294 results about "Quantized neural networks" patented technology

The invention discloses a mobile terminal streetscape image real-time segmentation method based on a lightweight neural network, and relates to the technical field of image segmentation. The method comprises the following steps: firstly, carrying out 320 * 320 adjustment, Z-score standardization, adaptive histogram equalization and 3 * 3 Gaussian filtering preprocessing on an input streetscape image; then, an improved MobileNetV3 backbone network is used, and a five-scale feature map is output in combination with DropBlock regularization through eight feature extraction stages including depth separable convolution and an SE attention module; multi-scale features are fused through a U-shaped structure, and a fusion feature map is generated through up-sampling, element-by-element addition of dimension reduction low-layer features and an attention gating module; and during reasoning, outputting a segmentation mask by using a convolutional layer, Softmax and a conditional random field, and finally performing knowledge distillation, weight pruning, 8-bit quantization and TensorRT optimization. According to the invention, high-precision real-time street view segmentation is realized, the robustness is high, and the method is suitable for different devices and scenes.

The invention discloses an intelligent electric equipment monitoring and optimizing method, and relates to the technical field of electric power, and the method comprises the steps: collecting high-dimensional voltage-current time sequence data and transient event marks, calculating topological invariant features, inputting the topological invariant features to a lightweight neural network model, and recognizing the features of all electric equipment; performing abnormal attribution and anti-fact energy efficiency prediction by using causal reasoning and dynamic regularization regression based on the identified characteristics of each electric device, and constructing a multi-objective optimization function through the abnormal attribution and anti-fact energy efficiency prediction; and based on the multi-objective optimization function, generating an equipment operation scheduling strategy through a deep reinforcement learning agent, based on the operation scheduling strategy, sending a control instruction to the electric equipment, and collecting an operation result feedback in real time for optimization and updating. According to the method, through fusion of topological features, causal reasoning and safety reinforcement learning, the precision, robustness and safety of monitoring and optimization of the intelligent electric equipment are improved.

The invention provides a distributed collaboration-based computing network intelligent agent system and a resource dynamic scheduling method thereof, and belongs to the technical field of computing network integration. The system comprises an edge agent used for sensing local computing power, network bandwidth and task load in real time, predicting task demand fluctuation by using a lightweight neural network, adjusting resource allocation weight in real time in combination with network topology change, and executing a preliminary task scheduling decision; the regional collaborative agent is used for aggregating multiple edge node states based on federated learning, generating a cross-node resource scheduling strategy, verifying the credibility of a computing power transaction smart contract and determining a cross-domain resource allocation scheme; and the cloud management agent is used for constructing a global resource portrait model according to the information provided by the edge agent and the regional collaborative agent, performing long-term strategy optimization, issuing global strategy information, and constructing and updating a computing power transaction smart contract based on a preset computing power transaction smart contract template. According to the invention, multi-level refined scheduling of computing network resources is realized.

The invention relates to the technical field of neural network model construction, in particular to a lightweight neural network model construction method, which comprises the following steps: based on a target task data set, in a lightweight basic operator library comprising a depth separable convolution, an inverted residual structure and an attention mechanism module, constructing a lightweight neural network model; and searching and jointly optimizing network structure parameters and weight parameters through the differentiable neural architecture to obtain an initial lightweight network model. And deploying the initial model in a target hardware simulation environment, and generating a Pareto optimal model cluster through structural re-parameterization and hardware-aware progressive channel pruning iterative optimization by taking model precision, reasoning delay and memory occupancy as collaborative optimization targets. And selecting a reference student model from the clusters according to deployment constraints, constructing a distillation framework taking the initial model as a teacher model, and performing fine adjustment by adopting a mixed strategy fusing multi-dimensional distillation loss to obtain a final model. The model gives consideration to precision and efficiency, and the detection efficiency and the quality control level are improved.

