61results about How to "Satisfy real-time" patented technology

The invention relates to the technical field of chemical process fault diagnosis, in particular to a chemical process fault diagnosis method based on multi-scale spatial-temporal feature fusion, which comprises the following steps: regularly acquiring fault data in a chemical production process, preprocessing the fault data and inputting the preprocessed fault data into an MTS-F model; the MTS-F network comprises a time sequence feature extraction module and a spatial feature extraction module which are parallel, and the time sequence feature extraction module comprises a long short-term memory network, a channel attention mechanism module, a one-dimensional convolution layer and a global maximum pooling layer which are connected in sequence; the spatial feature extraction module comprises a GS-CNN module, a maximum pooling layer, two groups of spatial separable convolution layers and maximum pooling layers, and a global maximum pooling layer which are connected in sequence; and after the outputs of the two parallel networks are spliced, a prediction result of fault classification is output through a discarding layer and a full connection layer. According to the method, the time sequence and space features are extracted simultaneously in a parallel mode, the fault diagnosis capability is improved, the model parameter quantity and the calculation quantity are reduced, and good generalization performance is achieved.

The invention discloses an automatic driving target detection method based on improved RT-DETR, and relates to the technical field of target detection and automatic driving, and the method comprises the steps: obtaining a disclosed automatic driving target detection data set SODA10M; a backbone network of an FSSD-DETR algorithm is constructed, and a BaseBlock optimization feature extraction process in a P5 layer is replaced with an FDC module; a small target detection layer is introduced into the neck network of the RT-DETR algorithm to improve the detection performance of a remote small target; a neck network of an FSSD-DETR algorithm is constructed, and a Zoomcat module and a ScalSeq module are used to optimize a feature fusion process; an up-sampling operator of the FSSD-DETR algorithm is constructed, and a DySample up-sampling operator is used to replace a nearest neighbor interpolation method so as to improve the image quality; therefore, good performance is maintained when complex traffic scenes such as small targets, dense vehicles and shielding are processed.

The invention discloses an electromagnetic transient simulation and mapping method and system oriented to multi-physics field coupling, and belongs to the technical field of substation digital operation and maintenance and electromagnetic simulation. Deploying multiple types of sensors in a transformer substation, preprocessing the acquired multi-physical field data, and fusing with the three-dimensional twinborn model; a coupling algorithm is adopted to carry out multi-physical field coupling analysis, multi-physical quantity display and thermodynamic diagram superposition are realized, and scene evolution change simulation under various typical working conditions is supported; a micro-service architecture is adopted to construct a digital twin system, double closed-loop verification of simulation and actual measurement is performed through actual measurement data and simulation data of a real-model test field, model optimization is completed, and real-time mapping of electromagnetic transient dynamic simulation and a transformer substation physical entity is realized. The method improves the precision of model and entity mapping, meets the requirements of real-time tracking and response of the electromagnetic transient change of the transformer substation, adapts to the digital transformation of different transformer substations, and provides a basis for the operation and maintenance of the transformer substations.

The invention discloses a method for quantitatively evaluating the immune level of subsequent commutation failure of an AC-DC hybrid system, and the method comprises the steps: carrying out the analysis of the transient process of the system based on the transient reactive dynamic characteristics of the AC-DC hybrid system on the basis of achieving the cross-regional power transmission through a high-voltage DC power transmission technology; the essential relation among the electrical quantities in the transient process is obtained; and deducing an online evaluation index GCFII for reflecting the immune level of the subsequent commutation failure of the system in combination with known system parameters and electrical signals acquired by the PMU. According to the method, the influence of the direct current control mode is considered by obtaining the online evaluation index GCFII, and real-time updating can be carried out along with the change of system operation parameters, so that the requirements of real-time performance and accuracy can be met, a theoretical basis is provided for power grid construction and operation of operation personnel, and safe operation of a power grid is effectively ensured.

