73 results about "Quantized neural networks" patented technology

The application discloses a satellite communication filtering method based on a handheld terminal and belongs to the technical field of satellite communication. The application firstly performs EMD decomposition on a satellite communication signal collected by the handheld terminal, obtains a plurality of intrinsic mode functions and a residual signal, and extracts energy signals of the intrinsic mode functions; then, first and second time-domain attenuation components are respectively calculated through neighborhood ratio detection and range impact detection, time-domain attenuation weight signals are obtained through fusion, and frequency-domain attenuation weight signals are calculated according to the energy signal in-band and out-of-band total energy difference value; each intrinsic mode function is processed by using a trained lightweight neural network, adaptive attention weighting is realized in combination with the time-domain and frequency-domain attenuation weight signals, a reconstructed signal is obtained, the reconstructed signal is superimposed with the residual signal, and finally, a filtered signal is output. The application effectively improves the satellite communication signal filtering precision through the time-domain and frequency-domain two-dimensional weight constraint combined with the lightweight neural network feature focusing.

A quantized neural network circuit. The circuit may include a neuron processing element, the neuron processing element including a first neuron cluster and a second neuron cluster. The first neuron cluster may include: a first binary neuron, having a first input network with a first number of inputs; a second binary neuron, having a first input network with a second number of inputs, the second number being different from the first number; a plurality of multiplexers, each having an output connected to a respective input of the inputs of the first input network of the first binary neuron; and a plurality of flip-flops, each having an output connected to an input of a respective multiplexer of the plurality of multiplexers.

The application discloses a water mist removal method and system of an unmanned sensing device, relates to detection control, and comprises the following steps: collecting temperature data and humidity data of a vehicle-mounted environment in real time through at least two temperature and humidity sensors; performing filtering processing on the collected temperature data and humidity data, and calculating a historical temperature change rate to form a feature vector; inputting the feature vector into a pre-trained lightweight neural network model, performing reasoning, and outputting a water mist level of a current environment, wherein the water mist level comprises four levels of safety, early warning, mildness and severity; determining corresponding air supply temperature parameters and fan rotating speed parameters through a preset mapping relationship table according to the water mist level; calculating a power control signal of a PTC heating sheet by using a fuzzy PID algorithm according to the air supply temperature parameters; and simultaneously generating a rotating speed control signal of the fan according to the fan rotating speed parameters. In view of the fact that a traditional passive demisting mode fails within several minutes in a high-humidity saturated underground environment, the application improves the active demisting effect through hierarchical control.

The application discloses a kind of non-cooperative target centroid intelligent positioning method and system fusing multi-dimensional physical constraint.The method is directed to sparse, noisy point cloud data, and constructs a comprehensive evaluation function containing four-dimensional information of geometry, orbital dynamics, time sequence continuity and surface physical properties;Using a two-stage solution framework of surrogate model and hybrid optimization, first, a lightweight neural network surrogate model is used with a differential evolution algorithm for global coarse search, and then switching to a high-fidelity orbit model combined with an adaptive particle swarm optimization algorithm for local fine search;Finally, the optimal state estimation and its covariance matrix, evaluation decomposition and other decision support information are output.The application improves the accuracy, robustness and computational efficiency of centroid positioning, and enhances the interpretability of the results, suitable for on-orbit servicing, space debris removal and other high-risk space missions.

The application discloses a camera support visual detection method, device and equipment, and belongs to the technical field of image processing and machine vision. The method comprises the following steps: fixing the to-be-detected support on a platform with a positioning mark to complete posture initialization; projecting a coded sinusoidal fringe pattern on the surface of the support; synchronously collecting a deformed fringe image through a monocular camera; calculating an initial wrapped phase by using an edge calculation unit, and performing phase unwrapping by using a lightweight neural network to generate a depth map; extracting a three-dimensional point cloud of a key feature region from the depth map; registering the measured point cloud with a pre-stored CAD model to calculate a spatial error field; performing qualified judgment and defect classification according to the error field, and outputting a control signal and a detection report. The application combines monocular structured light and deep learning to realize high-precision and high-efficiency three-dimensional deformation detection of a low-reflectivity material support, and the system has low cost and strong robustness, and is suitable for online automatic quality detection of a high-speed production line.