The invention discloses a millimeter-wave radar sparse point cloud-based human body posture recognition method, which belongs to the technical field of human body posture recognition, and comprises the following steps of: acquiring three-dimensional point cloud data of human body actions through a millimeter-wave radar, and preprocessing the three-dimensional point cloud data; and training a lightweight neural network model by using the preprocessed point cloud data, wherein the model comprises an edge convolution module and a grouping sparse Transform encoder module. The edge convolution module extracts spatial geometric features and detects dynamics, static postures are directly classified, dynamic postures capture a time sequence dependency relationship through a grouping sparse Transform module, attention calculation is only carried out on frames with feature changes exceeding a threshold value, and mean pooling aggregation is carried out on other frames. And finally, classifying the human body postures based on the spatial geometric features and the time sequence dependency relationship to obtain a classification result. The device is simple in structure, accurate in recognition and suitable for efficient deployment of edge equipment.

The invention relates to the technical field of intelligent gateways, and discloses a multi-mode industrial Internet of Things intelligent gateway based on edge computing and an implementation method thereof, and the method comprises the steps: obtaining sensor data streams of a door magnetic sensor, a human body sensor, a temperature and humidity sensor and a smoke detector in an intelligent region; constructing a space-time fusion data matrix based on the sensor data stream; inputting the space-time fusion data matrix into an edge layer quantization neural network, a fog layer recurrent neural network and a cloud layer large model for distributed reasoning to obtain a reasoning result set; and performing priority queue scheduling and zero-copy transmission on emergency events, important events and conventional events in combination with the reasoning result set to generate a control instruction sequence, so that parallel operation of data acquisition and processing is realized, and the intelligent level, the response performance and the operation reliability of an intelligent regional Internet of Things system are improved.

The invention relates to the technical field of gas detection, in particular to a gas detection system and method based on an AI technology, and the system comprises a differential sensing module, a drift compensation processor, a dynamic risk assessment unit and a cloud-edge collaborative learning framework. Compared with the prior art which adopts a single-sensor independent compensation scheme and has the defects that the false alarm rate is high and long-term drift is difficult to correct under environmental interference, the scheme innovatively introduces a dual-mode differential sensing architecture and an adaptive drift compensation algorithm, a reference sensor is packaged in an inert gas environment, and the reference sensor and a main sensor are subjected to real-time collaborative analysis; the pre-trained lightweight neural network is combined to dynamically eliminate temperature and humidity cross interference, the detection precision and stability under complex working conditions are remarkably improved, and the aging distortion problem of the semiconductor sensor is effectively solved.

The invention relates to a lightweight neural network model implementation method and system for small target detection in a complex aerial photography scene, and belongs to the technical field of computer vision and unmanned aerial vehicle target detection. In order to solve the problem of low detection precision caused by small target size, complex background, easy feature submerging and the like in an aerial image of an existing unmanned aerial vehicle, the method comprises the following steps: constructing a multi-scale adaptive hierarchical feature enhancement module MSAHFE, and combining multi-scale pooling and differential edge enhancement to improve feature sensitivity; constructing a lightweight feature selection module LAFS based on an attention mechanism, and screening high-correlation features by using a space and frequency double-domain attention mechanism; a lightweight detection head LWDeect is constructed, and shared packet convolution and self-calibration convolution are utilized to reduce the calculation complexity. The method has the advantages of being high in detection precision, small in parameter quantity, high in reasoning speed and the like, and is suitable for complex aerial photography and other application scenes needing real-time small target detection.