The invention discloses an engineering wake flow model dynamic correction method and system based on the actually measured wind speed of a wind power plant, and the method comprises the steps: constructing a dynamic wake flow model according to the time delay characteristic of wake flow propagation, and building an optimization model; performing first-order Taylor expansion on a dynamic single wake flow model and a dynamic superposition model in the dynamic wake flow model to obtain an increment form of the dynamic wake flow model, and forming a discrete state space matrix of a historical wind speed predicted value; and according to the discrete state space matrix of the historical wind speed predicted value, converting the optimization model into a discrete form, setting constraint conditions of a wake flow expansion coefficient and a superposition correction coefficient, converting the constraint conditions into a standard quadratic programming problem, and solving to obtain an optimal model parameter so as to realize dynamic real-time correction of the wake flow model. According to the method, the parameters of the engineering wake flow model are dynamically optimized according to the historical wind speed measured value, so that the calculation precision of the engineering wake flow model is improved, and meanwhile, the wake flow calculation speed and timeliness are guaranteed.

The invention provides a distributed energy information modeling method for on-site coordination control of an active power distribution network, and belongs to the technical field of power grid control. Comprising the following steps: based on intelligent electronic equipment and a distributed energy system, establishing an equipment side information model according to an IEC61850 standard; according to an IECCIM expansion principle, distributed energy CIM expansion is carried out on the power system information model at a master station side; performing static topology model fusion and measurement model fusion on the equipment side information model and the electric power system information model after CIM expansion; and constructing a fusion model by utilizing a protocol mapping mode, and mapping data of the equipment side information model and the power system information model by utilizing the fusion model. According to the invention, the problem of model fusion between the terminal equipment and the distributed energy and the master station system is solved, a unified lightweight object-oriented information model for distributed energy cooperative control is established, and reliable and real-time exchange of equipment-equipment and equipment-master station information is realized.

The invention relates to the technical field of medical robot control, in particular to a neural network path planning method and system for a nasal digestive tract catheterization robot. The method comprises the following steps: establishing a digestive tract three-dimensional point cloud model; determining the pose of a path planning starting point and an end point, and generating a path point sequence; constructing an encoder network for extracting point cloud features of the digestive tract, and obtaining a trained encoder network; training the planning network by adopting the path point sequence to obtain a trained planning network; an encoder network and a planning network are combined with an alimentary canal point cloud model obtained in real time, and a path point sequence is obtained; calling a path planning network to update a global path according to environment information sensed by a sensor; the system comprises an encoder network module, a planning network module, a neural network path planning module and a continuous learning framework module. Through the mode, the efficiency of the robot in a complex, narrow and dynamic change environment is improved, and the real-time performance in an operation is met.

The invention relates to the field of welding seam detection, and discloses a welding seam center line real-time tracking method and device, computer equipment and a storage medium, and the method comprises the steps: obtaining a laser tracking evaluation speed, an ultrasonic tracking evaluation speed and a tracking speed threshold value of a to-be-detected weldment; if the laser tracking evaluation speed and the ultrasonic tracking evaluation speed are both larger than the tracking speed threshold value, the weld joint center line of the to-be-detected weldment is tracked according to a laser and ultrasonic simultaneous inspection strategy; and if the laser tracking evaluation speed or the ultrasonic tracking evaluation speed is smaller than or equal to the tracking speed threshold value, the welding seam center line of the to-be-detected weldment is tracked according to a main laser and auxiliary ultrasonic strategy. According to the invention, the obvious limitation of a single laser tracking technology under the conditions of a complex surface and strong illumination is overcome, and high-precision and high-robustness real-time tracking is realized through acousto-optic fusion.

The present application relates to the technical field of unmanned aerial vehicle, and discloses a field splicing method of distributed photoelectric tracking system, first, a frequency spectrum monitor is used to find an unmanned aerial vehicle target and send the azimuth and distance data of the target to a target detection unit, after the target is found by the target detection unit, the target is automatically or manually locked, and the information of the locked target is transmitted to a target tracking unit, the target tracking unit extracts features in the area around the original position of the target in the last frame, finds the position with the highest matching degree of the target features and takes the position as the new position of the target, then adjusts the camera to the new position, automatically or manually controls the linked interference device, and transmits satellite navigation and remote control interference signals to the target area until the unmanned aerial vehicle is driven away. The present application solves the problem of too large target position error of the frequency spectrum monitor, and has accurate positioning, and meets the real-time and success rate requirements of target search.