This invention discloses a method for assessing hypoxia response based on dynamic evaluation of heart rate deceleration force and HRV, belonging to the field of biomedical signal processing. The method simultaneously acquires electrocardiogram (ECG) signals and environmental and blood oxygen parameters to dynamically correct the risk baseline; extracts frequency domain, time domain, and heart rate deceleration force features from the ECG signals and establishes a dual-channel cross-judgment logic; inputs the judgment results with the current blood oxygen saturation into a lightweight neural network inference output risk stratification; controls a wearable actuator to initiate multi-target stratified sequential stimulation including heat and vibration based on the stratification results, and adaptively updates the individualized hyperthermia prescription based on the intervention efficacy index, ultimately achieving dual-channel synergistic control of oxygen therapy and hyperthermia. This invention effectively solves the shortcomings of traditional methods, such as delayed early warning and inability to dynamically correct in non-networked environments, constructing a closed loop from precise monitoring to multi-target synergistic intervention, significantly improving the body's adaptability and resilience to hypoxia.

The neural network is trained to selectively quantize the weights of its filters into binary or ternary weights. Multiple training iterations are performed, each iteration comprising: quantizing a set of real-valued weights of the filters to generate a corresponding set of quantized weights; generating an output feature tensor based on a matrix multiplication of the input feature tensor with the set of quantized weights; calculating a loss based on the output feature tensor using a regularization function that minimizes the loss if (i) the quantized weights are close to binary weights or (ii) the quantized weights are close to ternary weights; calculating a gradient to minimize the loss; and updating the real-valued weights based on the calculated gradient. Upon completion of the training iteration, the set of weights quantized according to the updated real-valued weights is stored as either a set of binary weights or a set of ternary weights.

The application relates to the technical field of electronic communication, and particularly provides an audio indoor positioning base station system based on multi-parameter fusion. The system comprises an audio signal collection, preprocessing and feature extraction module, a local lightweight fusion module, a cloud collaborative computing engine and a positioning result output module. Through dynamic unloading of a computing load to the cloud and adoption of a lightweight neural network model for local preliminary positioning, intelligent switching of a cloud collaborative mode and a local autonomous mode is realized in combination with a dynamic load decision unit, so that the positioning precision and system robustness are guaranteed while the base station computing complexity and power consumption are reduced.

This invention discloses an AI-powered operating system and method for an ion chamber robot based on edge computing. The system includes an ion chamber body module, which integrates multiple physiological intervention units. These physiological intervention units include at least two or more of the following: a massage component, a bio-detection component, a red light therapy component, a negative oxygen ion generator component, a terahertz wave resonance component, and a graphene heating component. This edge-computing-based AI-powered operating system for an ion chamber robot addresses the pain point of real-time control, ensuring safety and continuity. Local deployment of edge computing nodes enables real-time local processing of user data, achieving millisecond-level latency in controlling the physiological intervention units, thus completely resolving the latency issue of cloud-based control. A built-in lightweight neural network model allows for independent safety judgment and emergency protection during network outages / weak network conditions. Combined with local priority switching logic for multi-mode communication, it ensures continuous and stable operation of the device, guaranteeing service continuity and user safety.

The application relates to the technical field of computer edge computing, and discloses a production equipment abnormality monitoring and self-adaptive control method based on real-time edge computing. The method comprises the following steps: collecting equipment operation, process, environment and maintenance data through a multi-source sensor; performing data preprocessing and multi-scale feature extraction on the edge side; realizing millisecond-level abnormality detection by using a lightweight neural network; positioning an abnormality source by combining a Bayesian causal diagram reasoning engine; and generating optimal control parameters based on a self-adaptive control strategy library or online reinforcement learning, and issuing the optimal control parameters to a PLC for execution. The system is deployed on an edge computing node, and the control loop delay is ensured to be no more than 10 milliseconds. The application realizes the collaborative control of high robustness, self-adaptability and energy efficiency optimization, and significantly improves the product qualification rate, reduces energy consumption and unplanned downtime.