The invention discloses a construction site three-dimensional modeling method based on multi-sensor fusion. The method comprises the following steps: synchronously acquiring data by using a laser radar, a camera, an infrared sensor, a GPS (Global Positioning System) and an IMU (Inertial Measurement Unit); the method comprises the following steps of: performing denoising and filtering operation on collected original data, aligning a timestamp and a space of the collected data, fusing the collected multi-source data, and aligning and fusing the processed multi-source point cloud data through a point cloud registration and fusion algorithm to generate a complete three-dimensional point cloud model; generating a grid model by using a surface reconstruction algorithm based on the point cloud data, and performing texture mapping in combination with texture information collected by a camera; the model is simplified and the precision of the model is improved through a model optimization technology, and a reliable three-dimensional visualization foundation for construction management and analysis is formed; real-time detection and modeling of dynamic targets such as mobile equipment and workers on a construction site are realized, and the real-time response capability of the system in a complex construction site environment is ensured through combination of a lightweight neural network architecture and an edge computing technology.

The invention discloses a power transmission tower attitude monitoring and vibration intrusion identification method based on multi-source data fusion, and the method comprises the steps: carrying out the deep fusion of GPS displacement data and accelerometer inclination angle data, employing a dynamic weight fusion mechanism: adjusting a fusion coefficient alpha in real time according to the GPS signal quality, guaranteeing the accuracy and stability of data fusion, and at the same time, carrying out the real-time adjustment of the fusion coefficient alpha, and carrying out the recognition of the vibration intrusion of a power transmission tower. And continuously optimizing the error propagation model by taking the final inclination angle theta final of the power transmission tower as a state variable, updating the attitude estimation of the power transmission tower, and realizing + / -0.01-degree high-precision attitude monitoring. According to the invention, the system can output a high-precision three-dimensional attitude angle in real time by combining the rapid dynamic response of the accelerometer and the long-term stability of the GPS. Besides, an LCD-MPE (feature extraction technology) is introduced, lightweight neural network classification is combined, vibration signal sampling, noise reduction and abnormal signal segment identification, analysis and extraction are carried out on the transmission tower through an MEMS accelerometer, and second-level detection and behavior type discrimination of invasion behaviors around the transmission tower are realized. The core technical barrier from'inability perception 'to'accurate identification' is overcome, and subversive upgrading of power facility safety protection from'passive disposal 'to'active defense' is promoted.

The present invention discloses a neural network construction method having average quantization mechanism that includes steps outlined below. A weight combination included in each of network layers of a neural network is retrieved. A loss function is generated according to the weight combination of all the network layers and target values. Corresponding to each network layers, a Gini coefficient of the weight combination is calculated and the Gini coefficients corresponding to all the network layers are accumulated as a regularized correction term. The loss function and the regularized correction term are merged as a regularized loss function to perform training on the neural network accordingly to generate a trained weight combination of each of the network layers. A quantization is performed on the trained weight combination of each of the network layers to generate a quantized neural network, in which each of the network layers thereof includes the trained weight combination.

The invention discloses a debris flow infrasound recognition method based on a Mel spectrogram and a lightweight neural network, and belongs to the field of image data processing, and the method comprises the steps: collecting infrasonic wave signals generated by on-site debris flow and other events, and generating a Mel spectrogram data set; constructing a channel alignment fusion bridge; constructing an improved MBConv block; constructing a debris flow infrasound recognition network based on the channel alignment fusion bridge and the improved MBConv block; and training by using the Mel spectrogram data set to obtain a debris flow infrasound identification model, and identifying the infrasound signal to be identified. According to the method, the characteristics of debris flow infrasonic waves are combined, samples in the data set are constructed into the infrasonic mel spectrogram, the debris flow infrasonic recognition network is designed based on the multi-stage lightweight feature extraction module, the infrasonic mel spectrogram is classified and recognized, the calculation complexity can be effectively reduced, the parameter quantity can be reduced, low-power-consumption equipment can be adapted, and the method is suitable for popularization and application. The method can improve the classification accuracy, and is especially suitable for the real-time early warning scene of debris flow.