This invention relates to a physical information neural network-based intelligent performance prediction method and system for automated multi-material fiber placement on complex curved surfaces in aircraft. The invention includes: collecting fiber placement process parameters; performing standardized preprocessing and physical constraint-enhanced sampling to obtain an enhanced training dataset; constructing a high-order nonlinear sparse regression candidate library based on the enhanced training dataset; jointly identifying piecewise smooth physical differential equation systems and their mode switching logic using sparse Bayesian regression and Hidden Markov Models to obtain an embeddable inverse mechanism model; constructing a Physical Information Embedded Generative Adversarial Network (PI-GAN); collaboratively optimizing the generator and discriminator through adversarial training to output surface defect indices and mechanical performance indices; and initiating incremental self-learning when introducing new process scenarios to achieve model self-evolution. This invention achieves end-to-end accurate prediction of multi-dimensional quality indices such as surface defects and mechanical performance.

The invention discloses a trajectory prediction system based on convolution and self-attention. The trajectory prediction system comprises an overall network architecture, a training method and a reasoning output algorithm. The whole network architecture takes a road grating map, a vehicle motion state and target information as inputs, the inputs are processed through a DenseNet layer, a self-attention coding layer, a full-connection coding layer and a decoding output layer in sequence, and a prediction track is output; according to the method, a road grating map containing information such as a collision box and a road marking and a vehicle motion state are used as input, road environment constraints are associated with a self-attention mechanism, a polar coordinate track description mode fitting vehicle kinematics characteristics is adopted, a track parameter value range is limited, common errors and abnormal tracks are effectively avoided, and the method is suitable for popularization and application. Traffic rules and driving safety requirements are met, and the reasonability and safety of formed track prediction are prominent; based on 16000-frame multi-dimensional training data collected by a Carla simulator, errors are reduced in a verification set after optimization training, and errors in a crossing turning scene are reduced.

The invention relates to a multifunctional intelligent aloe bin, and relates to the field of artificial climate boxes, the multifunctional intelligent aloe bin comprises a planting bin, a plurality of brackets vertically arranged in the planting bin at intervals, and planting trays arranged above the brackets, and the planting bin is divided into a plurality of cultivation bin chambers by the planting trays; the environment sensor is arranged in the planting bin and used for detecting environment parameters; the execution mechanism is arranged in the planting bin and used for adjusting the growth environment or supplying growth substances; the control terminal is connected to the environment sensor, is arranged on the planting bin or is in communication connection with the planting bin, comprises a human-computer interaction interface and is configured to establish communication connection with a remote experimenter terminal; a control instruction from the remote experimenter terminal is received and responded to drive the execution mechanism to act; and data detected by the environment sensor is sent to a remote experimenter terminal. According to the invention, the experiment requirements of real-time monitoring and flexible control are met, and the experiment accuracy is improved.

The application discloses a spherical decoding detection method and device, electronic equipment and a readable storage medium, and belongs to the technical field of wireless communication. The application obtains a Q matrix and an R matrix by performing orthogonal triangular QR decomposition on a channel response matrix, wherein each column in the Q matrix is arranged according to the size of channel energy; determines a target equalization signal according to the Q matrix and the R matrix; performs an ML path search process according to the target equalization signal to select N branches with minimum metrics as survivor paths; performs ML complement path search on the survivor paths to obtain ML complement paths; takes the survivor paths with minimum metrics as ML paths, and obtains the likelihood ratio information of each bit of each symbol at each layer according to the ML paths and the ML complement paths. The application can reduce the operation complexity of the spherical decoding algorithm while maintaining the decoding detection performance.

The present invention discloses an embedded lightweight remote sensing target detection system, including: an embedded device and a target detection model deployed on the embedded device; the target detection model adopts a deep learning model, and is deployed to the embedded device after training on a GPU platform to obtain parameters; the target detection model includes: a structural reparameterization module, which is used to absorb redundant residuals and convolutional structures in the deep learning model into a backbone network; an improved residual structure module, which is used to achieve the same dimension of input and output through an improved residual connection method. The embedded lightweight remote sensing target detection system, based on structural reparameterization, increases the use of residual connections, and improves the residual connection method when the input and output dimensions are different, reduces the FLOPs of the model, improves the model detection accuracy and maintains the inference speed of the model, so as to achieve the real-time performance of remote sensing target detection under limited computing power of the embedded device, and is convenient for improving work efficiency.