This invention discloses a vision-based robot inspection system for production line defect identification, belonging to the field of manufacturing automation technology. The system comprises: a visual acquisition, collaborative reasoning, execution storage, self-evolutionary update, and heatmap optimization module working collaboratively to achieve automatic identification, location, and marking of defects in production line workpieces. It utilizes lightweight neural networks for synchronous reasoning, combined with a confidence fusion algorithm to improve accuracy, and automatically updates the model based on inspection data. Simultaneously, it optimizes the robot inspection path and process through heatmap analysis. This invention achieves synchronous parallel processing and comprehensive judgment of visual positioning and defect identification, improving inspection accuracy and real-time performance. Its self-evolutionary capability adapts to defect changes, and it intelligently optimizes the inspection path and provides early warnings through heatmap analysis. Ultimately, it constructs a dynamic and intelligent closed-loop quality control system, comprehensively enhancing the reliability and adaptability of production line inspection.

Systems, apparatuses and methods may provide technology for optimizing an inference neural network model that performs asymmetric quantization by generating a quantized neural network, wherein model weights of the neural network are quantized as signed integer values, and wherein an input layer of the neural network is configured to quantize input values as unsigned integer values, generating a weights accumulation table based on the quantized model weights and a kernel size for the neural network, and generating an output restoration function for an output layer of the neural network based on the weights accumulation table and the kernel size. The technology may also perform per-input channel quantization. The technology may also perform mixed-precision auto-tuning.

This invention discloses a preparation and construction method for a high-strength, impermeable coal-based solid waste geopolymer grouting material for filling underground goaf areas, belonging to the technical field of underground engineering filling materials. The method includes three steps: material composition design, core preparation process, and AI dynamic proportion optimization. Using coal gangue and other solid waste as raw materials, component A is prepared through a cascade mechanical activation process, combined with a composite activator component B. A lightweight neural network model is constructed based on the DEGI-BPNN algorithm, and the ratio of components A and B is dynamically optimized in conjunction with underground hydrogeological conditions. This invention solves the problems of high carbon emissions, low solid waste activity, difficult coagulation control, and insufficient intelligent proportioning of traditional materials, achieving high-value utilization of solid waste. The material has a 3-day compressive strength ≥15MPa, a 28-day compressive strength ≥30MPa, and an apparent porosity ≤4%, possessing advantages of high strength, impermeability, and green low-carbon characteristics, suitable for the integrated needs of goaf seepage prevention and support.

This invention discloses a visual detection method for defect screening of dried blueberries, specifically relating to the field of visual inspection technology for agricultural products. The method involves acquiring real-time images of the dried blueberries to be tested and segmenting them into individual images; inputting the individual images into a lightweight neural network model to extract basic feature maps; and calculating an image complexity index that fuses content complexity and context complexity in real time based on the basic feature maps; comparing the image complexity index with a preset threshold; dynamically selecting different inference paths and outputting corresponding inference results; this invention dynamically allocates computational resources through the image complexity index, enabling rapid classification of simple images and fine segmentation of complex images, significantly reducing average computation time while ensuring detection accuracy, thus solving the problem of high-precision models being unable to detect in real time; achieving accurate quantification of detection difficulty through the fusion of content complexity and context complexity; and employing a lightweight model design to reduce hardware costs, meeting the speed requirements of industrial sorting.

This invention discloses a lightweight semantic point cloud reconstruction and visualization method and system for power corridors oriented towards tree obstacle analysis. The method includes: collecting and fusing laser point cloud and image data using a drone to generate an initial point cloud with texture; performing semantic segmentation on the initial point cloud using a lightweight neural network incorporating local feature aggregation and attention mechanisms to identify power line point clouds and vegetation point clouds; obtaining the actual support point coordinates of the towers at both ends of the power line and reconstructing a three-dimensional power line model based on a preset mechanical model; adaptively thinning the density of the vegetation point cloud based on local geometric features to generate a lightweight vegetation point cloud that retains key structures; identifying dangerous vegetation areas and conducting risk assessments based on the distance between vegetation and power lines calculated using the reconstructed model and the lightweight point cloud; and finally, fusing various data and assessment results to construct a three-dimensional visualization scene. This invention achieves high-precision, high-efficiency intelligent identification and visual early warning of tree obstacle hazards in power corridors.