The invention provides a bone mineral density measurement method based on a segmented large model and a lightweight semi-supervised model, which comprises the following steps of: acquiring an original hip joint X-ray image, and preprocessing the original hip joint X-ray image to obtain a data set; performing image segmentation on the data set by using a preset segmentation large model to obtain a final femoral head mask pattern; cutting and identifying the original hip joint X-ray image by using the final femoral head mask pattern to obtain a mixed supervision sample set only containing the unilateral femoral head; training a preset semi-supervised lightweight neural network model by using the mixed supervised sample set to obtain a trained semi-supervised lightweight neural network model; and predicting the bone mineral density of a newly input hip joint X-ray image by using the trained semi-supervised lightweight neural network model. According to the method, the segmentation large model SAM and an interactive prompt mechanism are introduced, the femoral head segmentation precision and generalization are remarkably improved, and high-precision prediction of the bone mineral density under the small sample condition can be realized by using the lightweight semi-supervised regression model.

The invention relates to the technical field of strip mine side slope slippage trend early warning, and particularly discloses a strip mine side slope slippage trend early warning system based on active micro-vibration excitation, which comprises the following components: an excitation unit which is composed of a frequency-adjustable micro-vibration exciter which periodically applies low-amplitude controllable mechanical disturbance signals to a side slope rock mass, the rock-soil medium structure response is excited, and the early warning sensitivity is improved; the sensing unit comprises a three-axis acceleration sensor, a displacement meter, a GNSS (Global Navigation Satellite System) sensor and a ground acoustic wave sensor, and is used for recording response change in a vibration disturbance signal propagation process and obtaining data; the analysis unit comprises an edge intelligent terminal, a slippage trend identification model built in the edge intelligent terminal and a lightweight neural network, and the slippage trend identification model performs sensing processing on the data and outputs a slippage trend label; and the communication control unit is used for ensuring that the data streams of the units are stable and cooperatively run with the modules.

The invention relates to a lightweight neural network target detection model optimization method for an embedded device, and belongs to the technical field of target detection, and the method comprises the steps: determining the hardware constraint and detection task boundary of the embedded device; adjusting a network infrastructure based on hardware and task characteristics; network parameter redundancy is eliminated; designing a scene adaptive dynamic feature selection mechanism; compressing parameter storage precision; optimizing calculation-intensive operation; adjusting a memory access mode; an embedded deployment framework is adapted; and constructing a closed-loop iterative optimization process. The method is beneficial for realizing lightweight and efficient operation of the model, improving detection precision and stability, reducing memory occupation and power consumption, and meeting hardware constraints of embedded equipment.

The invention relates to a longitudinal federal learning method and device based on feature importance and electronic equipment, and aims to realize a local privacy budget generator of a client through a lightweight neural network by taking local feature importance as drive. Therefore, a cooperative closed loop from'feature-level privacy budget 'to'embedded dimension disturbance' to'suggested privacy budget uploading 'to'server-side gradient noise addition' is realized, and meanwhile, discrete grading and suggested privacy budget disturbance mechanisms are introduced to prevent the server-side from reversely deducing privacy contents, so that the differential privacy security and practicability of the whole system are improved.

The application discloses a tea variety identification method and system based on machine vision, relates to the technical field of variety identification, and has the advantages that in the system operation, the surface and part of the internal structure information of tea are captured through a multi-band image acquisition device, original image data are acquired, the exposure characteristics in the real-time detection of environmental illumination and collected images are detected, the light source intensity and distribution are automatically adjusted, the camera parameters are calibrated, the complete contour and the foreground image area of tea are extracted, and are fused into a unified mixed feature vector, so that the core data expression of variety identification is provided, the multi-dimensional features of the tea image are identified and judged through a lightweight neural network model, the specific variety category and the confidence are output, the corresponding sorting control instruction is generated, the system decision control is carried out on the sorting device, including a mechanical arm and a pneumatic sorter, the tea is put into the corresponding packaging channel, and the track state is fed back, so that the closed-loop control is formed.