The application discloses a DSP lock-free parallel clustering method based on adaptive division and boundary summary, comprising the following steps: obtaining original spatial point cloud data, determining a global calculation region and putting the global calculation region into a to-be-processed queue as a root node, and calculating the region potential energy corresponding to each region in the to-be-processed queue; when the regions in the to-be-processed queue are subjected to bisection cutting, final partitions are obtained; according to the final partitions, the original spatial point cloud data is traversed to obtain an extended data set and an index list of core attribution points; the index list of the extended data set is subjected to density clustering to obtain local clusters and a preliminary global cluster label vector; according to the triangle inequality, cluster summaries are screened and a union-find set is processed to obtain a union-find set result; and the union-find set result is mapped and converted into a continuous integer label to obtain a final clustering result. The application can realize a strictly consistent DBSCAN clustering effect on a multi-core DSP platform, and can meet the requirements of real-time performance, low communication overhead, scalability and high engineering reliability.

The invention relates to the field of industrial internet-of-things data transmission, in particular to an industrial internet-of-things data secure transmission method, system, device and medium, and the method comprises the steps: pre-registering client information at a server, and generating a client unique identifier; the client generates a client key according to the public key obtained from the server and the unique identifier of the client, and registers the key to the server; the client uses the registered key to initiate security authentication to the server, and after the authentication is passed, the server generates or negotiates with the client to generate a session key for data encryption transmission, and distributes the session key to the client; the client and the server encrypt or decrypt interaction data by using the session key; according to the method, the data transmission security of the industrial Internet of Things can be ensured, the real-time performance, the reliability and the usability are considered, and the method is particularly suitable for industrial control, monitoring and cooperative processing scenes with relatively high requirements on the security and the real-time performance, and has relatively high practical value and popularization prospect.

The embodiment of the invention provides an iridium satellite monitoring and early warning system and method, and belongs to the wireless communication signal monitoring technology. The system comprises a signal acquisition subsystem used for acquiring wireless signals in an iridium communication frequency band; the signal processing subsystem is connected with the signal acquisition subsystem and is used for performing iridium communication characteristic-oriented processing on the wireless signal so as to obtain analyzable effective signal data; the information analysis subsystem is connected with the signal processing subsystem and is used for analyzing the effective signal data and generating communication information for representing iridium communication activity; and the monitoring and early warning subsystem is connected with the information analysis subsystem and is used for executing monitoring analysis based on the communication information and outputting a corresponding monitoring result or early warning information. According to the scheme of the invention, acquisition, processing, analysis, monitoring and early warning of iridium communication signals are integrated in a unified system architecture, and continuous monitoring, structured analysis and timely early warning output of iridium communication activities are realized.

The application discloses a kind of fast simulation method of ion scattering track cluster driven by bypass factor, including constructing basic radiation parameter table;According to the basic radiation parameter table, search discrete energy point, form lookup table;According to the variation of bypass factor of each discrete energy point in lookup table, determine the scattering angle distribution characteristic parameter corresponding to discrete energy point, and generate scattering angle sample based on scattering angle distribution characteristic parameter and store in scattering angle sample pool;Initialize three-dimensional grid, set the initial position and direction of incident ion;Carry out incident ion transport and energy deposition cycle;Parallel execution of the simulation task of multiple incident ions, respectively in independent three-dimensional grid copy incident ion transport and energy deposition cycle, finally all three-dimensional grid copy is obtained Total dose distribution.This method achieves optimal balance between calculation efficiency and accuracy, so as to generate high-fidelity three-dimensional dose distribution chart.

The invention discloses a distributed synchronization method for multi-agent cluster data. The method comprises the following steps: determining an agent fragment structure; state updating and node fault detection in the cluster are realized by using a Gossip / SWIM protocol, and an active agent list in the cluster is determined; the maximum message length of network transmission is determined according to the network state, and a TCP / UDP protocol is adopted to carry out data communication between the intelligent agents; updating data information in the intelligent agent cluster by utilizing incremental synchronization operation; and domain-based synchronization is realized by using a cluster structure and a data label mechanism. According to the method, high reliability, real-time performance and safety in the data transmission process between agent clusters are met, and expandability and maintainability are supported, so that subsequent expansion and maintenance of the system are facilitated.