The invention discloses an off-line environment-oriented mobile terminal health state assessment method and system and a mobile terminal, and belongs to the technical field of artificial intelligence and mobile medical treatment. The method comprises the following steps: locally acquiring time sequence physiological signals and symptom text data of a user at a mobile terminal; calling a lightweight neural network model which is preset locally and is subjected to 4-bit integer quantization for processing; fusing the physiological features and the text semantic features through a multi-modal feature fusion module in the model; and finally generating risk assessment and prompt locally. The system and the mobile terminal integrate corresponding modules to realize the method. According to the method, health state evaluation with complete offline, high privacy security and low resource consumption is realized, and the method is particularly suitable for network-free or weak network environments such as deep space exploration and remote areas.

The application discloses an edge computing method based on artificial intelligence, and particularly relates to the technical field of edge computing and artificial intelligence. The method comprises the following steps: deploying a light neural network model on an edge device to realize local data processing; dynamically optimizing the allocation strategy of a computing task between an edge node and a cloud end through a deep reinforcement learning algorithm; and realizing cross-device collaborative training by using an improved federated learning framework to guarantee data privacy. Through real-time data processing on the edge side, the average system response delay is reduced by more than 40%; the task offloading algorithm based on Q-learning can reduce network traffic by more than 30%; and the innovative differential privacy federated learning mechanism reduces the data leakage risk by 60% on the premise of ensuring model accuracy. Through multi-level technical cooperation, the application effectively solves the problems of high response delay, large network bandwidth occupation and insufficient data privacy protection in the traditional cloud computing architecture.

This invention discloses a brain-computer interface (BCI) teaching demonstration system and control method, belonging to the field of BCI and artificial intelligence education technology. The system includes: a multi-channel EEG headband for synchronously acquiring EEG signals and head posture data; an edge AI hub with a built-in NPU accelerator for dynamically gating and adaptively filtering EEG signals based on head posture data, generating a two-dimensional time-frequency graph through complex Morlet wavelet transform, and using a lightweight RepEEG-Net neural network with structural reparameterization and quantization processing to analyze user intentions in real time, while simultaneously generating visualized teaching data; and a microcontroller execution chassis equipped with a real-time operating system (RTOS) for controlling the actions of the execution mechanism through multi-priority task scheduling. This invention achieves low-latency, highly interference-resistant brain-controlled demonstrations by deploying lightweight intelligent algorithms at the edge, and visualizes the algorithm processing process, significantly improving the real-time performance, robustness, and intuitiveness of the teaching demonstration.

This application proposes a video stream cloud-edge collaborative analysis method and device. The method includes: an FPGA edge computing device acquiring a video stream captured by a camera device; the FPGA edge computing device performing real-time key region detection on the video stream using an internally deployed lightweight neural network model to obtain key region detection results and extracting low-resolution feature vectors of the key regions; the FPGA edge computing device, through its built-in video encoding and decoding unit, performing lossless encoding and high bit rate compression on the key region video stream based on the key region detection results, and performing Gaussian blur preprocessing and high quantization parameter compression on the non-key region video stream to obtain compressed video data; and uploading the compressed video data, key region location metadata, and low-resolution feature vectors to a cloud GPU server. This can improve real-time performance and network efficiency.