The invention relates to the technical field of energy management, and discloses a thermal management system energy consumption optimization method based on a lightweight neural network, which comprises the following steps: step 1, collecting operation data and environment information of an energy storage system, constructing a unified feature input set, performing timestamp alignment and normalization processing on the collected operation data and environment information, and obtaining a feature input set; periodic time features are extracted, and a structured feature matrix used for model input is formed; step 2, based on the structured feature matrix, inputting the structured feature matrix into a long-term trend prediction model to predict future charge and discharge starting time and deadline; and step 3, extracting a prediction result of the long-term trend prediction model. The dynamic cooling regulation and control strategy based on the real-time operation data and the environment parameters is adopted, the temperature prediction model and the cooling load demand analysis mechanism are established, the operation state of the cooling equipment is adjusted according to needs, and the technical effects of flexible regulation and control according to actual needs and reduction of overall energy consumption are achieved.

The invention provides a cable testing method and a small handheld cable tester. The method comprises the following steps: carrying out wavelet transform noise reduction, Z-score standardization and time sequence dynamic regularization on original cable data; then constructing a geometric deep learning network by using a dispersed self-organizing structure of the rotating cube set, and extracting multi-modal depth features; fusing different modal features through a spiking neural network gating mechanism and a cross-modal attention mechanism; and finally, fault classification is carried out based on a differential evolution optimized neural network, and fault location is realized by combining a time sequence generative adversarial network and a dynamic probability neighborhood growth clustering algorithm. The system also evaluates the data contribution degree of each modal through information entropy and mutual information analysis, and optimizes the model performance by using adaptive weight distribution and lightweight neural network technologies. According to the invention, various fault types such as cable breakage, short circuit, insulation aging, poor contact and the like can be identified and positioned with high precision, and the efficiency and the accuracy of cable maintenance are remarkably improved.

The invention relates to a video conference echo suppression method based on a low-delay adaptive learning model. The method comprises the following steps: acquiring acoustic scene noise features, and matching a lightweight LSTM neural network model by using the acoustic scene noise features; the method comprises the following steps: collecting real-time sound signals of a video conference, preprocessing and extracting multi-modal features of the real-time sound signals; performing adaptive filtering by using an NLMS algorithm to estimate an echo path between a near-end microphone signal and a far-end reference signal of the real-time sound signal, performing convolution on the echo path and the far-end reference signal to generate an echo estimation signal, and subtracting the echo estimation signal from the near-end microphone signal to obtain a preliminary residual signal; learning a gain by using the lightweight LSTM neural network model, and further suppressing the residual signal by using the gain to obtain a time domain enhanced signal; comfortable noise is introduced, and incremental learning is adaptively triggered. And efficient linear and nonlinear echo cancellation is supported.

Embodiments relate to converting functions or function call instructions of a first neural network (NN) model to graph modules. Relationships between inputs and outputs of the of graphing modules are analyzed and a second NN model in the form of a directed acyclic graph (DAG) using the graph modules corresponding to the first NN model. Markers are added to the graph modules in the second NN model. Calibration data is generated by collecting input values and output values of each of the graph modules by using the markers. A scale value and an offset value applicable to the second NN model is determined. Based on the scale value and the offset value, a third NN model including weight parameters quantized in the form of integers are generated and training is performed to update the weight parameters of the third NN model.

The invention relates to a multi-modal emotion recognition fusion and communication method oriented to real-time human-computer interaction, and aims to solve the problems of asynchronous modal time sequence, non-uniform feature dimension, single fusion mode, high communication delay and the like of an existing emotion recognition system and improve the recognition precision and real-time performance of the system. According to the method, voice, images and physiological signals are synchronously collected through a multi-modal input module, feature alignment is achieved through time sequence interpolation, dynamic time warping and space coordinate mapping, double-domain feature fusion is conducted in combination with a time path network and a space path network, emotional state judgment is completed through a lightweight neural network, and the emotional state recognition accuracy is improved. And outputting six types of basic emotions and confidence coefficients. Meanwhile, a low-delay communication protocol based on UDP clipping extension realizes rapid feedback of emotion data, and in combination with modal priority scheduling, data compression and bandwidth sensing mechanisms, high-efficiency and low-delay transmission is ensured, and the real-time interaction requirement in a weak network environment is met. The method has the advantages of high accuracy, low power consumption, low time delay and flexible deployment, is suitable for various real-time interaction application scenes such as voice assistants, virtual customer service, emotion accompanying and telemedicine, and has wide application value and market prospect.