This invention discloses a multifunctional biomimetic robot for field exploration, comprising a body structure, a variable-radius slewing wheel-leg mechanism, a variable-stiffness connection mechanism, a flexible tail mechanism, and a front-end actuator. The body structure includes multiple base plates; the variable-stiffness connection mechanism includes multiple variable-stiffness connection components, with adjacent base plates connected by these components; the variable-radius slewing wheel-leg mechanism includes multiple variable-radius slewing wheel-leg components, each capable of both wheel-based and leg-based movement; the flexible tail mechanism possesses degrees of freedom for circumferential rotation and vertical bending; and the front-end actuator has obstacle removal and / or data acquisition functions. This design achieves high movement speed on flat terrain while simultaneously meeting the requirement for high obstacle-crossing capability on complex terrain. Furthermore, the body stiffness can be automatically adjusted, and the tail structure exhibits high controllability and flexibility.

The invention discloses a side end multi-scene perception control system and method based on mechanical arm vision, and relates to the technical field of intelligent technology. Through the unified software development kit and the intelligent sensing control software with the body, the intelligent terminal can be used for cooperative work with multiple types of mechanical arms, multiple types of depth cameras and multiple types of sensors, and seamless integration and cooperative work of the multiple types of mechanical arms and the depth cameras are achieved; multi-scene intelligent sensing and mechanical arm operation control under the limited computing power of a terminal are achieved, the intelligent sensing and operation control capacity of the system is improved, meanwhile, decoupling type integration of the mechanical arm, the depth camera and multiple sensors is achieved through a unified interface layer and a calling layer, the coupling degree and adaptation cost of the system are remarkably reduced, and plug and play is supported.

This invention discloses a method for inferring pilot stress and emotional states based on video, audio, and QAR data, belonging to the field of intelligent perception technology for multimodal pilot psychological states. The method includes temporal alignment, noise reduction, and data cleaning of the acquired data; unified modeling of multimodal features; outputting results and generating interpretive visualization reports. This invention improves the usability and deployment convenience of the system in real flight or simulated training environments; enhances the accuracy and robustness of psychological state recognition; strengthens the model's cross-human generalization ability; and meets the stringent real-time and reliability requirements of aviation scenarios.

The invention provides an analogue simulation method and device for determining a risk field based on a particle diffusion model, and the method comprises the steps: carrying out the analysis of the current sensing data collected by each target sensing device on a target vehicle in a simulation scene, and determining the risk particles included in each current sensing data; performing particle diffusion model updating processing by using each risk particle and the current environmental influence data, and determining at least one candidate particle diffusion model; for each candidate particle diffusion model, re-updating the candidate particle diffusion model by using effective secondary particles generated by chain reaction after collision to obtain at least one optimal particle diffusion model; fusing all the optimal particle diffusion models to determine a target particle diffusion model; and outputting a target risk field according to particle information in the target particle diffusion model. Therefore, the particle diffusion model and dynamics are combined to realize dynamic interactive modeling, so that the dynamic accurate evaluation of the risk field can be effectively realized.

The application discloses a building image segmentation method and system based on a SAM-GAN, and relates to the technical field of image data processing. The method comprises the following steps: acquiring an original remote sensing image, and preprocessing the original remote sensing image to obtain a preprocessed image; generating a preliminary building segmentation result based on a preset SAM-GAN model according to the preprocessed image; the SAM-GAN model comprises a SAM generator and a double discriminator network, the double discriminator network is used for optimizing the SAM generator, and the double discriminator network is a double discriminator architecture that fuses semantic and topological constraints; and performing geometric rule adaptive optimization post-processing on the preliminary building segmentation result to obtain a final building segmentation result. The application improves the accuracy, quality and generalization ability of the building image segmentation result.

The application discloses a laser radar point cloud object detection method and system, comprising the following steps: obtaining a three-dimensional point cloud of a laser radar, and after filtering interference point clouds and compensating for missing point clouds, obtaining a to-be-segmented point cloud; generating a two-dimensional distance image from the to-be-segmented point cloud; setting an angle constraint; the angle constraint is that the included angle between the line connecting two adjacent points in the two-dimensional distance image and the line connecting the radar center and the point farther away from the radar center among the two adjacent points is not less than an angle threshold; setting a distance constraint; the distance constraint is that the distance between the two adjacent points in the two-dimensional distance image is not greater than a distance threshold; traversing the points in the two-dimensional distance image, judging whether the two adjacent points satisfy the angle constraint or the distance constraint, if any constraint is satisfied, marking the two adjacent points as the same object, otherwise, marking the two adjacent points as different objects, and thus obtaining an object segmentation result. The object can be quickly detected under sparse point cloud data.

