the storage process does not affect the execution efficiency of the encryption algorithm.

The utility model discloses a mobile invoicing handheld device equipped with a solar charging panel, and belongs to the technical field of invoicing and printing. Comprising a mobile invoicing handheld device body, a data input button, a data display screen, a remote data transmission module, a charging hole, an invoicing paper storage bin, a solar panel, a solar panel connector, a power supply connecting line, an invoicing paper rotating motor, an electronic board, an invoicing outlet, a rotating shaft base, a printing module, a device battery, an invoicing paper fixing rod, an invoicing paper mounting column and a storage bin sealing cover. Solar energy can be utilized to charge the billing device through connection of solar panels, charging energy consumption is reduced, meanwhile, a remote data transmission module is arranged, billing data can be uploaded in real time, backtracking is facilitated, and meanwhile, the billing paper storage bin, the storage bin sealing cover and the sealing cover are arranged to be connected in an inserted mode, so that the billing device is convenient to use. And the storage bin can be conveniently opened to supplement billing paper.

This article discloses a method, device, computer storage medium and terminal for updating channel parameters. In an embodiment of the present invention, for a beam in a stable tracking state, according to the satellite to which the beam belongs, the beam in a stable tracking state is divided into different beam groups; for each divided beam group, more than one first beam is determined according to the beam signal quality; and the channel parameters of the beam group where the first beam is located are updated according to the channel parameters of the first beam. In each beam group of the embodiment of the present invention, it is only necessary to update the channel parameters of other beams according to the channel parameters of the first beam, thereby reducing the calculation of the channel parameters of other beams in the beam group and reducing the amount of calculation of the channel parameters by the processor.