The invention relates to the technical field of industrial intelligence, edge computing, embedded artificial intelligence and the like, and provides an industrial intelligent edge computing all-in-one machine integrated with various lightweight neural network models and an implementation method. A plurality of neural network models are trained on a host to analyze industrial field signals, the models are compressed and quantized, and a plurality of lightweight neural network models are obtained; the compressed model file is subjected to format conversion and deployed on a microcontroller platform, a microcontroller collects insulation partial discharge signals, electrical parameters, mechanical vibration and other signals, and when the microcontroller platform operates, neural network reasoning is executed to achieve edge AI calculation. A reasoning result is transmitted to the control output module through the edge AI module. The module generates corresponding control instructions or alarm information according to different model results, sends a maintenance prompt to equipment with serious insulation aging, adjusts the power level of production equipment according to an energy efficiency optimization result, reminds emergency maintenance of a circuit breaker fault, and triggers filtering compensation or alarms for electric energy quality abnormity.

The invention provides a visual tactile sensor, a geometric calibration method and computer equipment, and the visual tactile sensor comprises a base which is internally provided with an imaging unit; the light source module is arranged on the base and comprises 3N single-point light sources which are arranged around the imaging unit and can be independently controlled, and N is larger than or equal to 1; and the elastic housing is arranged on the base in a covering manner, and the imaging unit and the light source module are wrapped in the elastic housing. By constructing the novel curved surface visual tactile sensor and the matched geometric calibration method thereof, the problems that a traditional plane sensor is weak in multi-direction contact sensing capacity, difficult in curved surface calibration and the like are effectively solved, data generated by a physical model are used for training a lightweight neural network, and the accuracy of the curved surface visual tactile sensor is improved. Real-time surface normal prediction under single-frame three-color image input is realized, and the precision and reasoning speed of three-dimensional reconstruction are considered while hardware dependence and calibration cost are remarkably reduced.

The invention discloses a vegetation type real-time monitoring method for unmanned aerial vehicle platform lightweight, and relates to the cross technical field of unmanned aerial vehicle remote sensing and edge intelligent calculation. In order to solve the defects of high model calculation complexity, difficulty in real-time operation at an unmanned aerial vehicle end, low vegetation small target detection precision, poor classification algorithm robustness and lack of a dynamic adjustment mechanism adaptive to unmanned aerial vehicle resource constraints in the prior art, the technical scheme provided by the invention comprises the following steps: receiving a multispectral image and pose information; synchronous pairing is carried out; executing dynamic illumination compensation on the multispectral image; calculating and fusing a vegetation index based on a compensation result; inputting the fused image into a lightweight neural network model for classified reasoning; generating a vegetation type distribution map and completing registration; and a result image is transmitted back to the ground station through the communication module. The method is suitable for carrying out high-precision and low-delay real-time remote sensing monitoring work on various vegetation types on a resource-limited unmanned aerial vehicle platform.

The invention provides a multi-sensor adaptive anti-interference working well environment detection system and method, and the system comprises a single-chip microcomputer control panel, a temperature and humidity monitoring sensor, a four-in-one combustible gas detector, a dust particle detector, a differential pressure transmitter, and a cloud server. The temperature and humidity monitoring sensor, the four-in-one combustible gas detector, the dust particle detector and the differential pressure transmitter are in circuit connection with the single-chip microcomputer control panel, the single-chip microcomputer control panel is in wireless communication with the cloud server, the detection method dynamically adjusts the weight distribution of the sensor through the feature level attention module, a lightweight neural network and a hard redundancy check rule are combined, and the detection accuracy is improved. Compared with the prior art, the method has the advantages that the low false alarm rate is kept, the model size is low, and the method is suitable for industrial field deployment.