The present application relates to the oil exploitation operation area-oriented bird voiceprint and visual fusion real-time identification method, belongs to the oil exploitation operation technical field, the method includes: collecting the current monitoring area of the set nature reserve in the current time segment in the field bird song signal, and converting into voiceprint time-frequency graph; acquire the pertinence visual data of each frame of field monitoring picture of the current monitoring area in the current time segment; using the corresponding lightweight neural network intelligent identification of the set nature reserve in the current time segment in the current monitoring area of various different bird data; the present application aims at the technical problem that bird identification lacks scene adaptability, time integrity and data comprehensiveness, for the set nature reserve distributed with oil exploitation operation area, using the lightweight neural network of custom structure design, at the same time, the intelligent identification of various different bird data is completed, so that the above technical problems are solved.

This invention relates to the field of unmanned aerial vehicle (UAV) technology, and more particularly to a lightweight neural network model optimization method for UAV airborne platforms. The method includes: performing structured pruning on the acquired original neural network model to obtain a lightweight model; during the structured pruning process, dual supervision is applied using task loss and knowledge distillation loss; and based on feedback information from the target hardware platform, the pruning strategy is jointly searched and optimized; a dedicated runtime engine is constructed based on the lightweight model; the dedicated runtime engine performs graph optimization and static memory planning on the lightweight model; the lightweight model integrates an adaptive inference scheduling mechanism; the adaptive inference scheduling mechanism dynamically selects exits at different computational depths in the lightweight model to achieve an optimal balance between accuracy and speed. This invention provides a complete solution from algorithm compression, hardware co-optimization, dedicated runtime deployment to dynamic inference scheduling.

This invention relates to the field of deep learning technology, and more particularly to a method and system for post-training channel-mixed precision quantization of neural networks. For each weight channel in each layer of the pre-trained model to be quantized, the invention calculates the initial scaling factor of the weight channel at different bit widths based on the weight range of the weight tensor of that weight channel; optimizes the initial scaling factor of each weight channel in each layer of the pre-trained model to be quantized at different bit widths to obtain the target scaling factor of the weight channel at different bit widths; constructs an optimal bit width allocation integer linear programming problem based on the obtained target scaling factor; obtains the optimal bit width allocated to each weight channel by solving the optimal bit width allocation integer linear programming problem; and quantizes the pre-trained model to be quantized based on the obtained optimal bit width of the channels. This invention effectively improves the processing accuracy of the quantized neural network model for data such as text, images, and audio.

This invention relates to a CBAM-FasterNet fusion module, a personnel trajectory prediction method, and a system in the fields of computer vision target detection, lightweight neural networks, and engineering machinery safety protection. The fusion module uses the CBAM attention module to weight and label feature maps, determines a fixed number of channels by combining information retention ratio, dynamic threshold calculation, and percentile statistics, filters target convolutional channels, and achieves lightweight feature extraction through channel rearrangement and FasterNet differential convolution. An improved YOLOv8n network is constructed, paired with a ByteTrack correlation network and a Transformer prediction module, effectively solving the technical problems of blind channel selection, low feature utilization, poor anti-interference capability in complex industrial environments, and unreasonable computational power allocation that exist when traditional FasterNet replaces the YOLOv8n backbone network C2f module with fixed continuous channels.

This invention belongs to the field of machine vision calibration technology, and relates to a collaborative robot hand-eye calibration method, device, and hand-eye calibration model training method. The collaborative robot hand-eye calibration method includes: acquiring multi-frame calibration board image data obtained by a camera module capturing images of a calibration board; determining the three-dimensional coordinate data of the corner points of the calibration board relative to the camera coordinate system and the six-dimensional pose data relative to the base coordinate system based on the calibration board image data; performing steady-state discrimination and anomaly detection on the three-dimensional coordinate data and six-dimensional pose data corresponding to the multi-frame calibration board image data to obtain steady-state data; performing dimensionality upscaling and condition label embedding on the steady-state data to obtain high-dimensional fusion features; and calculating the high-dimensional fusion features using a hand-eye calibration model to obtain a hand-eye calibration matrix. This invention satisfies targeted data augmentation, filtering, and multi-dimensional information mining based on a small number of calibration points, and uses a lightweight neural network to achieve accurate conversion between the camera and robot coordinate systems based on a small number of calibration points.