The application relates to the field of power system data acquisition and edge computing, and discloses a multi-source heterogeneous data pre-acquisition and edge intelligent processing method, which comprises the following steps: using a modal division mechanism, combining the operation state of power equipment, and performing condition correction on an adaptive sampling frequency to obtain a corrected sampling frequency, so as to acquire multi-source heterogeneous data; extracting fusion core features from multi-dimensional feature vectors; substituting the fusion core features into a general fault judgment model and a time sequence prediction model to respectively output equipment fault prediction results and time sequence prediction results, and obtaining early warning information; based on the fusion core features, the early warning information and the equipment fault prediction results of each acquisition node, and combined with the Lymerzi complexity theory, the spatio-temporal coupling law is identified; and based on the spatio-temporal coupling law and the time sequence prediction results, edge end decision suggestions are provided for source-grid-load-storage collaborative regulation. The application effectively improves the power system data processing efficiency and the power grid operation and maintenance regulation level.

The application provides a power load adaptation and output regulation control method, equipment and storage medium, and belongs to the technical field of power load regulation control, and the method comprises the following steps: calculating dynamic parameters based on load electrical parameters; generating an initial solution population of control parameters based on a chaotic mapping mechanism and a dynamic parameter value boundary; iteratively optimizing the initial solution population by using an improved artificial bee colony algorithm to obtain an optimal control parameter set matched with the current load characteristics; fusing the optimal control parameter set as a feedforward control amount and a feedback correction amount to generate a comprehensive control instruction; adjusting a power output unit according to the instruction; determining whether the power output reaches a target set value within a preset adjustment time window; if not, updating an iterative optimization input condition and repeatedly executing the iterative optimization; and if yes, maintaining the comprehensive control instruction. Based on the method, corresponding equipment and storage medium are also provided. The application realizes accurate and stable output of a power system within the first sine wave cycle after load mutation.

The invention discloses a motor sweeper operation path generation method and device, and the method comprises the steps: constructing a road connected graph based on a to-be-swept region map road boundary and a topological connection relation; coarsening the connected graph according to a preset merging strategy, and merging the fine-grained nodes into logic region nodes; constructing a final cleaning task area based on the coarsened connected graph; planning a sub-region cleaning sequence in combination with the current position of the cleaning vehicle; and constructing a global operation path. According to the invention, whole-process automation of cleaning area division and sequence planning is realized, manual intervention is not needed, and operation efficiency and reliability are improved; complex road topologies such as roundabout and the like are adapted, and sub-region division is more reasonable; a heuristic weight and an improved threshold acceptance algorithm are fused, the solving efficiency and the planning quality are balanced, and the requirements for the real-time performance and the accuracy of outdoor cleaning are met.

The present application relates to the technical field of online learning data security, and relates to an online learning platform data security access control method.The present application calls the historical operation behavior of a user before the user applies for data access after identity verification, analyzes the antecedent association of the applied access operation behavior sequence and the historical operation behavior, extracts the reference access operation behavior sequence of other users with the same attribute data in accessing the current data, compares and outputs the matching degree of the applied access operation behavior sequence, dynamically adjusts the access permission according to the analyzed risk value of the user in accessing the current data, increases the risk recognition accuracy of the user operation behavior, and according to the deviation degree of the access operation data and the user behavior portrait after the adjustment of the user access permission, determines the dynamic trust score in combination with the historical trust score of the user, realizes the quick response means when abnormal access operation occurs, and provides fine user trust management for the shuttle machine in the learning platform.

The invention belongs to the technical field of target recognition, and relates to an abnormal target recognition method, device and equipment and a readable storage medium. The method comprises the following steps: inputting a to-be-detected picture into an initial convolution module, and outputting an initial feature map; inputting the initial feature map into four feature extraction modules which are sequentially connected in series to obtain a first feature map, a second feature map, a third feature map and a fourth feature map; inputting the first feature map and the second feature map into a first feature fusion module, and outputting a first fusion feature map; inputting the second feature map and the third feature map into a second feature fusion module, and outputting a second fusion feature map; inputting the third feature map and the fourth feature map into a third feature fusion module, and outputting a third fusion feature map; respectively inputting the first feature map, the first fusion feature map, the second fusion feature map and the third fusion feature map into a first detection head, a second detection head, a third detection head and a fourth detection head; and obtaining an abnormal target identification result of the to-be-detected picture based on the output of all the detection heads.