The invention discloses an industrial image super-resolution reconstruction method based on a physical consistency self-supervision mechanism, and belongs to the technical field of industrial automation and machine vision. In order to obtain a high-resolution industrial image, the non-uniform degraded image of the industrial scene image is generated by the non-uniform degraded image generation method for constructing the industrial scene image; constructing an imaging fuzzy degradation model based on a point spread function, and simulating a fuzzy degradation process of an industrial scene image; constructing a lightweight neural network composed of a multi-layer convolution module and a pixel recombination module; constructing a composite loss function of basic reconstruction loss, structure edge perception loss and optical blur consistency loss, collecting industrial scene images, preprocessing the industrial scene images, and inputting the industrial scene images into a lightweight neural network for training to obtain a trained lightweight neural network; and collecting an industrial scene image, processing the industrial scene image by adopting a multi-view spatial transformation and integration mechanism, inputting the industrial scene image into the trained lightweight neural network, and outputting an enhanced image of the industrial scene image.

The invention discloses a full-closed-loop control method for a humanoid robot joint module based on an FPGA (Field Programmable Gate Array), which relates to the technical field of robot control and comprises the following steps of: performing second-order active low-pass filtering and programmable gain amplification on analog signals of an encoder, a gyroscope and a force sensor to convert the analog signals into digital signals; a multi-dimensional feature sequence is constructed, temperature signals of a joint driver and an encoder are collected through a temperature sensor interface, a temperature drift model is constructed, and zero drift compensation parameters are obtained; further obtaining a joint elastic deformation error, and applying the joint elastic deformation error to an encoder signal for online offset correction; outputting a reconfiguration region selection signal based on the lightweight neural network; and triggering bit stream local refresh corresponding to the PR Region according to a region selection signal output by the lightweight neural network. The problem that a control loop is difficult to eliminate flexible errors due to the fact that a traditional rigid model or a constant parameter model is difficult to reflect deformation characteristics under the heat-lubrication-load combined action in real time is solved.

The invention relates to a forklift battery SOC prediction method based on a neural network. The method comprises the steps that a voltage signal of a forklift battery end and forklift working condition state data are acquired; based on the voltage signal and the forklift working condition state data, a multi-dimensional feature vector is obtained, and the multi-dimensional feature vector comprises an original voltage value, a voltage change rate, a working condition state code and historical reference voltage; reasoning the multi-dimensional feature vector based on a lightweight neural network model to obtain an initial SOC predicted value; and based on the standing open-circuit voltage tag and the charging event tag, correcting the initial SOC predicted value to obtain a target SOC value. According to the method, the SOC of the lead-acid battery of the forklift can be better estimated in combination with the actual working condition of the forklift, and the estimation accuracy is improved.

The invention discloses an acoustic signal partial discharge detection system of a distribution network transformer, which is characterized by comprising a multi-sensor acquisition unit, a synchronous clock control unit, a signal preprocessing unit, an edge calculation processing unit and a communication and storage unit, the edge calculation processing unit integrates a cross-band denoising module, a multi-scale feature extraction module, a lightweight neural network diagnosis module and a diagnosis result fusion module, wherein the cross-band denoising module is used for performing joint denoising on audible sound and ultrasonic frequency band signals. According to the invention, a set of partial discharge detection scheme for a distribution network transformer scene is established, and the scheme flow comprises denoising, feature extraction, neural network training and diagnosis. According to the scheme, audible audio frequency band and ultrasonic frequency band data are covered, and interference generated by mechanical vibration of the transformer in the ultrasonic frequency band is removed through a soft threshold value of wavelet transformation and a pulse detection denoising algorithm.