This invention discloses a method for resolving and scheduling airspace conflicts among multiple unmanned aerial vehicles (UAVs) based on counterfactual causal reasoning, comprising the following steps: S1, constructing a multi-leader-multi-follower Stackelberg game model; S2, performing equilibrium analysis on the Stackelberg game model using backward induction; S3, modeling the Stackelberg game model as a partially observable Markov decision process; S4, obtaining the counterfactual individual contribution of each UAV within the framework of the partially observable Markov decision process; S5, incorporating the counterfactual individual contribution as an intrinsic reward into the policy optimization objective; S6, obtaining the neuron importance ranking and dynamic pruning threshold; S7, updating the pruning binary mask and reconstructing the policy network; S8, obtaining a lightweight neural network model adapted for real-time deployment on UAV airborne platforms. This invention has the advantages of strong game modeling capabilities, interpretable causal contributions, and ease of real-time airborne deployment.

The application discloses a kind of battery health state cloud edge coordination online monitoring system and method based on high-frequency ripple feature extraction, it is related to battery management technical field, including the following steps: the voltage signal of battery in the process of charging and discharging is sampled, high-frequency ripple voltage signal is obtained, and according to the amplitude of high-frequency ripple voltage signal, the amplification multiple is adaptively adjusted, to output the original ripple voltage signal sequence in the preset quantization range.The application improves signal quality by wavelet threshold and LMS two-stage noise reduction, combined with 12-dimensional multi-feature enhanced information utilization;Real-time monitoring is realized by using lightweight neural network with low computing power;Cloud edge coordination and incremental update mechanism are constructed to improve continuous learning ability;Deep learning is used to enhance prediction accuracy and generalization ability;Based on existing sampling hardware, non-invasive monitoring is realized, and performance and cost are considered.

The application provides a power distribution network state sensing and fault diagnosis method based on embedded edge computing, which comprises deploying an embedded edge computing node on a power distribution network terminal device, collecting multi-element operation data of the power distribution network in real time by the edge computing node through high-precision sensors, and performing local preprocessing on the collected multi-element operation data; deploying a lightweight neural network model on the edge computing node, and realizing real-time sensing of the operation state of the power distribution network and AI / ML fault diagnosis through real-time inference; and interacting the edge computing node with a central station through an edge-cloud collaborative mechanism, wherein the edge-cloud collaborative mechanism comprises a data selective uploading strategy, a federated learning type model optimization, and a central station deep diagnosis supplement, so as to optimize the diagnosis result and the model parameter. The application can realize high-precision real-time sensing of the operation state of the power distribution network and rapid and accurate diagnosis of faults.

The application discloses an automatic driving target detection method based on a lightweight Transform neural network, and comprises the following steps: acquiring RGB image data; inputting the RGB image data into a CBRM module in a trained backbone network for feature extraction, then performing deep feature extraction through a ShuffleBlock stacking structure in three stages in the trained backbone network, processing through an Agent attention module in the trained backbone network, and then performing processing through an SPPF module in the trained backbone network to obtain output features of the trained backbone network; performing up-sampling on the output features of the trained backbone network, fusing the output features with part of output features of the ShuffleBlock stacking structure in three stages in the trained backbone network, then performing optimization through a C2f module in a trained Head network, then performing processing through an AKConv module in the trained Head network, and then performing target detection through a trained detection head to obtain a target detection result. The application can efficiently recognize and detect targets.

The application provides a scientific research project whole life cycle management system, relates to the technical field of project management, and comprises a project data digitization module, which is used for carrying out digital structured processing on original data of a standing project classification link, a contract management link, a cost management link, a progress management link, a performance management link, an application management link and an archive management link, and obtaining corresponding standardized data; a sub-item data synchronization module, which is used for carrying out internal cross-module and external cross-business system synchronization processing on the standardized data of each management link, and obtaining project management integrated data; a project risk intelligent constraint and deviation correction module, which is used for analyzing and intelligently reasoning the standardized data, the project management integrated data and project historical data based on a multi-dimensional strategy reasoning model and a lightweight neural network model, and obtaining a project risk closed-loop management result. The application can improve efficiency.